Optics and Photonics: A Practical Reference

A laser diode is a forward-biased semiconductor junction with a resonator built into the chip. Electrons and holes recombine in a thin active layer, the emitted photons stimulate more recombination, and cleaved facets or a grating provide feedback. Edge-emitting diodes send light out of the chip's side as a strongly asymmetric beam. Because they are made on wafers by the same processes that make LEDs, they are the cheapest and most numerous lasers in existence, and almost every other laser on this sheet is pumped by one.

Electrical-to-optical efficiency reaches 50–70%, far above any other laser type, they modulate directly at gigahertz rates, and single units cost from a few dollars to a few hundred. Wavelengths run from about 375 nm to beyond 3 um depending on the material system. The weaknesses are beam quality and stability. The emitting aperture is a fraction of a micron in one axis and tens of microns in the other, so the output diverges asymmetrically and needs collimation. Wavelength drifts with temperature and current, typically 0.3 nm per degree, so anything spectroscopic needs stabilization, and facets fail catastrophically if driven too hard.

Use a diode whenever the wavelength exists and the beam quality is adequate, which covers most sensing, communications, and consumer applications. It is also the right pump for a fiber or solid-state laser, which is how the high-power market actually works. Move to a fiber or DPSS laser when you need a diffraction-limited beam, high pulse energy, or a wavelength diodes cannot reach efficiently. For a wavelength-critical measurement, budget for a distributed-feedback diode and temperature control rather than a bare Fabry-Perot device.

Wall-plug efficiency 50–70%, the highest of any laser type · single emitters from milliwatts to about 25 W, bars and stacks to kilowatts · wavelength drift roughly 0.3 nm per degree C · direct modulation to tens of GHz · unit cost from a few dollars for consumer parts to hundreds for single-frequency devices.

Pump diodes at 915 and 976 nm behind essentially every fiber laser; 1,310 and 1,550 nm distributed-feedback diodes in telecom transmitters; blue diodes in projectors and welding of copper; the diode bars behind industrial direct-diode cutting systems.

Diode lasers follow semiconductor economics: cost per watt has fallen by more than an order of magnitude since 2000 as fiber laser demand pulled pump volume up, and that decline is what made kilowatt fiber lasers affordable. The market splits sharply between commodity emitters sold by the thousand and specialty single-frequency devices that cost hundreds each, and the two barely compete.

7 minHow Laser Diodes Work - The Learning Circuitelement14 presents · 100k+ views

8 minHow does semiconductor laser work? | Explained with animationVROOK Learning · 100k+ views

7 minHow a LASER DIODE Works ⚡What is a LASER DIODEVirtualBrain [ENG] · 100k+ views

Progress of Edge-Emitting Diode Lasers Based on Coupled-Waveguide Concept (Micromachines via PMC) · High Brightness Diode Lasers Based on Beam Splitting and Polarization Combining (Applied Sciences)

A vertical-cavity surface-emitting laser puts the resonator perpendicular to the wafer instead of along it. Two stacks of quarter-wave layers form mirrors above and below a very thin active region, and light leaves through the top surface as a round, low-divergence beam. Because the cavity is short, a VCSEL naturally runs on a single longitudinal mode, and because it emits upward it can be tested on the wafer before dicing, which changes the manufacturing economics entirely.

Wafer-level test and a circular beam are the two properties that made VCSELs the sensing laser of choice: they cost cents at volume, need no beam-shaping optic, and can be built as two-dimensional arrays with thousands of emitters on one die. Threshold current is low, so power consumption is modest, and wavelength drifts only about 0.06 nm per degree, five times less than an edge emitter. The limits are power and wavelength. A single aperture makes milliwatts to tens of milliwatts, and the material systems work well around 850–940 nm with 1,300–1,550 nm devices harder and more expensive.

Choose VCSELs for short-reach optical links, optical mice, proximity sensing, and any time-of-flight or structured-light system that needs an array of emitters. The consumer face-recognition dot projector is the canonical high-volume case. Use an edge-emitting diode instead when you need hundreds of milliwatts or more from one emitter, or a wavelength outside the comfortable VCSEL range. For LiDAR at longer range, the choice between 940 nm VCSEL arrays and 1,550 nm fiber lasers is really a choice between cost and eye-safe power budget.

Single aperture typically 1–10 mW, arrays to tens of watts · circular beam with roughly 15–25 degree divergence, needing no correction · wavelength drift about 0.06 nm per degree C · testable at wafer level, which is the main cost advantage · commodity 850 and 940 nm devices cost cents at volume.

Apple's Face ID dot projector and proximity sensors, which took VCSELs to billions of units; 850 nm arrays in short-reach datacenter links; automotive and consumer LiDAR illuminators; laser mice, the first mass consumer application.

The cost advantage is wafer-level test, and it is worth being precise about why. An edge emitter has to be cleaved and diced before anyone can tell whether it lases, so a bad die has already absorbed its packaging cost; a VCSEL is probed on the wafer, so the yield loss lands before the expensive step. That is what gets commodity 850 and 940 nm parts to cents apiece, and it only works at consumer volume: the capacity built for face recognition and proximity sensing is what everyone else now buys from. A buyer ordering thousands rather than millions is really ordering a custom array, where mask and qualification costs dominate and the cents-per-part figure does not apply. Moving to 1,300 or 1,550 nm means a different material system and a much smaller market, which is why those devices stay expensive.

5 minWhat is VCSEL Laser (Vertical Cavity Surface Emitting Laser)?Fiber Optics For Sale Co. · 100k+ views

5 minKent Choquette: Introduction to Vertical-Cavity Surface-Emitting Lasers (VCSELs) and ApplicationsSPIETV · 5k+ views

5 minWhat is VCSEL Laser Vertical Cavity Surface Emitting LaserProf. HC Kuo · 10k+ views

Vertical Cavity Surface-emitting Lasers (RP Photonics Encyclopedia) · Harnessing the capabilities of VCSELs: unlocking the potential for advanced integrated photonic devices and systems (Light: Science & Applications)

An ordinary diode laser emits a photon when an electron crosses the bandgap, which fixes the wavelength to the material. A quantum cascade laser instead uses transitions between engineered sub-bands inside a stack of quantum wells, so the wavelength is set by layer thickness rather than by chemistry. One electron cascades through 25–75 stages and emits a photon at each, which is where the gain comes from. The result is a semiconductor laser that reaches 3–12 um and beyond, where nothing else compact operates.

It opens the mid-infrared, which is where most molecules have their strongest absorption features, so a QCL plus a detector makes a gas sensor with parts-per-billion sensitivity in a package the size of a matchbox. Wavelength is designed rather than discovered, and tunable external-cavity versions sweep hundreds of wavenumbers. The costs are thermal and financial. Wall-plug efficiency is typically 10–20%, so most of the drive power becomes heat that a thermoelectric cooler must remove, and devices cost thousands to tens of thousands of dollars because volumes are small and the epitaxy is demanding.

Use a QCL when you need a coherent mid-infrared source: trace gas detection, standoff chemical identification, infrared countermeasures, and spectroscopic imaging. It is the only practical laser for most of that band. Do not use it where a broadband source and a filter would do, since a thermal emitter costs a hundredth as much, and do not assume room-temperature continuous operation at the far end of the range, where devices often need pulsed drive or deep cooling. In the terahertz region QCLs work but generally require cryogenic cooling.

Emits roughly 3–12 um in the mid-infrared and into the terahertz · one electron produces a photon at each of 25–75 cascade stages · wall-plug efficiency typically 10–20% · continuous room-temperature operation across much of the mid-infrared, with cooling needed at the extremes · unit cost commonly $5,000–30,000.

Trace gas analyzers for methane, ammonia, and nitrous oxide; infrared countermeasure systems on aircraft; breath analysis and standoff explosives detection research; Alpes Lasers, Block Engineering, and Leonardo DRS as long-standing suppliers.

The price sits in the epitaxy and in the volumes rather than in the materials. A device is a stack of 25 to 75 precisely grown stages, the yield on that is what it is, and nobody makes them by the million, so $5,000–30,000 is the normal range and the laser is usually the largest line item in the analyzer built around it. The consequence for anyone selling a QCL instrument is that the sensitivity has to be the product, because a thermal source and a filter cost a hundredth as much and answer most gas-detection questions well enough. So the market has mostly settled on regulatory emissions monitoring and defense countermeasures, where parts-per-billion sensitivity or a band nothing else reaches is the whole reason for the purchase. Wall-plug efficiency of 10–20% is a second-order cost, but it sets the cooler and the power budget, so it decides whether a product can be handheld and battery-powered. If QCL prices ever fall tenfold it will be because one high-volume application arrived, not because the epitaxy got cheaper on its own.

3 minQuantum Cascade Laser (QCL) TechnologyLeonardo DRS · 5k+ views

10 minUnderstanding Quantum Cascade LaserHyunwoo Y. · 1k+ views

24 minExplorations of the Quantum Cascade Laser — with Jerome FaistPhotonics Media · under 1k views

Quantum Cascade Lasers (RP Photonics Encyclopedia) · Quantum cascade laser roadmap (arXiv)

A fiber laser uses a length of glass fiber doped with a rare earth as the gain medium. Diodes pump the fiber through its cladding, the dopant absorbs that light and amplifies at its own wavelength, and fiber Bragg gratings written into the core form the mirrors. Ytterbium at 1,030–1,080 nm dominates industrial use; erbium at 1,550 nm serves telecom and eye-safe applications; thulium reaches 2 um. Because the gain medium is a long thin waveguide, the beam is defined by the fiber rather than by alignment.

The geometry is what makes it good: enormous surface area for cooling, a beam whose quality is set by the fiber core, and no free-space alignment to drift. Wall-plug efficiency of 30–40% is far above lamp-pumped systems, maintenance is minimal, and multi-kilowatt units are ordinary industrial equipment. The limits come from intensity in a small core. Nonlinear effects such as stimulated Raman and Brillouin scattering cap peak power and narrow linewidths, and transverse mode instability limits single-mode power. Delivering the beam through fiber also means back-reflection from shiny metal can damage the laser.

Fiber lasers are the default for industrial cutting, welding, and marking of metals, and for anything needing a good beam at high average power with low maintenance. At 1,550 nm they are the standard for long-range eye-safe LiDAR and for directed-energy work. Choose a DPSS or disk laser instead when you need high pulse energy at low repetition rate, where nonlinearities in a fiber become the limit, or a wavelength ytterbium and erbium cannot reach. For thick-section cutting, compare against CO2, which still has an advantage on some non-metals.

Industrial units from 100 W to over 100 kW, with 1–20 kW covering most cutting and welding · wall-plug efficiency 30–40% · near-diffraction-limited beam from single-mode fiber · pump diodes at 915 or 976 nm · maintenance intervals measured in tens of thousands of hours.

IPG Photonics, which built the industrial fiber laser market; nLIGHT and Coherent in the same space; erbium-doped fiber amplifiers, the same technology that made long-haul optical communications possible; 1,550 nm fiber lasers in automotive LiDAR.

Cost per kilowatt has fallen roughly tenfold since 2010, driven by pump diode prices and by Chinese entrants such as Raycus and Max Photonics taking a large share of the mid-power market. That decline moved fiber lasers from a premium option to the default for sheet metal cutting, and it displaced CO2 lasers from most metalworking. The remaining differentiation is at very high power and in beam-shaping for welding.

13 minHow a Fiber Laser Worksnuferncorporation · 500k+ views

5 minHow does the CNC Fiber laser cutting machine work? - FactoriesFactories · 100k+ views

14 minEverything You Need to Know About Fiber Laser Metal CuttingThat Mom With A Laser · 10k+ views

Fiber Lasers (RP Photonics Encyclopedia) · Transverse Mode Instability (arXiv)

A diode-pumped solid-state laser uses a crystal, usually neodymium-doped YAG or vanadate, pumped by diode bars instead of a flashlamp. The crystal lases at 1,064 nm, and nonlinear crystals frequency-double or triple that to 532 and 355 nm, which is how most green and ultraviolet industrial laser light is made. Geometry varies: rod, slab, and thin disk, with the disk arrangement spreading heat over a large face so it can reach high average power with good beam quality.

A solid crystal stores far more energy than a fiber core, so DPSS systems produce high pulse energies at low repetition rate, which is exactly where fiber lasers struggle. The output is easy to frequency-convert, giving access to visible and ultraviolet wavelengths. Efficiency of 20–30% is well above the 1–3% of the lamp-pumped lasers they replaced. The weaknesses are thermal and mechanical: heat in the crystal causes thermal lensing that changes with power, so beam quality varies with duty cycle, and free-space optics need alignment and a clean environment.

Choose DPSS when you need high pulse energy, ultraviolet or visible output, or a wavelength that fiber cannot reach. Marking and micromachining at 355 nm, laser rangefinding, and scientific pump sources are the standard cases. Prefer a fiber laser for continuous industrial cutting, where it is cheaper and needs less maintenance. When comparing quotes, ask about thermal lensing across the operating range rather than at one setpoint, because a beam specified at full power can be considerably worse at half.

Typically 1,064 nm from Nd:YAG or Nd:YVO4, doubled to 532 nm and tripled to 355 nm · efficiency 20–30% against 1–3% for lamp-pumped · pulse energies from millijoules to joules, far above fiber at low repetition rate · thin-disk geometry reaches multi-kilowatt average power with good beam quality · unit cost commonly $10,000–100,000.

355 nm DPSS lasers marking and drilling electronics; Nd:YAG rangefinders in military systems; TRUMPF thin-disk lasers in automotive welding; the green laser pointer, which is a tiny DPSS laser with a frequency-doubling crystal.

DPSS lost the continuous-wave industrial market to fiber lasers and kept the two jobs fiber cannot do: high pulse energy at low repetition rate, and visible or ultraviolet output from frequency conversion. That split is largely about assembly cost. A DPSS head is free-space optics, so a large share of what a buyer pays for is alignment labor and the mounts that hold the alignment, and that content falls slowly with volume, while a fiber laser is spliced together and carries almost none of it. Ongoing cost matters too, since nonlinear crystals degrade under ultraviolet, so a 355 nm system carries a service and consumables budget a 1,064 nm fiber laser does not. The $10,000–100,000 range holds because nothing cheap emits at 355 nm directly, so the customer is paying for the conversion chain rather than for the photons. If a direct or fiber-based source ever reaches the same pulse energy in the ultraviolet, expect DPSS to lose that market the way it lost cutting.

3 minLasers - Direct Diode vs Diode-Pumped Solid-State (DPSS)Laser Focus · 10k+ views

22 minHigh Power Diode Pumped LaserTech Ingredients · 50k+ views

5 minCoherent Laser Munich 2017 - Diode-Pumped Solid-State LasersCoherent Corp. · 1k+ views

μ-MOPA Architecture for Photonic Integrated Solid State Laser (arXiv) · Diode-pumped Nd:YVO4 Laser and Nonlinear Optics (University of New Mexico)

A gas laser excites a gas mixture with an electrical discharge and takes light from transitions in the gas itself. Three types still matter commercially. CO2 lasers emit at 10.6 um in the long-wave infrared and dominate cutting and engraving of non-metals. Excimer lasers use a noble-gas halide to emit deep ultraviolet at 193 or 248 nm, and are the light source for semiconductor lithography. Helium-neon lasers at 633 nm survive as cheap, extremely stable alignment and metrology references.

Each type persists because it does something no solid-state laser does as well. CO2's 10.6 um wavelength is absorbed strongly by organics, so it cuts acrylic, wood, and fabric cleanly where a 1 um fiber laser passes straight through. Excimer's deep ultraviolet photon energy breaks bonds directly, which is what makes it right for lithography and for corneal surgery. HeNe has exceptional frequency stability for the money. The weaknesses are shared: gas needs replacing, discharge electrodes and optics erode, wall-plug efficiency is 5–20%, and the systems are physically large.

Use CO2 for non-metals and for thick-section cutting where its absorption suits the material, and accept the maintenance. Use excimer where the application needs deep ultraviolet photon energy rather than merely short wavelength, which in practice means lithography, ophthalmic surgery, and annealing. Use HeNe as an alignment and interferometry reference where a stable visible line matters more than power. Everywhere else, a diode, fiber, or DPSS laser is smaller, more efficient, and needs less service.

CO2 emits at 10.6 um with wall-plug efficiency around 10–20% and powers to several kilowatts · excimer at 193 nm (ArF) and 248 nm (KrF), the deep ultraviolet used in lithography · HeNe at 633 nm, milliwatts, with excellent frequency stability · consumable gas and optics are ongoing costs · systems are bench or cabinet scale rather than modules.

CO2 laser cutters across sign making and packaging; ArF excimer sources inside every deep ultraviolet lithography scanner; LASIK systems using 193 nm excimer; HeNe references in dimensional metrology and in classic interferometers.

Three separate businesses hide under one heading and they have different economics. CO2 is a commodity: fiber lasers took the metal cutting, so what is left is non-metals, and that end is served by low-cost tubes going into sign-making and hobby machines where purchase price decides the sale. Excimer is the opposite, a small number of very expensive sources sold to a handful of lithography scanner makers, so the revenue tracks semiconductor capital spending rather than laser demand. HeNe survives on stability per dollar in metrology, in volumes that keep shrinking, because nothing cheaper holds a visible line that well. What all three share is that purchase price understates the cost: gas needs replacing, electrodes and optics erode, and wall-plug efficiency of 5–20% against 30–40% for a fiber laser means several times the electricity and cooling for the same output. If you are comparing a quote against a solid-state alternative, put consumables and service on the same line as the capital cost, or the comparison will favor the gas laser for the wrong reason.

8 minWhat is Co2 Laser? How does it work? | Physics | Explained with animationVROOK Learning · 100k+ views

1 minThe principle of CO2 lasersLuxinar · 10k+ views

16 minHow a CO2 Laser Works *in up to 22 available languages.Inventors Workshop · 5k+ views

Gas Lasers (RP Photonics Encyclopedia) · State-Of-The-Art and Trends in CO2 Laser Cutting of Polymeric Materials (Materials)

An ultrafast laser emits pulses of picoseconds or femtoseconds by mode locking: forcing many longitudinal modes to oscillate with a fixed phase relationship so they interfere into a single short burst circulating in the cavity. A 100 fs pulse containing a microjoule reaches gigawatts of peak power while averaging only a watt. Because the pulse is shorter than the time heat takes to diffuse out of the illuminated volume, material is removed by ionization rather than melting, which is the whole point.

Cold ablation gives cuts with no recast layer, no burr, and almost no heat-affected zone, which is why ultrafast lasers do ophthalmic surgery, cut stents, and drill fuel injector nozzles. High peak intensity also drives nonlinear processes that no other source reaches, including multiphoton microscopy and attosecond generation. The costs are price and throughput. Systems run tens to hundreds of thousands of dollars, average power is modest so removal rates are low, and the pulses need dispersion management through every optic in the beam path.

Use an ultrafast laser when thermal damage is unacceptable or when the process is inherently nonlinear. Transparent-material processing, medical device manufacturing, and multiphoton imaging are the durable applications. Do not use one where a nanosecond pulse would do, because you are paying a large premium for pulse duration you do not need. When comparing systems, compare average power and repetition rate rather than pulse duration alone, since throughput is set by how many pulses per second you can deliver, not by how short each one is.

Pulses from about 10 fs to tens of ps · peak powers of megawatts to gigawatts at watt-level average power · pulse energy from nanojoules to millijoules · almost no heat-affected zone, which is what distinguishes the process · system cost commonly $50,000–500,000.

Femtosecond lasers cutting corneal flaps in LASIK; stent and injector nozzle machining; multiphoton microscopy in neuroscience; the chirped pulse amplification technique that won the 2018 Nobel Prize in Physics and made high-energy ultrafast systems possible.

The premium buys pulse duration, and it has to come back either as an outcome no other process delivers or as scrap the customer stops making. Throughput is the weak point: average power is modest, so removal rates are low, and at $50,000–500,000 a system the cost per part only works out when the part is small, valuable, or currently failing at another process. That is why the durable markets are medical devices, where a stent or a corneal flap is worth a great deal per square millimeter, rather than general machining. The number to watch is average power, since throughput is pulses per second and every increment of average power lowers cost per part and brings a wider class of parts into range. If you are evaluating a purchase, work out parts per year first and divide the system and service cost by it, because a comparison on pulse duration alone says nothing about whether the machine is worth buying.

20 minThe Incredible Femtosecond LaserAsianometry · 100k+ views

4 minPRINCIPLES OF MODE-LOCKING - PASSIVELY MODE-LOCKED LASERSOptcomm · 50k+ views

3 minMode Locking explained!Your Favourite TA · 10k+ views

Mode Locking (RP Photonics Encyclopedia) · Scientific background: Groundbreaking inventions in laser physics (Nobel Prize)

A light-emitting diode is the same junction as a laser diode without a resonator: carriers recombine and emit photons spontaneously, in all directions and over a spectral width of tens of nanometers. White LEDs are blue emitters with a phosphor that converts part of the output. The blue gallium nitride emitter, which took decades to make efficient and won the 2014 Nobel Prize, is what turned LEDs from indicators into the dominant lighting technology.

Efficacy of 150–200 lumens per watt beats every other light source, lifetimes run tens of thousands of hours, and the devices switch in nanoseconds, which is why they double as sources for optical sensing and short-range communication. Cost is measured in cents. The limits matter when you want to do optics with them. LEDs emit into a wide angle from an extended area, so they cannot be collimated tightly or focused to a small spot, and their spectrum is far too broad for interferometry or narrow spectroscopy. Output falls with junction temperature, so thermal design decides real-world performance.

Use LEDs whenever you need light rather than a beam: illumination, indicators, machine-vision lighting, fluorescence excitation with a filter, and UV curing. They are also the right choice for high-speed switching at low cost. Move to a laser diode when the light has to be collimated, focused to a small spot, or made narrowband. For fluorescence work, an LED plus a good excitation filter now competes well with a laser at a fraction of the cost.

Efficacy 150–200 lm/W for white devices, against roughly 15 for incandescent · spectral width typically 20–40 nm · lifetime of tens of thousands of hours, defined as the point at which output falls to 70% · output falls measurably as junction temperature rises, so heatsinking sets real performance · unit cost in cents at volume.

General lighting, which LEDs took over almost completely between 2010 and 2020; machine vision ring lights and bar illuminators; UV-C LEDs for disinfection; 365–405 nm arrays curing adhesives and resins in manufacturing.

General lighting is a commodity business now, and most of the improvement that drove it has already been taken: going from roughly 15 lm/W to 150–200 lm/W was better than a tenfold gain, and what is left comes a few percent at a time. Long life makes that harder for whoever is selling, since tens of thousands of hours means each customer buys again rarely, so the 2010–2020 replacement wave was a one-time market rather than a recurring one. Margin has moved to the parts of a luminaire that are not the emitter, and to specialty devices such as UV-C, horticultural, and automotive LEDs, where the specification and the reliability data are worth paying for and volumes are smaller. For a system designer the useful economic fact is the comparison against a laser: an LED plus a good excitation filter costs on the order of a hundredth of a laser source, so the question to settle before specifying a laser is whether the application needs a beam or only needs light.

19 minHow LED Works - Unravel the Mysteries of How LEDs Work!The Engineering Mindset · 1m+ views

9 minHow LED works ⚡ What is a LED (Light Emitting Diode)VirtualBrain [ENG] · 1m+ views

Light-emitting Diodes (RP Photonics Encyclopedia) · LED Basics (US Department of Energy)

A supercontinuum source takes short high-peak-power pulses and sends them into a nonlinear fiber, usually a photonic crystal fiber whose dispersion is engineered to keep the pulse intense over a long interaction length. Self-phase modulation, soliton fission, and four-wave mixing spread the narrow input spectrum across hundreds or thousands of nanometers. The output is white light with a laser's spatial coherence: it can be focused to a diffraction-limited spot and coupled into a single-mode fiber, which no lamp can do.

One box replaces a rack of lasers when an experiment needs many wavelengths, and because the beam is spatially coherent it can be filtered to any line and still focus tightly. Brightness per nanometer far exceeds a lamp. The weaknesses are cost, noise, and spectral shape. Systems run tens of thousands of dollars, the output is not flat and falls off at the edges, and pulse-to-pulse intensity noise is significantly higher than a continuous-wave laser's, which matters for sensitive measurements. Power in any one narrow band is modest once you filter.

Choose a supercontinuum when an instrument needs tunable or multi-wavelength light with laser-like focusing: confocal and flow-cytometry excitation, broadband spectroscopy, optical coherence tomography, and component characterization across a wide band. Use an LED or lamp where spatial coherence does not matter, since it costs a hundredth as much. Use a tunable diode where you need one wavelength at a time with low noise. Check the noise specification against the measurement before assuming a supercontinuum can replace a stabilized laser.

Output spanning roughly 400–2,400 nm from one source · spatially coherent, so it couples into single-mode fiber and focuses to a diffraction-limited spot · spectral power density of milliwatts per nanometer · pulse-to-pulse noise considerably higher than a continuous-wave laser · system cost commonly $30,000–100,000.

NKT Photonics SuperK systems, the reference product in this category; confocal microscopy and flow cytometry using a filtered supercontinuum instead of several lasers; broadband optical coherence tomography; fiber and coating characterization in photonics labs.

The purchase case is substitution, so the comparison is against every laser the box replaces plus its controller and combining optics, not against one laser. That arithmetic works when an instrument genuinely needs many wavelengths and stops working at two or three, where a pair of diodes is cheaper and quieter. Most of the unit cost is the ultrafast pump laser and the photonic crystal fiber, so the price tracks what ultrafast sources cost (007) rather than anything specific to supercontinuum generation. Volumes are instrument-scale, spread across microscopy, flow cytometry, and optical coherence tomography rather than anything larger, which is why the $30,000–100,000 band has been stable and why no volume mechanism is in sight that would break it. Supply is concentrated enough to be worth asking about at design time, since NKT Photonics has been the reference product in this category for most of its commercial life.

10 minRobert Alfano on the supercontinuum: history and future applicationsSPIETV · 5k+ views

2 minSave time with the tunable SuperK CHROMATUNE supercontinuum white light laserHamamatsuPhotonicsAS · 1k+ views

4 minAdvantages of replacing lamps with SuperK supercontinuum white-light lasers.HamamatsuPhotonicsAS · 1k+ views

Supercontinuum Generation (RP Photonics Encyclopedia) · Recent advances in supercontinuum generation in specialty optical fibers (arXiv)

The terahertz band sits between microwaves and the infrared, roughly 0.1–10 THz, and is awkward because it falls between the two technologies that work well on either side. Electronic oscillators run out of speed below it and semiconductor lasers run out of photon energy above it, which is why it was called the terahertz gap. Three approaches now fill it: photoconductive antennas gated by femtosecond pulses, optical rectification and difference-frequency mixing in nonlinear crystals, and quantum cascade lasers at the upper end, usually cooled.

Terahertz radiation passes through paper, plastic, ceramics, and clothing while reflecting from metal and being absorbed strongly by water, which is a useful combination for inspection. Photon energy is far too low to ionize, so it is safe for people. Many molecular and lattice vibrations fall in this band, giving spectroscopic contrast that infrared misses. The problems are power and cost. Sources deliver microwatts to milliwatts, atmospheric water vapor absorbs strongly over any distance, and a time-domain spectroscopy system costs $100,000 or more.

Use terahertz for non-destructive inspection of non-conducting materials: paint thickness on cars, coatings on pharmaceutical tablets, voids in composites and foam. It is also established for security screening at short range. Do not expect it to work through walls or over long distances outdoors, because water vapor absorbs it heavily, and do not use it where an X-ray or ultrasound method already answers the question more cheaply. Treat any claim of long-range terahertz imaging with the atmospheric absorption spectrum in hand.

Band roughly 0.1–10 THz, corresponding to 3 mm to 30 um wavelength · typical source power microwatts to milliwatts · non-ionizing, so no radiation safety case is needed · atmospheric water vapor absorbs strongly, limiting outdoor range to meters · time-domain spectroscopy systems commonly $100,000+.

Terahertz paint thickness gauges used on automotive lines; pharmaceutical tablet coating inspection; airport body scanners operating at millimeter and submillimeter wavelengths; TeraView and Toptica as instrument suppliers.

This is the standard problem for an emerging instrument: a system at $100,000 and up has to displace a measurement the customer already makes some other way, and usually that other way is cheaper. Terahertz wins where nothing else measures the thing at all without destroying it, which in practice means wet paint thickness on a car body and coating thickness on pharmaceutical tablets. Those are narrow markets, but the measurement is worth a lot per unit, and that is what supports the price. Most of the system cost is the femtosecond pump laser and the timing hardware around it, so prices come down when ultrafast sources get cheaper rather than through any volume of terahertz systems, which stay instrument-scale. The largest imagined application, standoff imaging at range, is closed off by water vapor absorption rather than by cost, so no cost reduction opens it. Treat a plan that assumes terahertz gets cheap through scale with some skepticism, because the scale would have to come from an application nobody has found yet.

2 minTerahertz waves: The missing electromagnetic wavesFaculty of Engineering Tel-Aviv · 50k+ views

2 minA plasma-based nanodevice to generate powerful terahertz wavesEPFL · 10k+ views

4 minHyperTerahertz ImagingHyper Terahertz · 1k+ views

Terahertz Sources (RP Photonics Encyclopedia) · Intense terahertz radiation: generation and application (Frontiers of Optoelectronics)

A lens bends light by refraction, and every real lens fails to focus perfectly in ways that have names. Spherical aberration comes from the surface shape, chromatic aberration from the material's index varying with wavelength, and coma, astigmatism, and field curvature from off-axis geometry. Practical optics correct these by combining elements: an achromatic doublet cancels first-order color, an apochromat corrects three wavelengths, and a microscope objective or camera lens stacks six to twenty elements to control all of them at once.

Refractive optics are compact, achromatic designs are cheap in volume, and molded aspheres have made single-element collimation of laser diodes routine. Glass catalogs offer hundreds of index and dispersion combinations to design with. The limits are material and wavelength. Ordinary glass stops transmitting below about 350 nm and above about 2.5 um, so ultraviolet needs fused silica or calcium fluoride and thermal infrared needs germanium or zinc selenide, which cost far more. Refractive systems also absorb, which matters at high power, and every surface reflects unless coated.

Use refractive optics wherever the wavelength band and the power allow, which is most of visible and near-infrared work. Specify an achromat rather than a singlet whenever more than one wavelength is present, and an aspheric singlet when you are collimating one wavelength from a small source. Move to mirrors when the band is very wide, when the wavelength is outside common glass transmission, or when absorbed power would distort the element. For anything imaging, resolve the specification into numerical aperture and field of view first, since those set the design difficulty.

Common glass transmits roughly 350–2,500 nm; fused silica extends into the ultraviolet and germanium into the long-wave infrared · an uncoated glass surface reflects about 4% per face · achromatic doublets correct color at two wavelengths, apochromats at three · numerical aperture sets both resolution and depth of field · cost from under a dollar for a molded asphere to thousands for a research objective.

Achromatic doublets as the workhorse of optical benches; molded aspheres collimating laser diodes in every optical drive and sensor; microscope objectives with a dozen elements and immersion; germanium lenses in thermal cameras, where material cost dominates.

Two things set what a lens costs, and neither of them is the glass. The first is whether the surface can be molded, since a molded asphere costs under a dollar with one tool amortized over millions of parts and the same shape ground and polished individually costs hundreds, so the question at design time is whether annual volume justifies a mold. The second is element count, because every element is two surfaces to polish, coat, center, and mount, and assembly labor scales with it, which is how a research objective with a dozen elements reaches thousands of dollars while a doublet bought by the thousand does not. Material sits on top of both: staying inside roughly 350–2,500 nm keeps you in ordinary glass, and stepping out to fused silica, calcium fluoride, germanium, or zinc selenide costs money and lead time, far enough that in a thermal camera the lens often costs more than the sensor. If a design is drifting toward more elements or an exotic material, that is the point to check whether the specification really requires it, because both changes multiply cost rather than adding to it.

3 minHow Lenses FunctionCanon Imaging Asia · 1m+ views

3 minChromatic AberrationQuick Biochemistry Basics · 10k+ views

13 minDiopters, Aberration, and the Human Eye | Geometric optics | Physics | Khan Academykhanacademymedicine · 100k+ views

Lenses (RP Photonics Encyclopedia) · Introduction to Microscope Objectives (Nikon MicroscopyU)

A mirror reflects rather than transmits, which means the light never passes through a material and therefore never disperses. Two families exist. Metal-coated mirrors, aluminum, silver, or gold on a substrate, reflect broadly across a wide band at 90–98%. Dielectric mirrors stack dozens of quarter-wave layers whose reflections interfere constructively, reaching above 99.99% but only over a designed band and angle. Curved mirrors, off-axis parabolas among them, focus without any chromatic aberration at all.

No dispersion is the structural advantage: one mirror works identically at every wavelength it reflects, which is why every large telescope and every broadband instrument is reflective. Dielectric coatings reach reflectivity no transmissive element can match, which is what makes laser cavities and gravitational-wave interferometers possible. The awkwardness is geometric. Reflective systems fold the beam, so the optic gets in its own way and off-axis designs introduce their own aberrations and alignment sensitivity. Metal coatings tarnish and scratch easily, and gold works only above about 600 nm.

Use mirrors when the band is wide, the wavelength is outside good transmissive materials, or the power would damage or distort a lens. Use an off-axis parabola to focus broadband or ultrafast light without introducing dispersion. Choose dielectric coatings for a fixed wavelength and angle where high reflectivity matters, and metal for broadband work where a percent or two of loss is acceptable. Check the specified angle of incidence, because a dielectric mirror used at the wrong angle behaves very differently from its datasheet.

Protected aluminum reflects about 90% across the visible; silver about 98%; gold above 96% beyond 600 nm · dielectric stacks exceed 99.99% over a designed band · no chromatic aberration at all, since light never enters a material · dielectric performance depends on angle of incidence and polarization · substrate flatness commonly specified in fractions of a wavelength.

Off-axis parabolas focusing ultrafast pulses without adding dispersion; laser cavity mirrors with dielectric stacks; the 39-meter segmented primary of the Extremely Large Telescope; the ultra-low-loss coatings on LIGO's test masses.

Cost sits in two places, the polish and the coating run, and both scale badly with the specification. Polishing time rises steeply as the flatness spec tightens, so a catalog flat at a quarter wave is inexpensive and tightening it to a twentieth of a wave usually multiplies the price with no change in material. Coating is a batch process: a chamber run has a fixed setup and a fixed load, so the price per mirror is that run divided by how many fit in it. That is why a standard diameter with a catalog dielectric coating is cheap while a custom design in a custom size pays for the whole run and waits weeks for it. The practical rule is to take a catalog coating and a catalog diameter unless the design genuinely cannot, since the substrate is rarely the expensive part. At the extreme end, the coatings on the Extremely Large Telescope's segments and on LIGO's test masses are supply questions rather than price ones, because only a few facilities in the world can deposit to that loss and uniformity.

7 minRefractor vs Reflector telescope explained for beginnersAlien Tech · 50k+ views

6 minSpherical & parabolic mirrorsKhan Academy India - English · 100k+ views

6 minAstronomy - Ch. 6: Telescopes (11 of 25) The Newtonian Reflecting TelescopeMichel van Biezen · 10k+ views

Mirrors (RP Photonics Encyclopedia) · LIGO Optics (Caltech)

EUV lithography (0.33 NA)Semiconductor Manufacturing

A diffraction grating is a surface ruled or etched with a periodic structure, typically 300–3,600 lines per millimeter. Light scattered from each groove interferes, and the directions where the path differences are whole wavelengths depend on wavelength, so the grating spreads a beam into a spectrum. Blazing the groove profile tilts the reflected energy into one diffraction order, which is what makes a grating efficient rather than merely dispersive. Ruled gratings are cut mechanically; holographic ones are recorded in photoresist by interfering beams.

A grating disperses far more strongly than a prism and does so linearly in wavelength, which makes spectrometer design tractable. It works from the extreme ultraviolet to the far infrared, and volume-phase holographic versions reach very high efficiency in a band. Gratings also compress and stretch ultrashort pulses, which is what chirped pulse amplification requires. The weaknesses are stray light and orders. A fraction of the light scatters where it should not, limiting dynamic range in a spectrometer, and overlapping diffraction orders need an order-sorting filter. Efficiency is also polarization-dependent.

Use a grating whenever you need to separate wavelengths: any spectrometer, any wavelength-division multiplexer with more than a few channels, any tunable external-cavity laser. Use a prism instead when the priority is very low stray light over a modest band, or when the beam must stay in a straight line. In pulse compression, gratings are essentially the only option at high energy. Specify blaze wavelength and groove density against the band you actually care about, since efficiency falls away from the blaze.

Groove densities from about 300 to 3,600 lines per mm · dispersion far higher than a prism, and linear in wavelength · blazed gratings put 60–90% of the light into the chosen order · overlapping orders require order-sorting filters · stray light rather than resolution limits dynamic range in most spectrometers.

The dispersing element in nearly every laboratory spectrometer; grating pairs in chirped pulse amplification systems; echelle gratings in astronomical spectrographs; arrayed waveguide gratings performing the same function on a photonic chip.

Almost all of a grating's cost sits in the master rather than in the part you buy. A ruling engine cuts one master mechanically and replicas are cast off it, so a catalog grating is mostly the master amortized over everything sold from it, while a custom groove density or blaze wavelength means paying for a new master. Ruling engines are rare enough that custom work goes to a short supplier list with long lead times, which is one reason holographic recording, done with interfering beams rather than a machine tool, is the more available route for new designs. The rest of the cost lands downstream: efficiency falls away from the blaze, so a grating specified for the wrong band gets paid for in a brighter source or a more sensitive detector. The same function made lithographically on a photonic chip, as an arrayed waveguide grating, follows semiconductor economics instead and is cheap per channel at telecom volume, which is why chip-scale multiplexing displaced free-space gratings in that application.

2 minHow does a Grating work? | Raman for Beginners | Optical GratingsBruker · 10k+ views

15 mindiffraction gratings: Using the principles of interferences to examine lightPhysicsHigh · 50k+ views

4 minDiffraction gratingGetAClass - Physics · 50k+ views

Diffraction Gratings (RP Photonics Encyclopedia) · Diffraction Gratings Ruled and Holographic (HORIBA Scientific)

An optical coating is a stack of thin dielectric layers, each a fraction of a wavelength thick, deposited to control reflection through interference. Choose the thicknesses so reflections cancel and you get an antireflection coating; choose them to add and you get a high-reflector. The same tool makes bandpass filters, beamsplitters, and dichroics. Deposition method decides quality: ion beam sputtering gives the densest, most stable films and the lowest loss, while evaporation is cheaper and produces films that shift with humidity.

Coatings are what make multi-element optics possible at all. An uncoated glass surface loses 4% per face, so a ten-element lens would transmit under two thirds of the light; a good broadband coating brings that to a few tenths of a percent per surface. Laser mirrors above 99.99% and filters with steep edges both come from the same technique. The costs are environmental and mechanical. Coating performance depends on angle and polarization, films can absorb enough to matter at high power, and adhesion and abrasion resistance vary sharply with process.

Specify a coating on every optical surface in any system where throughput or stray light matters, and specify it against the real angle of incidence rather than normal. Use ion beam sputtered coatings where stability, low loss, or high laser damage threshold is required, and evaporated coatings where cost dominates and conditions are benign. Ask for the laser damage threshold at your pulse duration and wavelength, because it is the specification most often quoted at conditions that do not match the application.

An uncoated glass surface reflects about 4% per face · a good broadband antireflection coating leaves under 0.5% per surface, and a single-wavelength V-coat under 0.25% · high reflectors exceed 99.99% in a designed band · ion beam sputtered films are denser and more humidity-stable than evaporated ones · performance shifts with angle of incidence and polarization.

Antireflection coatings on every camera and eyeglass lens; dielectric high reflectors in laser cavities; dichroic beamsplitters separating excitation from emission in fluorescence microscopes; the ultra-low-loss coatings developed for gravitational-wave detectors.

The unit of cost in coating is a chamber run, not a part. One run coats every substrate that fits on the fixture, so the same recipe costs a fraction as much per piece on a tray of small lenses as on a handful of large ones, and a single first article carries the whole setup on its own. A custom design adds recipe development and a multi-week lead time, which is why a catalog coating is the right default whenever one exists for your band and angle of incidence. Ion beam sputtering costs more per run than evaporation largely because it deposits slowly, so what a buyer is paying for when specifying stability, low loss, and a high damage threshold is chamber hours. If a design genuinely needs a custom coating, the way to make it affordable is to order a year of demand in one lot rather than to negotiate the unit price.

8 minAntireflection coatingPhysics Unlimitted · 10k+ views

4 minHow Antireflective Coatings Work for SemiconductorsBrewer Science · 10k+ views

5 minPHYS 201 | Thin Film Interference , Antireflection CoatingProfessor Hafner · under 1k views

High-performance near- and mid-infrared crystalline coatings (Optica) · Development of ion-beam sputtered silicon nitride thin films for low-noise mirror coatings of gravitational-wave detectors (arXiv)

An optical filter passes some wavelengths and blocks others. Absorptive filters use colored glass or dye and work at any angle, with gentle edges. Interference filters use a dielectric stack and can be made with edges a few nanometers wide and blocking of six orders of magnitude, at the cost of shifting toward shorter wavelengths as the angle of incidence increases. Dichroic filters are interference filters used at 45 degrees to split a beam by wavelength, which is how every fluorescence microscope separates excitation from emission.

Modern hard-coated interference filters transmit above 95% in band while blocking at optical density 6 or better out of band, which is what makes single-molecule fluorescence possible. They are stable and do not degrade like the soft-coated filters they replaced. The limitations are angular and thermal. Passband shifts with angle, so a filter in a converging beam broadens and shifts; it also shifts slightly with temperature. Very narrow filters have correspondingly tight manufacturing tolerances and cost, and in a high-power beam the blocked light has to go somewhere.

Use interference filters wherever the separation between what you want and what you must reject is demanding, which is nearly all fluorescence, Raman, and spectroscopic work. Use absorptive filters for broad rejection in uncollimated light, where angle sensitivity would spoil an interference design. Always specify blocking as an optical density over a stated range, not merely as a passband, since out-of-band leakage is what limits real measurements. And put filters in collimated space if the design allows it.

Hard-coated interference filters transmit above 95% in band with blocking at optical density 6 or better · passband shifts toward shorter wavelengths as angle of incidence increases · edges as steep as a few nanometers between 90% transmission and deep blocking · absorptive filters are angle-insensitive with much gentler edges · cost from tens of dollars to over a thousand for narrow demanding parts.

Fluorescence filter cubes pairing excitation, dichroic, and emission filters; Raman edge filters rejecting the laser line while passing shifted light a few wavenumbers away; near-infrared bandpass filters in front of LiDAR detectors to reject sunlight; color filters on camera sensors.

Filter price tracks layer count, and layer count tracks how narrow and how steep you ask for. A wide bandpass with modest blocking is a short stack that runs quickly and yields well; a few-nanometer edge with optical density 6 needs a long deposition run and thickness control tight enough that part of every batch misses the spec, which is how the same size part goes from tens of dollars to over a thousand. So the cheapest move available is to specify the widest passband and the least blocking the measurement actually needs, since each step tighter moves the part up a price band. The other lever costs nothing: passband shifts with angle, so putting the filter in collimated space usually buys more real performance than buying a better filter and using it in a converging beam. At the far end of the volume curve the same physics is nearly free, since the color filters on a camera sensor are patterned lithographically at wafer scale rather than coated one substrate at a time.

2 minOptical Bandpass Filter Explained in 1 minuteSyronOptics · under 1k views

18 min#1046 Strange Optical FiltersIMSAI Guy · 5k+ views

1 minOptical Filter Guide 1 : Optical Filters and TypesSyronOptics · under 1k views

Optical Filters (RP Photonics Encyclopedia) · Handbook of Optical Filters for Fluorescence Microscopy (Chroma Technology)

Optical fiber guides light in a glass core surrounded by a lower-index cladding, so rays inside a certain angle are totally internally reflected and stay in the core. Single-mode fiber has a core around 9 um and supports one spatial mode, which is what long-distance communication needs. Multimode fiber has a 50 or 62.5 um core, carries many modes, and is easier to couple into but suffers modal dispersion. Specialty fibers add doping for gain, hollow cores for low nonlinearity, or stress rods to hold polarization.

Loss of about 0.2 dB per kilometer at 1,550 nm is the number that built the internet: light travels 100 km and still delivers over 1% of its power. Bandwidth is enormous, the medium is immune to electromagnetic interference, and the raw material is sand. The weaknesses show up in handling and physics. Fiber is fragile at small bend radius, joints need fusion splicing or precision connectors, and dirt on a connector end face is the most common fault in any installation. Chromatic and polarization-mode dispersion limit reach and have to be compensated.

Use single-mode fiber for anything long, anything at high data rate, and anything where beam quality out of the fiber matters. Use multimode for short links inside a building or a rack where cheaper transceivers and easier alignment win. Use specialty fiber when a physical property is the point: polarization-maintaining for interferometry, hollow-core where nonlinearity or latency matters, doped fiber where the fiber is the gain medium. Budget for cleaning and inspection of connectors, because that is where installed systems actually fail.

Attenuation about 0.2 dB/km at 1,550 nm, the loss minimum for silica · single-mode core about 9 um; multimode 50 or 62.5 um · bend radius limits are typically 10–30 mm for standard fiber, less for bend-insensitive types · fusion splice loss under 0.05 dB · connector contamination is the leading cause of link faults.

Submarine cables carrying nearly all intercontinental traffic; erbium-doped fiber amplifiers spaced along those routes; distributed acoustic sensing turning a buried fiber into thousands of vibration sensors; hollow-core fiber now deployed on latency-critical financial routes.

The glass is the cheap part. Standard single-mode fiber is drawn by the kilometer from a material that starts as sand, and in an installed link the cable is a small fraction of the cost next to trenching, duct, permits, and splicing labor. That is why operators pull far more fibers than they need on the first install: adding strands to a cable already going into the ground costs very little, and going back later means paying the whole civil cost twice. It also explains where the recurring cost sits, which is labor rather than parts, since connector contamination is the leading fault and cleaning and inspection are what prevent it. Specialty fiber inverts the picture. Polarization-maintaining, hollow-core, and rare-earth-doped fiber come from small specialty draws, so they are priced by the meter rather than by the kilometer, and a design that needs one should budget the fiber as a component rather than as cable.

6 minHow Does LIGHT Carry Data? - Fiber Optics ExplainedTechquickie · 1m+ views

2 minTotal Internal Reflection Demo: Optical FibersPhysics Demos · 1m+ views

6 minHow Fiber Optics Works 🌎VirtualBrain [ENG] · 100k+ views

Fibers (RP Photonics Encyclopedia) · G.652: Characteristics of a single-mode optical fibre and cable (ITU-T)

An acousto-optic modulator drives a piezoelectric transducer against a crystal, launching an ultrasonic wave that compresses and rarefies the material. The resulting periodic index variation acts as a moving diffraction grating, so part of the beam is deflected into the first order and shifted in frequency by exactly the acoustic drive frequency. Turning the radio-frequency drive on and off switches the diffracted beam, and changing the frequency steers it, which makes one device a switch, an attenuator, a frequency shifter, and a small-angle scanner.

It handles high optical power without damage, gives extinction ratios of 1,000:1 or better, and shifts frequency by a precisely known amount, which is what heterodyne detection and laser cooling depend on. There are no moving parts. The limits are speed and efficiency. Switching time is set by how long the acoustic wave takes to cross the beam, typically tens to hundreds of nanoseconds, so an AOM is far slower than an electro-optic modulator. Diffraction efficiency of 70–90% means real loss, and the deflection angle is small.

Use an AOM for pulse picking, fast shuttering, power stabilization, and frequency shifting in a laser system, especially at high power where an electro-optic crystal would be at risk. It is standard in atomic physics and in laser machining. Move to an electro-optic modulator when you need gigahertz modulation, and to a mechanical shutter when milliseconds are fine and you want no loss at all. Remember that an AOM in a beam path always deflects and always shifts, so the layout has to accommodate both.

Rise time of tens to hundreds of nanoseconds, set by the acoustic transit across the beam · diffraction efficiency typically 70–90% · frequency shift equals the drive frequency, commonly 40–400 MHz · extinction ratio of 1,000:1 or better · handles far more optical power than an electro-optic modulator.

Pulse picking in amplified ultrafast systems; laser cooling and trapping setups, where AOMs set the exact detuning of every beam; power stabilization loops in metrology; acousto-optic tunable filters used as electronically selectable bandpass elements.

An AOM has no expensive physics in it, and the price mostly reflects how few get made. The parts are a crystal, a bonded piezoelectric transducer, and an impedance-matching network, made in small batches for laser labs and machining systems rather than in the volumes that make a diode cheap. Budget the radio-frequency driver as a second line item, since it often costs about as much as the head, and a laser cooling setup running a dozen AOMs usually spends more on RF electronics than on the optics they drive. The competitive question is what else could do the job: if millisecond switching is enough, a mechanical shutter costs one or two orders less and adds no loss, and above about 10 MHz an electro-optic modulator is the only option. So the AOM's market is the middle of that range, where the beam is powerful enough to put an electro-optic crystal at risk, the switching is in tens to hundreds of nanoseconds, and the frequency shift is wanted rather than merely tolerated.

14 minModulation Of Light | Acousto Optic Modulation | Optoelectronics Devices And SystemsENGINEERING TUTORIAL · 10k+ views

5 minAcousto-Optic Modulation for Sound TransmissionTMU Physics · 10k+ views

20 minHomemade Acousto-optic controller for Under $100!Les' Lab · 10k+ views

Acousto-optic Modulators (RP Photonics Encyclopedia) · Acousto-Optic Modulation, Application Note AN0510 (Isomet)

An electro-optic modulator exploits the Pockels effect: in crystals such as lithium niobate the refractive index changes linearly with applied electric field. Put that crystal in one arm of a Mach-Zehnder interferometer and a voltage becomes a phase shift, which the interferometer converts into an amplitude change. This is how essentially every long-haul optical transmitter encodes data. Thin-film lithium niobate, where the guiding layer is submicron rather than diffused, has recently pushed bandwidths past 100 GHz on a chip.

Speed is the reason it exists: modulation to tens of gigahertz is routine and beyond 100 GHz is demonstrated, orders of magnitude faster than acousto-optic. The response is linear in field and the device introduces no frequency shift. The costs are drive voltage and material. Bulk modulators need tens to hundreds of volts for a half-wave shift, which means a real driver amplifier, they are polarization-sensitive and temperature-sensitive, and the bias point of a Mach-Zehnder drifts and needs a control loop. Optical power handling is well below an AOM's.

Use an electro-optic modulator whenever the modulation rate is above about 10 MHz, which covers all telecom and datacom transmitters, phase locking, and sideband generation for laser stabilization. Bulk Pockels cells remain the standard for Q-switching and pulse slicing at high energy. Use an AOM instead when the beam is powerful, the switching contrast has to be near-total, or a frequency shift is wanted. On a photonic chip, weigh thin-film lithium niobate against silicon modulators: better performance against a less mature supply chain.

Bandwidths of tens of GHz routinely, and above 100 GHz in thin-film lithium niobate · half-wave voltage of a few volts on chip and tens to hundreds of volts in bulk devices · no frequency shift, unlike an acousto-optic modulator · Mach-Zehnder bias drifts and normally needs a control loop · optical power handling well below an AOM's.

Mach-Zehnder modulators inside coherent transceivers on every long-haul route; Pockels cells Q-switching Nd:YAG lasers; phase modulators generating sidebands for Pound-Drever-Hall laser locking; thin-film lithium niobate devices now entering commercial transceivers.

Telecom volume pays for this part. Mach-Zehnder modulators at 1,310 and 1,550 nm are built in numbers no other electro-optic device approaches, and the same physics at 1,064 nm or in the visible costs several times more because nobody is making millions of them. Within the telecom part, fiber pigtailing, RF packaging, and the alignment they require take a large share of unit cost, which is the usual pattern once light has to get into and out of a waveguide. The driver is the other line item, and it is why half-wave voltage matters commercially as well as technically: a few volts on chip can be driven straight from CMOS, while the tens to hundreds of volts a bulk device needs means paying for a dedicated amplifier. That is the real case for thin-film lithium niobate, lower drive voltage and more bandwidth per lane, weighed against a supply chain with few wafer suppliers and little foundry capacity. If that capacity arrives, the buyers are the transceiver makers who need more bandwidth per lane than a silicon modulator can give them; if it does not, silicon modulators absorb the volume.

8 minMastering Lithium Niobate Modulators: Essential Design & TheoryOzen Engineering, Inc - A Member of SimuTech Group · 1k+ views

1 minPhotonic Integrated Circuits - Mach-Zehnder ModulatorAIM Photonics Academy · 10k+ views

4 minNTT Research: Advancing Thin Film Lithium NiobateNextGenInfra · 1k+ views

Integrated lithium niobate electro-optic modulators: when performance meets scalability (Optica) · High-Speed Electro-Optic Modulators Based on Thin-Film Lithium Niobate (Nanomaterials via PMC)

A spatial light modulator imposes a pattern on a beam's amplitude or phase across its cross-section rather than in time. The dominant technology is liquid crystal on silicon: a liquid crystal layer over a CMOS backplane, where the voltage at each pixel rotates the local molecules and changes the optical path length. Load a computed hologram onto the device and the far field becomes whatever pattern you designed. Digital micromirror devices do the amplitude version by tilting millions of tiny mirrors.

Full programmable control of the wavefront is a capability nothing else offers: one device can split a beam into hundreds of traps, correct aberrations in a microscope, or shape a machining spot into a ring. Phase-only modulation wastes no light in principle. The costs are speed and efficiency. Liquid crystal devices update at 60–500 Hz, far slower than a scanner, they are strongly polarization-dependent, and diffraction efficiency into the desired order is typically 60–80% with the rest going into unwanted orders and a zero-order spot that has to be blocked.

Use an SLM when the wavefront itself is the thing you want to control: holographic optical tweezers, adaptive optics in microscopy, multi-spot laser processing, and structured illumination. Use a galvo or MEMS scanner instead when you need one spot moved quickly, because scanners are orders of magnitude faster. Budget optical power carefully, since liquid crystal devices have modest damage thresholds, and plan for the zero-order beam in the layout rather than discovering it during alignment.

Frame rates typically 60–500 Hz for liquid crystal devices, kilohertz for digital micromirror arrays · resolution commonly 1,920 by 1,080 pixels or better, with pixel pitch of a few microns · diffraction efficiency 60–80% into the designed order · phase modulation of at least 2 pi at the design wavelength · strongly polarization-dependent, so the input must be conditioned.

Holographic optical tweezers manipulating many particles at once; adaptive optics correcting sample-induced aberrations in deep-tissue microscopy; multi-beam laser processing splitting one source into dozens of spots; digital micromirror devices in projectors and in maskless lithography.

The panel inside a scientific SLM is close to the panel inside a projector, and most of the price difference is volume and calibration. Consumer LCOS and micromirror engines are made by the million for projectors and headsets, while a phase-only SLM sold to a laboratory is a low-volume variant with a thicker liquid crystal layer, a per-device wavefront calibration, and support from the vendor, which is what puts it in the thousands rather than the tens of dollars. If an application can live with an off-the-shelf projector engine, that substitution is the largest saving available. Two costs beyond the purchase price are worth budgeting. Diffraction efficiency of 60–80% means the rest of the laser power is thrown away, so at high power the buyer pays again in source cost, and the 60–500 Hz update rate caps throughput, so in laser processing an SLM is competing against simply buying more galvo scanners and makes sense only when the job needs many spots at once rather than one spot moved quickly.

3 minLCOS Spatial Light Modulator working principleHOLOEYE Photonics AG · 10k+ views

23 minEXULUS Spatial Light Modulators – Principles and ApplicationsThorlabs · 10k+ views

5 minPolarization Dependence of Phase-Only Spatial Light Modulators (SLM) | Thorlabs Insights Topic FocusThorlabs · 10k+ views

Fringe-Based Structured-Light 3D Reconstruction: Principles, Projection Technologies, and Deep Learning Integration (Sensors via PMC) · Metasurface-enabled polarization-independent LCoS spatial light modulator for 4K resolution and beyond (Light: Science and Applications via PMC)

A beam scanner points a beam somewhere. Galvanometer scanners mount a mirror on a limited-rotation motor with a position sensor and a servo loop; two of them in series give X and Y, and an F-theta lens turns angle into a flat focal plane. Polygon scanners spin a many-faceted mirror continuously for very fast unidirectional sweeps. MEMS mirrors do the same job on a chip, electrostatically or magnetically actuated, in a package small enough for a consumer device. Fast steering mirrors trade range for precision and stabilize a beam rather than sweeping it.

Galvos are the workhorse: several kilohertz small-signal bandwidth, microradian repeatability, and enough aperture for kilowatt beams. Polygons reach far higher line rates. MEMS mirrors are tiny and cheap, which is what makes automotive LiDAR and pico-projectors possible. The trade-offs run along the same axis in each case: aperture against speed. A large mirror handles a big beam and moves slowly; a small one moves fast and clips a large beam. MEMS devices in particular have apertures of a few millimeters, which caps the range of a LiDAR built around them.

Use galvo scanners for laser marking, cutting, welding, and confocal microscopy, where a few kilohertz and a centimeter-scale aperture suit the job. Use a polygon when line rate dominates, as in high-speed printing and some inspection. Use MEMS where size and cost dominate and the beam is small. Use a fast steering mirror when the task is holding a beam still against disturbance rather than moving it. If neither speed nor aperture can be compromised, consider splitting the beam and using several scanners.

Galvo small-signal bandwidth of a few kHz, with step-and-settle in the hundreds of microseconds · repeatability in microradians · aperture from about 3 to 50 mm, trading against speed · MEMS mirrors typically 1–5 mm aperture with resonant scan rates of tens of kHz · polygon scanners reach line rates far above galvos in one direction only.

Galvo heads on every laser marking and cutting system; polygon scanners in laser printers and web inspection; MEMS mirrors in automotive LiDAR and pico-projectors; fast steering mirrors stabilizing beams in free-space optical links and telescopes.

In an industrial system the scanner is rarely the expensive part. A galvo pair and an F-theta lens sit in the low thousands next to a fiber laser at $10k–500k, so the head is a small share of a marking or cutting machine and the decision is about throughput rather than purchase price: a few kilohertz of bandwidth and step-and-settle in the hundreds of microseconds set parts per hour on a machine whose capital cost is mostly the laser. For a job shop, that usually makes the faster head worth buying. Within galvos, price climbs with aperture, since a bigger mirror needs a bigger motor and a stiffer servo to hold the same bandwidth, so specifying a beam larger than the process needs raises the cost of the head and the F-theta lens together. MEMS mirrors are the opposite case, made lithographically in wafer volume and cheap enough for a consumer product, which is the only reason pico-projectors and low-cost automotive LiDAR exist. That cheapness has a price elsewhere in the system: a 1–5 mm aperture caps how much return light the receiver collects, so the saving comes back as shorter range or as a more sensitive detector.

5 minLaser Beam Stabilization with Fast Steering MirrorsNewport Corp by MKS · 10k+ views

4 minM3-RS-U Developer's Kits for Beam Steering and Mirror PositioningNew Scale Technologies Inc · 10k+ views

3 minFast Steering Mirror SetupDon Henderson · 1k+ views

MEMS Micromirror Actuation Techniques: A Comprehensive Review of Trends, Innovations, and Future Prospects (Micromachines) · MEMS Mirrors for LiDAR: A Review (Micromachines via PMC)

Polarization is the orientation of the light's electric field, and a small set of components controls it. Polarizers pass one orientation and reject the other. Waveplates retard one axis relative to the other, turning linear into circular polarization or rotating linear by a chosen angle. Faraday rotators use a magnetic field in a glass to rotate polarization in a direction that does not reverse when the light comes back, which is what makes an optical isolator possible: a one-way valve for light.

Isolators are what stop a laser from being destabilized or destroyed by its own reflected light, which in high-power and single-frequency systems is not optional. Polarization control also makes beam combining and switching possible without loss, since two polarizations can share a path and be separated cleanly. The costs are wavelength and power. Waveplates are wavelength-specific and temperature-sensitive, isolators need a strong permanent magnet and a Faraday material with real absorption, and at high power that absorption causes thermal lensing and depolarization inside the isolator itself.

Put an isolator immediately after any laser whose output can see a reflective target, which includes essentially all fiber lasers, single-frequency sources, and amplifier chains. Use polarizing beamsplitters and waveplates whenever you need lossless combining or switching, and to build variable attenuators that do not change beam pointing. Specify components at the working wavelength rather than a nearby one, and for high-power isolators ask about thermal depolarization at your average power, not just the small-signal isolation.

Isolation typically 30–40 dB for a single stage, 50–60 dB for a double stage · insertion loss around 0.5–1 dB · waveplates are specified at one wavelength and drift with temperature · Faraday rotators need a strong permanent magnet, which sets the size · high average power causes thermal depolarization and reduces isolation.

Isolators protecting every fiber laser and telecom transmitter; polarizing beamsplitters combining two diode bars into one beam; quarter-wave plates converting to circular polarization so that reflected light returns in the orthogonal state; polarization-maintaining fiber in interferometric sensors.

An isolator is usually the cheapest insurance in a laser system. A part in the hundreds of dollars protects a source costing $10k–500k from its own back-reflection, which is why essentially every fiber laser and single-frequency system ships with one. Cost inside the isolator sits in two places, the Faraday crystal and the permanent magnet, and both scale badly with power: at high average power the crystal absorbs enough to lens and depolarize, so holding 30–40 dB of isolation forces a larger aperture, better material, and thermal management, and price climbs faster than the power rating. Wavelength is the other driver, and it is the familiar one, since isolators and waveplates at 1,550 nm are made in telecom volume and are commodity parts while the same function at 1,064 nm or in the visible comes from far smaller runs. Waveplate cost is mostly polishing labor, because a true zero-order plate is a very thin piece of quartz held to a fraction of a wavelength, which is why multi-order and compound versions exist at all.

5 minOptics: Optical isolator | MIT Video Demonstrations in Lasers and OpticsMIT OpenCourseWare · 10k+ views

6 minFaraday Isolator - TDTR Short CourseJoseph Feser · 1k+ views

5 minOptics. Faraday effectPhysics with Andrés Aragoneses · 5k+ views

Selection of Magneto-Optical Material for a Faraday Isolator Operating in High-Power Laser Radiation (Magnetochemistry) · Recent progress in metasurface-enabled optical waveplates (Nanophotonics via PMC)

A photodiode is a reverse-biased junction in which an absorbed photon creates an electron-hole pair that the field sweeps out as current. The p-i-n structure adds an intrinsic region to widen the absorbing depletion layer and lower capacitance, which is what makes fast devices possible. Material sets the band: silicon covers 200–1,100 nm, InGaAs covers 900–1,700 nm and is the telecom detector, germanium spans similar ground more cheaply and noisily, and extended-InGaAs reaches 2.6 um.

Photodiodes are linear over many decades, quantum efficiency reaches 80–95% with an antireflection coating, they need only a bias and a transimpedance amplifier, and they cost from cents to tens of dollars. Bandwidths of tens of gigahertz are commercial. The limits are gain and noise. A photodiode has no internal gain, so at very low light the amplifier's noise dominates and the device is useless below roughly a thousand photons per measurement. Dark current doubles roughly every 10 °C, which matters for InGaAs and germanium in warm environments.

Use a plain photodiode wherever there is enough light, which is most power monitoring, optical communications receivers above a certain level, machine vision, and consumer sensing. It is the default detector and everything else is a response to its limitations. Move to an avalanche photodiode when signal is a few tens of photons and bandwidth still matters, to a photomultiplier or SPAD when you are counting individual photons, and to a thermal detector when the wavelength is beyond what any semiconductor absorbs.

Silicon responds 200–1,100 nm, InGaAs 900–1,700 nm, extended InGaAs to 2.6 um · quantum efficiency 80–95% with a good coating · bandwidths to tens of GHz in p-i-n devices · no internal gain, so amplifier noise sets the floor · dark current roughly doubles per 10 °C rise.

InGaAs photodiodes in every fiber-optic receiver; silicon photodiodes as power monitors on laser systems and in cameras' exposure control; quadrant photodiodes in beam-position servos; large-area devices in pulse oximeters and other consumer medical sensors.

Silicon is the reason photodiodes are cheap. A silicon detector is made by ordinary silicon processing in very high volume and sells for cents, while an InGaAs detector covering 900–1,700 nm is grown on small indium phosphide wafers in much lower volume and sells for tens of dollars to do the same job. That step at 1,100 nm is the largest single cost decision in a detection chain, and it is worth checking whether the measurement can be moved below it before accepting it. Past the die the money moves to the circuit, since a transimpedance amplifier, its board, and the shielding around it commonly cost more than the diode, and at low signal the noise of that front end rather than the detector sets what the system can measure. The exception is telecom, where InGaAs receivers at 1,310 and 1,550 nm are made in enough volume that a packaged photodiode there is far cheaper than a catalog InGaAs detector of the same area.

12 minPhotodiodes (working & why it's reverse biased) | Class 12 | Semiconductors | Physics | Khan AcademyKhan Academy India - English · 100k+ views

7 minWhat is Photodiode? | Explained its Working and ApplicationAutomatedo · 50k+ views

3 minConstruction & Working of a Photodiode - Optical DevicesEzEd Channel · 100k+ views

Reduction in Dark Current in Photodiodes: A Review (Micromachines via PMC) · Infrared avalanche photodiodes from bulk to 2D materials (Light: Science and Applications via PMC)

An avalanche photodiode is a photodiode run at high reverse bias, where a photogenerated carrier is accelerated hard enough to knock more carriers loose, which do the same in turn. The result is internal gain of 10–100 before the signal ever reaches an amplifier, so the amplifier's noise is no longer what limits sensitivity. The price is that the multiplication is a random process, and its randomness appears as excess noise that grows with gain. There is therefore an optimum gain for any given system, not a maximum.

Internal gain buys 10–20 dB of sensitivity over a p-i-n diode at the same bandwidth, which is what makes long-range LiDAR and long-span optical receivers work. Response stays fast, into the gigahertz. The costs are bias and temperature. APDs need 100–400 V depending on material, gain varies sharply with both bias and temperature so a compensation loop is needed, and excess noise means gain past the optimum makes the measurement worse rather than better. Silicon APDs have much better excess noise behavior than InGaAs ones.

Use an APD when the signal is weak but not down to individual photons, and bandwidth still matters: automotive and industrial LiDAR, long-span fiber receivers, scintillation detection, and rangefinding. Use a plain photodiode if there is enough light, since it is simpler and cheaper. Use a SPAD or photomultiplier if you are counting single photons, where an APD's excess noise makes it the wrong tool. Always design in temperature compensation of the bias, because uncompensated gain drift will swamp the sensitivity you bought.

Internal gain typically 10–100, giving 10–20 dB sensitivity improvement over a p-i-n · bias of 100–400 V depending on material · excess noise factor rises with gain, so there is an optimum rather than a maximum · gain varies several percent per degree C, requiring compensation · unit cost tens to hundreds of dollars.

APD receivers in long-reach fiber links; automotive LiDAR detectors at 905 nm; silicon photomultipliers, which are arrays of Geiger-mode APDs, in medical imaging and particle physics; laser rangefinders in surveying and defense.

The die inside an APD is not much more expensive to make than a photodiode's, and most of the price difference is selection and the electronics around it. Devices are binned for breakdown voltage and dark current, and the system has to supply 100–400 V and a temperature compensation loop, so the tens to hundreds of dollars on the datasheet understates what an APD adds to a board. Size that spend against the alternatives that buy the same 10–20 dB: more laser power, a larger receive aperture, or a lower data rate. On a long fiber span or a long-range LiDAR, 10–20 dB is usually cheaper to buy here than anywhere else in the link, which is why APDs stayed standard in both; at short range the arithmetic flips and a plain photodiode with a brighter source is the cheaper system. Silicon is where the volume is, pulled along by 905 nm automotive LiDAR and by silicon photomultipliers in medical imaging, while InGaAs APDs carry both the compound-semiconductor cost and worse excess noise, which is part of why long-range sensing at 1,550 nm is expensive.

4 minWhat is Avalanche Photodiode?Fiber Optics For Sale Co. · 50k+ views

14 minAvalanche Photo Diode Explained | Basics, Structure, Working, Gain, Noise, Pros, Cons & ApplicationsEngineering Funda · 1k+ views

Silicon-Based Avalanche Photodiodes: Advancements and Applications in Medical Imaging (Nanomaterials) · Technical Note: Si APD (Hamamatsu Photonics)

Two technologies count individual photons. A single-photon avalanche diode is an APD biased above breakdown, so one absorbed photon triggers a self-sustaining avalanche that produces a large digital pulse; a quench circuit then resets it. A superconducting nanowire single-photon detector is a meander of superconducting wire held near 1 K and biased just below its critical current, where one photon creates a resistive hotspot and a voltage pulse. The two occupy opposite ends of the performance and cost range.

SPADs are compact, work at or near room temperature, cost tens of dollars, and are made in CMOS arrays for direct time-of-flight imaging. Superconducting nanowires reach 90–98% detection efficiency at 1,550 nm with dark counts of a few per second and timing jitter under 20 picoseconds, which is far beyond any semiconductor device. The costs mirror that. SPADs at telecom wavelengths have poor efficiency and high dark counts, and both types have dead time after each detection that limits count rate. A nanowire system needs a cryostat and costs six figures.

Use SPADs where integration and cost dominate: direct time-of-flight LiDAR, fluorescence lifetime imaging, and consumer proximity sensing. Use superconducting nanowires where efficiency and timing at 1,550 nm decide the experiment, which means quantum key distribution, quantum networking, and photon-starved astronomy, and accept the cryostat. Check dead time against the expected count rate in both cases, since a detector that saturates gives a nonlinear answer rather than a noisy one.

SPAD detection efficiency 30–50% in the visible, much lower at 1,550 nm · superconducting nanowires reach 90–98% at 1,550 nm with jitter under 20 ps · nanowire dark counts of a few per second against thousands for a SPAD · dead time of tens of nanoseconds limits count rate · nanowire systems need cooling to around 1 K and cost well into six figures.

SPAD arrays in smartphone time-of-flight sensors and in automotive LiDAR; superconducting nanowire detectors in quantum key distribution links and in NASA's deep-space optical communication receivers; fluorescence lifetime imaging systems built on SPAD arrays.

These two technologies sit four orders of magnitude apart on price, and neither number is really about the detector. A SPAD is CMOS, so smartphone proximity sensing and time-of-flight paid for the process development and an array now costs tens of dollars; counting photons is not expensive once the counting happens on a silicon die. A superconducting nanowire is also a cheap piece of film, and the six figures a system costs are the closed-cycle cryostat holding it near 1 K, the bias and readout electronics, and the integration labor. That has a useful consequence for anyone budgeting one: the marginal cost of another channel is far below the cost of the first, which is why quantum networking testbeds and deep-space optical receivers are built with many channels rather than one. The economics improve only if cryocoolers get cheaper, so compare cooler options and channel count before comparing nanowire specifications. If the photons are visible rather than at 1,550 nm, a SPAD does the job for tens of dollars and the question does not arise.

12 minThe Basics of Single Photon Avalanche DiodesKevin Zhang · 10k+ views

14 minSuperconducting Nanowire Single Photon DetectorsErick Elizalde · under 1k views

5 minSingle Photon Avalanche Diode (SPAD) – Complete Explanation | Principle, Construction & WorkingGHS · under 1k views

Superconducting nanowire single-photon detectors: physics and applications (arXiv) · Superconducting Nanowire Single-Photon Detectors for Quantum Information (arXiv)

A photomultiplier tube is a vacuum device. A photon strikes a photocathode and ejects an electron, which is accelerated onto a dynode where it knocks out several more, and that repeats through 8–12 stages to produce a gain of a million or more. The output is a current pulse large enough to see without a low-noise amplifier. The technology is older than the transistor and remains the standard where a large sensitive area is needed in the ultraviolet and visible.

Gain of 10^6 with very low noise means a PMT resolves single photons over an active area of many square centimeters, which no semiconductor matches at that size. Response is fast, in the nanoseconds. The weaknesses are physical and spectral. PMTs need 1–2 kV, they are glass vacuum tubes that break and that are sensitive to magnetic fields, and the photocathode's response dies beyond about 900 nm, so they are useless in the infrared. Exposure to room light while powered can destroy them, which makes them unforgiving in the field.

Use a PMT where you need single-photon sensitivity over a large area in the ultraviolet or visible: scintillation counting, flow cytometry, confocal microscopy, and large physics detectors. Use a silicon photomultiplier instead when the environment is rough, a magnetic field is present, or low voltage and small size matter, since it now matches PMT gain in a solid-state package. Use an InGaAs APD or superconducting detector beyond 900 nm, where a PMT simply does not respond.

Gain of 10^6 to 10^7 across 8–12 dynode stages · quantum efficiency 20–40% in the visible, falling to nothing beyond about 900 nm · active areas from millimeters to over 20 inches in the largest tubes · needs 1–2 kV bias · vulnerable to magnetic fields and to light exposure while powered.

Photomultipliers in gamma cameras and positron emission tomography scanners; the 20-inch tubes lining Super-Kamiokande's neutrino detector; detection channels in confocal microscopes and flow cytometers; silicon photomultipliers steadily replacing them in medical imaging.

A photomultiplier is a hand-assembled glass vacuum tube, so there is no semiconductor learning curve underneath its price and it has been roughly flat for decades. The number that matters is cost per unit of sensitive area rather than cost per device: at $100–1k for a tube covering many square centimeters, and up to 20 inches across in the largest ones, nothing else is close when the job is watching a large area in the ultraviolet or visible. Silicon photomultipliers are on the opposite curve, falling with wafer volume, which is why they displaced tubes in the small-area applications first and why the large physics detectors are still full of glass. Budget for what the tube needs around it, too, since 1–2 kV bias, magnetic shielding, and the risk of destroying it by switching on the room lights all cost money the unit price does not show. A good rule of thumb: if a design calls for many small PMTs rather than a few large ones, quote it with silicon photomultipliers before committing.

8 minWhat is a Photomultiplier Tube? [L19]Molecular Imaging & Therapy · 10k+ views

1 minPhotomultiplier TubeJefferson Lab · 50k+ views

4 minPhototubes and PhotomultipliersPeter Tusar [WEB3] · 50k+ views

Photomultipliers (RP Photonics Encyclopedia) · Photomultiplier Tubes: Basics and Applications, Fourth Edition (Hamamatsu Photonics)

An image sensor is a two-dimensional array of photodiodes with readout electronics. CCDs shift accumulated charge across the chip to one amplifier, giving very uniform response; CMOS sensors put an amplifier in every pixel and read them out in parallel, which is faster, cheaper, and lower power. CMOS has won nearly everywhere. Modern devices are backside-illuminated so light does not pass through the wiring, and stacked, with the pixel array bonded onto a separate logic die.

The economics are extraordinary: a sensor with tens of millions of pixels costs a few dollars because phone volumes pay for the fab. Quantum efficiency above 80% and read noise around one electron make scientific CMOS competitive with cooled CCDs for most work. The limits are material and dynamic range. Silicon stops absorbing beyond about 1,100 nm, so short-wave infrared needs InGaAs arrays that cost a thousand times more per pixel. Rolling shutter distorts fast motion unless the sensor has a global shutter, and dark current still forces cooling for long exposures.

Use a CMOS sensor for essentially all imaging in the ultraviolet through near-infrared. Choose scientific CMOS when read noise and speed matter together, and a cooled deep-well CCD only for very long exposures where its uniformity still wins. Specify global shutter if the scene or the camera moves quickly. Go to InGaAs only when the wavelength genuinely requires it, since the cost step is enormous, and consider whether an illuminator at a silicon-friendly wavelength would solve the problem instead.

Silicon responds 200–1,100 nm; InGaAs arrays extend to 1,700 nm at roughly a thousand times the cost per pixel · quantum efficiency above 80% in backside-illuminated devices · read noise around 1 electron in scientific CMOS · pixel pitch from under 1 um in phones to 10–20 um in scientific sensors · dark current halves roughly every 6–8 °C of cooling.

Sony's stacked backside-illuminated sensors in most smartphones; scientific CMOS cameras that displaced electron-multiplying CCDs in microscopy; InGaAs arrays for short-wave infrared inspection and surveillance; event cameras, which report per-pixel changes instead of frames.

Phone volume pays for everything else. A sensor with tens of millions of pixels sells for a few dollars because the fab, the process, and the stacking technology were all developed for the phone market, and scientific, machine-vision, and automotive cameras are built on processes somebody else's volume already paid to develop. The cost cliff is at 1,100 nm. InGaAs arrays run roughly a thousand times more per pixel, partly because indium phosphide has nothing like silicon's manufacturing base, and partly because the array has to be bonded to a separate readout chip instead of being made monolithically. That ratio is large enough that the first question on any short-wave infrared design should be whether an illuminator at a silicon-friendly wavelength solves the problem, since changing the wavelength is usually cheaper than changing the detector material. For a company, competing on volume silicon sensors is not realistic, so the value sits in the SWIR gap, in the optics, or in what gets done with the image.

9 minImage Sensors Explained: How CCD and CMOS Sensors works? CCD vs CMOSALL ABOUT ELECTRONICS · 100k+ views

18 minThe Chips That See: Rise of the Image SensorAsianometry · 100k+ views

15 min#743 Basics: How Image Sensors WorkIMSAI Guy · 10k+ views

CMOS Image Sensors for High Speed Applications (Sensors via PMC) · A Review of the Pinned Photodiode for CCD and CMOS Image Sensors (Dartmouth Digital Commons)

A thermal detector does not care about photon energy; it absorbs radiation, warms up, and reports the temperature change. A microbolometer array is the common form: each pixel is a thin membrane of vanadium oxide or amorphous silicon suspended on legs above the readout circuit, so it is thermally isolated and its resistance changes as it heats. Because the mechanism is heat rather than band-to-band absorption, one device responds across the whole infrared, including 8–14 um where photon detectors need cryogenic cooling.

Uncooled operation is what made thermal imaging ordinary: no cryocooler means a camera that costs hundreds rather than tens of thousands, runs on a battery, and starts instantly. Response is flat over a wide band. The costs are speed and sensitivity. Thermal time constants of 5–15 ms cap frame rates near 60 Hz and blur fast motion, and noise-equivalent temperature difference of 30–50 mK is well behind a cooled photon detector's few mK. Germanium optics, needed at these wavelengths, often cost more than the sensor.

Use an uncooled microbolometer for almost all thermal imaging: building inspection, firefighting, automotive night vision, drone payloads, and process monitoring. Move to a cooled photon detector, mercury cadmium telluride or indium antimonide, when you need to resolve small temperature differences, image fast events, or do quantitative mid-wave radiometry, and accept the cryocooler and the cost. Remember that the lens may dominate the bill of materials, which changes how much sensor performance is worth buying.

Uncooled microbolometers respond across 8–14 um without cooling · thermal time constant 5–15 ms, capping frame rate near 60 Hz · noise-equivalent temperature difference typically 30–50 mK, against a few mK for cooled photon detectors · common array formats 320 by 240 up to 1,280 by 1,024 · germanium optics often cost more than the detector.

FLIR and Teledyne uncooled cores in handheld and drone thermal cameras; automotive night vision systems; cooled mercury cadmium telluride detectors in missile seekers and scientific instruments; pyroelectric detectors as cheap motion sensors and laser power meters.

Removing the cryocooler is what took thermal imaging from tens of thousands of dollars to hundreds, and once it was gone the detector stopped being the expensive part. Germanium optics often cost more than the sensor, and germanium metal roughly tripled in price between 2021 and 2025 after China put it under export licensing, so the item that dominates the bill of materials is the one with no volume curve underneath it. Resolution costs twice over: going from 320 by 240 to 1,280 by 1,024 is about seventeen times the pixels and also a larger optical format, which means a larger germanium lens. That is why cost reduction in this business tends to come from wafer-level packaging and from requalifying onto chalcogenide glasses, rather than from improving a noise-equivalent temperature difference that is already adequate at 30–50 mK for building inspection, firefighting, and drone work. Cooled mercury cadmium telluride sits at a different price point for structural reasons: the cryocooler is expensive, it wears out, and export control narrows the supplier list before price does.

4 minAsk an Expert: What is a Microbolometer?Sierra-Olympia Technologies · 10k+ views

15 min#1486 Micro Bolometer Thermal Camera SensorIMSAI Guy · 5k+ views

5 minTech Talk: MWIR & LWIR OverviewSierra-Olympia Technologies · 10k+ views

Readout IC Architectures and Strategies for Uncooled Micro-Bolometers Infrared Focal Plane Arrays: A Review (Sensors via PMC) · Low-Cost Microbolometer Type Infrared Detectors (Micromachines)

Optical microscopy resolves small things using visible light, and diffraction sets the limit at roughly half a wavelength divided by numerical aperture, about 200 nm in practice. Fluorescence microscopy labels specific molecules with dyes and images only their emission, which turns a transparent cell into a map of one protein. Confocal microscopy adds a pinhole that rejects light from outside the focal plane, giving optical sections. Multiphoton microscopy uses an ultrafast laser so that only the focal volume gets enough intensity to excite, which allows imaging deep in tissue.

Nothing else images living samples at this resolution with molecular specificity, and the labeling chemistry is what makes it powerful rather than the optics. Confocal sectioning removes out-of-focus haze; multiphoton reaches hundreds of microns into scattering tissue. The costs are photons and time. Fluorophores bleach and cells are damaged by illumination, so every experiment trades signal against sample health, and point-scanning methods are slow because they visit each voxel in turn. Systems run from tens of thousands to well over half a million dollars.

Use widefield fluorescence for thin samples and speed, confocal when out-of-plane light is spoiling the image, and multiphoton when imaging deep in scattering tissue or when photodamage outside the focus must be avoided. Consider light-sheet microscopy for large living samples where total dose is the constraint. Go to super-resolution methods only when the question genuinely needs sub-200 nm detail, since they cost throughput and add reconstruction assumptions. For hard surfaces, interferometric and confocal profilometry answer height questions better than any fluorescence method.

Diffraction limits lateral resolution to roughly 200 nm in the visible · confocal pinhole gives optical sections around 500–800 nm thick · multiphoton reaches several hundred microns into scattering tissue · super-resolution methods reach 20–50 nm at the cost of speed and labeling constraints · systems from tens of thousands to over half a million dollars.

Confocal systems from Zeiss, Leica, and Nikon as laboratory standards; two-photon microscopy in neuroscience imaging of live brain tissue; light-sheet systems for developmental biology; the 2014 Nobel Prize in Chemistry for super-resolution fluorescence microscopy.

The price of a research microscope is set mostly by two items, the objectives and the light source, and which one dominates depends on the method. A widefield or confocal system is priced by its objectives and scan head; a multiphoton system is priced by its ultrafast laser, which at $50k–500k can cost more than everything else on the table. That is the whole reason the same optical question is answered by an instrument in the tens of thousands or by one over half a million, depending on how deep into scattering tissue it has to look. The other half of the economics is utilization. These are capital purchases carrying service contracts, and an instrument booked all day in a shared core facility costs a small fraction per experiment of the identical instrument sitting in one lab and used a few hours a week. If you are specifying a system, work out how many hours a week it will actually run before arguing about resolution, because utilization usually moves cost per result more than any specification does.

10 minWidefield and Confocal Fluorescence MicroscopyProfessor Dave Explains · 50k+ views

4 minConfocal MicroscopyQuick Biochemistry Basics · 100k+ views

2 minFluorescence Microscopy Animationmitedustar · 100k+ views

Optical medical imaging (NIST) · The Microscope Optical Train (Nikon MicroscopyU)

Both techniques read molecular vibrations, from opposite directions. Infrared spectroscopy shines broadband infrared through a sample and measures which frequencies are absorbed, which happens when a vibration changes the molecule's dipole moment. Raman illuminates with a single laser line and measures the tiny fraction of light scattered at shifted frequencies, which happens when a vibration changes the molecule's polarizability. The selection rules differ, so the two are complementary: symmetric bonds show up strongly in Raman and weakly in infrared, and vice versa.

Both give a molecular fingerprint without consuming the sample. Raman works through glass and plastic packaging and is almost blind to water, which makes it ideal for aqueous and biological samples and for identifying a substance inside a sealed container. Infrared is far more sensitive per photon and instruments are cheaper. The weaknesses are symmetric. Raman scattering is extraordinarily weak, roughly one photon in ten million, and fluorescence from the sample can bury the signal entirely. Infrared is defeated by water absorption and needs sample preparation or an attenuated-reflection accessory.

Use Raman for aqueous samples, for measurement through packaging, and where no sample preparation is possible: pharmaceutical incoming inspection, hazardous material identification in the field, and polymorph screening. Use Fourier-transform infrared for routine organic identification, thin films, and anywhere the sample is dry and cheap instruments matter. If fluorescence swamps a Raman spectrum, move the excitation to 785 or 1,064 nm before concluding the technique does not work. For trace gases, a quantum cascade laser plus a long path beats both.

Raman scattering yields roughly one shifted photon in 10^7 · common Raman excitation at 532, 633, 785, and 1,064 nm, with longer wavelengths reducing fluorescence · infrared fingerprint region 400–4,000 cm-1 · Raman is weakly sensitive to water, infrared strongly · handheld Raman units cost around $20,000–50,000, benchtop systems considerably more.

Handheld Raman analyzers used by first responders to identify unknown substances through a bag; pharmaceutical raw material verification at goods receipt; Fourier-transform infrared as the standard organic identification tool in every analytical lab; Raman microscopy mapping composition across a sample.

Cost follows the weakness of the signal. Raman returns roughly one shifted photon in 10^7, so the instrument has to combine a wavelength-stabilized laser, filters that reject the laser line by many orders of magnitude, and a low-noise detector, and every one of those is a specialty part bought in small volume. Infrared has no such problem: a Fourier-transform instrument is a Michelson interferometer with a moving mirror, the design has been competed on for decades, and that is why FTIR sits in every analytical lab while Raman does not. On a handheld Raman unit at $20,000–50,000, the hardware is only part of what the buyer is paying for. A first responder is buying an identification, which means the spectral library, the matching software, and the validation behind them, and those take years to build rather than a bill of materials. If you are evaluating a Raman company, ask what share of the price is the box and what share is the library, because the box gets cheaper and the library does not get copied.

7 minRaman Basics | Principles of Raman Spectroscopy | 7 Minute TutorialBruker · 100k+ views

4 minWhat is Raman Spectroscopy?METTLER TOLEDO AutoChem · 100k+ views

6 minRaman BasicsThermo Scientific Spectroscopy & Materials Analysis · 100k+ views

Raman spectroscopy for in-line water quality monitoring: instrumentation and potential (Sensors via PMC) · Vibrational Spectroscopy (Chemistry LibreTexts)

A normal camera records three broad color channels. A hyperspectral imager records a full spectrum at every pixel, typically 100–300 contiguous bands, producing a data cube with two spatial dimensions and one spectral. Most instruments build the cube by scanning: a pushbroom sensor images one line at a time through a slit and a grating while the platform moves, which suits aircraft and conveyor belts. Snapshot designs capture the whole cube at once at lower resolution.

Materials that look identical often have very different spectra, so hyperspectral imaging separates things no color camera can: crop stress before it is visible, plastic types on a sorting line, mineral species from the air, bruised fruit. It is non-contact and needs no sample preparation. The costs are data and light. A single flight produces terabytes, and the analysis, not the sensor, is usually the hard part. Splitting light into hundreds of bands means each one gets little, so the technique needs bright illumination or long integration, and calibration against a reference is essential.

Use hyperspectral imaging when the distinction you need is spectral rather than spatial and a handful of filters will not separate the classes. Agriculture, mineral exploration, recycling, and food inspection are the established markets. Try multispectral first, with five to ten well-chosen bands: if that separates the classes it is far cheaper, faster, and easier to deploy. Build the classification pipeline before buying the sensor, because the instrument is rarely the limiting factor and an unused data cube is expensive.

Typically 100–300 contiguous spectral bands per pixel · pushbroom scanning requires relative motion between sensor and scene · data rates of gigabytes per minute, and terabytes per survey · each band receives a small fraction of the light, so illumination and integration time matter · systems from about $25,000 for industrial line-scan to hundreds of thousands for airborne.

Airborne surveys for mineral exploration and pipeline monitoring; Pixxel and Planet hyperspectral satellites; plastic sorting lines separating polymer types optically; food inspection systems detecting bruising and foreign material before it is visible.

The sensor is usually the smallest line item. Systems run from about $25,000 for an industrial line-scan camera to hundreds of thousands for an airborne one, but a survey produces terabytes, and the storage, the compute, and a trained classification pipeline generally cost more over a program's life than the instrument did. Multispectral is the competitor that decides most deals: five to ten well-chosen bands are cheaper to buy, cheaper to fly, and far cheaper to process, so if they separate the classes then 200 bands are an expensive way to reach the same answer. That leaves two business models with very different margins. Selling data cubes means competing against a cheaper sensor while leaving the hard part to the customer; selling the answer, meaning polymer type or crop stress or mineral class, is priced against what the decision is worth. If you are evaluating a company here, ask what fraction of revenue comes from the analysis rather than the hardware.

3 minWhat is hyperspectral imaging - TutorialSpecimSpectral · 50k+ views

3 minHyperspectral and Multispectral Imaging - TRENDING IN OPTICSEdmund Optics · 10k+ views

3 minWhat is hyperspectral imaging: use cases, capabilities and benefits?Pixxel · 10k+ views

Modern Trends and Recent Applications of Hyperspectral Imaging: A Review (Technologies) · AVIRIS - Airborne Visible / Infrared Imaging Spectrometer: General Overview (NASA Jet Propulsion Laboratory)

Interferometry splits a beam, sends the halves along different paths, recombines them, and reads the resulting fringe pattern. Because a fringe corresponds to half a wavelength of path difference, the technique measures distance in units of light. A Michelson interferometer measures displacement along one axis; a Fizeau interferometer compares a test surface against a reference flat and maps its shape; a Fabry-Perot uses many round trips between two mirrors to resolve wavelengths finely.

Sub-nanometer sensitivity from a table-top instrument is the reason interferometry underlies dimensional metrology, optical fabrication, and precision positioning. It is traceable, because the wavelength of a stabilized laser is a known standard. The weaknesses are environmental and interpretive. Air temperature, pressure, and humidity change the refractive index, so an uncompensated meter-long path drifts by hundreds of nanometers over a normal day, and vibration washes fringes out. Phase measurements are also ambiguous by whole wavelengths, so tracking motion requires counting fringes continuously.

Use interferometry when the measurement has to be traceable and better than a micron: stage position feedback in lithography and machine tools, surface figure testing of optics, flatness and film thickness. Use white-light or low-coherence interferometry when absolute distance rather than relative motion is needed, since it removes the fringe ambiguity. Control the environment or measure it and compensate, because on any path longer than a few centimeters the air is part of your instrument.

One fringe corresponds to half a wavelength of path change, about 316 nm at 633 nm · sub-nanometer resolution with phase interpolation · air index changes roughly 1 part per million per degree C, which is 1 um per meter of path · vibration isolation is normally required · instruments from tens of thousands to several hundred thousand dollars.

Zygo and 4D Technology interferometers used throughout optical manufacturing; heterodyne laser interferometers as position feedback in lithography stages; LIGO, a 4 km Michelson interferometer that measures displacements smaller than a proton; Fourier-transform infrared spectrometers, which are Michelson interferometers with a moving mirror.

The instrument is rarely the whole purchase. Air index changes by roughly 1 part per million per degree C, which is 1 um per meter of path, so a meter-scale measurement needs a temperature-controlled room, vibration isolation, and environmental compensation, and that installation can approach the cost of the interferometer itself. Against prices of tens of thousands to several hundred thousand dollars, that is the number a first-time buyer usually gets wrong. Two quite different businesses sell into this. A benchtop Fizeau goes to an optics shop as capital equipment bought once every several years, from a short list of vendors that compete on capability rather than price; heterodyne interferometers designed into lithography stages ship as OEM components in volumes that support serious engineering investment. For an optics manufacturer the right way to book it is as fixed overhead, since a surface that cannot be measured cannot be polished to specification, and the cost spreads over every part the shop makes.

17 minLaser Interferometer - Part 1: The Optical Design.Diffraction Limited · 50k+ views

6 minInterferometer - How We Measure Tiny Physical ChangesNPS Physics · 10k+ views

6 minlaser interferometerDr Joji Thomas · 10k+ views

A Review of Optical Interferometry for High-Precision Length Measurement (Micromachines via PMC) · OPTI 513: Optical Testing (University of Arizona)

Optical coherence tomography is interferometry used as a depth-resolved imaging method. A broadband source has a short coherence length, so interference occurs only when the sample and reference path lengths match to within a few microns. Scanning that match point through the sample builds a depth profile, and scanning laterally builds a cross-sectional image. Modern systems work in the Fourier domain, either by dispersing the interference onto a line camera or by sweeping a laser's wavelength, which recovers the whole depth profile at once and is orders of magnitude faster than the original time-domain approach.

It produces micron-scale cross-sections of living tissue without contact, radiation, or contrast agent, which is why it became the standard of care in ophthalmology within a decade. Axial resolution comes from source bandwidth rather than from the lens, so it is decoupled from lateral resolution. The limit is depth. Scattering confines useful imaging to 1–3 mm in tissue, so it images the retina and the inside of an artery well and cannot see through a limb. Motion during a scan produces artifacts, and highly scattering tissue attenuates quickly.

Use OCT when you need micron-resolution structure a millimeter or two below a surface without touching it: retinal imaging, intravascular imaging, dermatology, and non-destructive inspection of coatings, films, and layered materials. Use ultrasound when you need centimeters of depth and can accept coarser resolution. In industrial work, compare against confocal profilometry, which measures surfaces well but does not see subsurface layers the way OCT does.

Axial resolution of 1–15 um, set by source bandwidth rather than by the objective · imaging depth 1–3 mm in scattering tissue · swept-source systems reach hundreds of thousands of depth scans per second · no contact, no ionizing radiation, no contrast agent · ophthalmic systems commonly $50,000–150,000.

Retinal OCT in nearly every ophthalmology clinic, now the most common imaging procedure in the specialty; intravascular OCT guiding coronary stent placement; OCT-based inspection of pharmaceutical tablet coatings and multilayer films; art conservation imaging of varnish layers.

Most of an OCT system is cheap by the standards of medical imaging. It works in the near-infrared and short-wave infrared, so the fiber, couplers, and detectors come out of the telecom supply chain at telecom prices, and the expensive items are the swept-source laser or the line camera. At $50,000–150,000 an ophthalmic system is a small fraction of what an MRI costs, and a clinic pays for it out of scan volume against a reimbursed procedure, which is why retinal OCT spread through the specialty within about a decade. Industrial OCT uses much the same hardware with none of that pull: the buyer is comparing it against confocal profilometry on a productivity case rather than a billing code, so the sales cycle is longer and the price ceiling is lower. Those are two different businesses that happen to share a physics, and when evaluating an OCT company the first thing to establish is which one it is in.

3 minOptical Coherence Tomography: A New Way of SeeingAlbert and Mary Lasker Foundation · 50k+ views

3 minOptical Coherence Tomography (OCT) Principle DemonstratedStarFish Medical · 10k+ views

5 minOptical Coherence Tomography (OCT) Full System AssemblyThorlabs · 5k+ views

Optical Coherence Tomography (RP Photonics Encyclopedia) · The Development, Commercialization, and Impact of Optical Coherence Tomography (Investigative Ophthalmology & Visual Science)

A photonic integrated circuit puts waveguides, splitters, modulators, filters, and detectors on one chip, the way an electronic integrated circuit puts transistors on one die. Silicon photonics is the dominant platform because it reuses CMOS fabs: silicon guides light at 1,310 and 1,550 nm with high index contrast, so bends can be micron-scale. Silicon cannot emit light, so a laser is bonded on or coupled in from outside. Indium phosphide can make the laser natively but has no equivalent manufacturing base, and silicon nitride offers much lower loss for sensing and visible work.

Wafer-scale manufacturing changes the economics of any optical function that can be integrated, replacing a bench of aligned parts with a die that costs tens of dollars in volume. Density is enormous and alignment stability comes for free. The obstacles are light in and light out. Coupling from a single-mode fiber to a submicron waveguide costs 1–3 dB and demands submicron alignment, packaging is often more than half the cost, and silicon's thermo-optic sensitivity means most devices need heaters and control loops to stay on wavelength.

Use integrated photonics where volume justifies the development: datacenter transceivers, LiDAR beam steering, and biosensing cartridges. Choose silicon for density and CMOS compatibility, indium phosphide when an on-chip laser is essential, and silicon nitride when loss or visible wavelengths matter. Do not integrate a one-off instrument, where discrete components are far cheaper and easier to change. And budget for packaging and test explicitly, because that is where integrated photonics projects usually run over.

Silicon waveguide loss around 1–2 dB/cm; silicon nitride below 0.1 dB/cm · fiber-to-chip coupling loss of 1–3 dB with submicron alignment tolerance · packaging and test often exceed half of unit cost · silicon's index shifts strongly with temperature, so wavelength-critical devices need active control · foundry runs available through shared multi-project wafers.

Intel and Cisco silicon photonics transceivers shipping in volume; Aeva and SiLC frequency-modulated LiDAR built on photonic chips; SMART Photonics and imec offering open foundry access; optical gyroscopes and biosensors implemented in silicon nitride.

The die is the cheap part. A silicon photonic chip costs tens of dollars in volume, and packaging and test often exceed half the unit cost, because getting a single-mode fiber onto a submicron waveguide costs 1–3 dB and takes active alignment done one device at a time. Wafer-scale economics do not reach a step performed serially by a machine and an operator, so the cost curve of an integrated photonics product follows how far the packaging has been automated rather than how small the die is. Multi-project wafer runs make a first chip cheap, which is part of why there are many photonics startups with working devices and few with a shipping product: a demonstration costs a shuttle run, and a product costs a packaging line. Thermal control adds to the bill on every unit, since heaters and their control loops consume power and take test time. If you are evaluating one of these companies, ask for packaged cost at volume and test time per part, not die cost.

16 minSilicon Photonics: The Next Silicon Revolution?Asianometry · 100k+ views

3 minWhat is Silicon Photonics? | Intel BusinessIntel Business · 50k+ views

3 minHow our photonic chips workSMART Photonics · 10k+ views

Ultra-wideband integrated photonic devices on silicon platform: from visible to mid-IR (Nanophotonics via PMC) · Roadmapping the next generation of silicon photonics (Nature Communications)

An optical transceiver is a pluggable module that converts electrical signals to light and back. Standard form factors, SFP, QSFP, QSFP-DD, and OSFP, define the mechanics, the electrical interface, and the power envelope so that any compliant module works in any compliant port. Inside is a laser or an array of them, a modulator or directly modulated source, a receiver, and a digital signal processor. Short-reach modules use directly modulated VCSELs over multimode fiber; long-reach coherent modules use external modulators and elaborate digital signal processing.

Pluggability is the product: an operator can change reach or rate by swapping a module rather than a switch, and competition across a standardized interface has driven cost per gigabit down relentlessly. The constraint is now power. An 800G module dissipates 15–25 W, and with dozens per switch the optics can approach the switch ASIC's own consumption, which is exactly the pressure driving co-packaged optics. Reach, rate, and power trade against each other, and the cheapest module that meets the link budget is usually the right one rather than the best-specified.

Choose the shortest-reach module that closes the link, since reach costs money and power: multimode for tens of meters inside a rack row, single-mode short reach for a few kilometers, coherent for anything longer. In AI clusters, count the optics power budget per rack early rather than late. Keep pluggables where flexibility and multi-vendor supply matter, which is most deployments, and consider co-packaged optics only where power density has genuinely become the binding constraint.

Rates from 10G through 400G and 800G, with 1.6T emerging · an 800G module dissipates roughly 15–25 W · form factors SFP, QSFP, QSFP-DD, and OSFP define mechanics and power envelope · multimode reaches tens to hundreds of meters; single-mode kilometers; coherent hundreds to thousands · cost per gigabit has fallen consistently, though power per bit has fallen more slowly.

QSFP-DD and OSFP 800G modules in current AI cluster fabrics; coherent pluggables at 400G ZR carrying metro traffic without a separate transport box; the Multi-Source Agreements that define each form factor and keep the market multi-vendor.

Standardization sets the margins here. The Multi-Source Agreements fix the mechanics, the electrical interface, and the power envelope so that any compliant module works in any compliant port, which is exactly what operators want and exactly what keeps prices falling; a vendor makes its margin in the window between shipping a new rate and having competitors qualified into the same sockets, and that window is short. Cost per gigabit has fallen consistently, but power per bit has fallen more slowly, and that gap is the story at 800G. A module dissipating 15–25 W, times dozens per switch, turns optics into a power and cooling line item rather than only a capital one, which is the pressure behind co-packaged optics. The buying rule follows from the same arithmetic: take the shortest-reach module that closes the link, because reach costs money once and then costs power for as long as the link is up.

3 minWhy do AI Data Centers Need 400G/800G Optical Transceivers?QSFPTEK · 1k+ views

2 minCoherent | 800G OSFP Pluggable Transceiver Demonstration at OFC 2021Coherent Corp. · 1k+ views

5 minCore Application Scenarios of 800G Optical ModulesQSFPTEK · under 1k views

Fiber-optic Links (RP Photonics Encyclopedia) · 400ZR Implementation Agreement (Optical Internetworking Forum)

Wavelength-division multiplexing runs many independent channels down one fiber, each on its own wavelength, combined by a multiplexer and separated at the far end by a demultiplexer. Coarse WDM spaces channels 20 nm apart and needs no temperature control, so it is cheap. Dense WDM spaces them 50 or 100 GHz apart in the erbium amplifier band around 1,550 nm, fitting 80 or more channels on one fiber, and requires stabilized lasers and temperature-controlled filters. Reconfigurable add-drop multiplexers let an operator route individual wavelengths without touching the traffic on the others.

It multiplies the capacity of fiber already in the ground, which is the cheapest capacity in telecommunications by a wide margin: lighting another wavelength costs a transceiver, while laying another fiber costs a construction project. One erbium amplifier boosts every channel at once. The costs are optical rather than electrical. Channels interact through fiber nonlinearity, so total launch power is capped, dispersion and polarization effects accumulate over distance, and per-channel provisioning needs precise wavelength control and a management system that understands the optical layer.

Use dense WDM on any long-haul or metro route where fiber is scarce or expensive, which is nearly all of them, and coarse WDM on short enterprise links where cost matters more than channel count. In datacenter interconnect, weigh WDM against simply lighting parallel fibers, since inside a campus fiber is often cheap enough that parallelism wins. Plan the optical layer's power and dispersion budget as a whole, because a channel added without re-checking the budget can degrade every other channel on the fiber.

Dense WDM channel spacing 50 or 100 GHz, giving 40–96 channels in the C band · coarse WDM spacing 20 nm, needing no temperature control · one erbium-doped fiber amplifier boosts all channels simultaneously · total launch power is limited by fiber nonlinearity, not by amplifier capability · adding a wavelength costs a transceiver rather than a new cable.

Dense WDM on every long-haul terrestrial and submarine route; coarse WDM in mobile fronthaul and enterprise links; reconfigurable optical add-drop multiplexers in metro rings; 400G ZR coherent pluggables putting WDM transport directly into router ports.

The whole case is one comparison: a transceiver against a construction project. Lighting another wavelength on installed fiber costs a transceiver, while laying another fiber costs permits, trenching, and years, so dense WDM makes sense on any route where the civil work dominates, which is most long-haul and metro. Inside a campus the comparison flips, since pulling more fiber a few hundred meters is cheap and parallel fibers skip the wavelength control entirely. That control is most of the cost difference between the two flavors: coarse WDM at 20 nm spacing needs no temperature control and is cheap enough for enterprise links, while dense WDM at 50 or 100 GHz needs stabilized lasers, temperature-controlled filters, and a management system that understands the optical layer. One erbium amplifier boosting 40–96 channels at once is what makes cost per channel fall as channels are added, so an operator lighting only a handful of wavelengths on a dense WDM line is carrying most of the fixed cost and collecting little of the benefit.

5 minWhat is WDM (Wavelength Division Multiplexer)? - FO4SALE.COMFiber Optics For Sale Co. · 100k+ views

9 minFOA Lecture 31 Wavelength Division Multiplexing (WDM)thefoainc · 50k+ views

2 minDense Wavelength Division Multiplexing (DWDM)Champion ONE · 10k+ views

Wavelength Division Multiplexing (RP Photonics Encyclopedia) · Spectral grids for WDM applications: DWDM frequency grid, G.694.1 (ITU-T)

In a conventional switch, the optics sit in pluggable modules at the faceplate and electrical signals travel across the board to get there. At 800G and above, driving those centimeters of copper costs real power and real signal integrity effort. Co-packaged optics moves the optical engines onto the same substrate as the switch ASIC, so the electrical path shrinks to millimeters. Light arrives and leaves through fiber connected at the package, often supplied by an external laser source so the heat-sensitive laser is not sitting next to a hot ASIC.

Shortening the electrical link cuts interface power by roughly 30–50%, which at rack scale is a meaningful fraction of the switch's budget, and it removes the signal-integrity work that limits how fast pluggable interfaces can go. Density improves because the faceplate is no longer the constraint. The costs are serviceability and supply chain. A failed optical engine can no longer be swapped in seconds; it is part of the package, so the whole assembly is the field-replaceable unit. Yield of a package containing an ASIC and several optical engines is the commercial question, and the ecosystem is young.

Watch it and pilot it if you operate at a scale where optics power per rack is a design constraint, which today means the largest AI fabrics. Everyone else should stay with pluggables, where multi-vendor supply, field replacement, and mature test infrastructure are worth more than the power saving. When evaluating, ask specifically about the failure and repair model and about laser sourcing, because those decide operating cost more than the efficiency number does.

Cuts electrical interface power by roughly 30–50% against front-panel pluggables · electrical path shrinks from centimeters to millimeters · an 800G pluggable dissipates 15–25 W, and a switch carries dozens · the field-replaceable unit becomes the package rather than the module · deployments today are early and concentrated in the largest AI fabrics.

Broadcom's Bailly co-packaged Tomahawk switch; NVIDIA's co-packaged silicon photonics switches announced for AI fabrics; Intel and Ayar Labs optical chiplet work; the Open Compute Project's co-packaged optics workstream defining interfaces.

The trade is capital and yield against operating power. Cutting interface power by 30–50% is worth real money where an 800G pluggable dissipates 15–25 W and a switch carries dozens of them, but the saving has to cover a harder manufacturing problem: a package holding an expensive switch ASIC and several optical engines only ships if every piece in it is good, so a bad engine now puts the ASIC at risk instead of one pluggable module. Serviceability moves the same way, because the field-replaceable unit becomes the package and a failure that used to be a module swap taking seconds becomes a switch RMA. Operators also give up multi-vendor supply, which is much of why pluggables took over in the first place. That is why the economics work today only in the largest AI fabrics, where power per rack is the binding constraint and the buyer is large enough to negotiate a repair model. If package yield improves and external laser sourcing gets settled, the comparison changes for everyone else; until then, the power number on its own does not decide it.

13 minCo Packaged Optics In The Data CenterSemiconductor Engineering · 5k+ views

7 minThe Rise of Co-Packaged Optics in AI Data CentersIDTechEx · under 1k views

2 minCo-Packaged Optics (CPO): The Technical Breakthrough Powering the Future of AIBroadcom Inc. · 1k+ views

Implementation Agreement for a 3.2Tb/s Co-Packaged (CPO) Module (Optical Internetworking Forum) · Co-Packaging Framework Document (Optical Internetworking Forum)