SWIR imaging
SWIR imaging for semiconductor and silicon wafer inspection
Silicon stops absorbing light just past 1100 nm. That single fact turns an opaque wafer into a window, and it is the reason short wave infrared imaging has become standard practice for bond voids, cracks, vias, die attach and backside alignment. This guide covers the physics that sets your wavelength, the three geometries worth knowing, how to choose a camera, and the practical failures that stop a first attempt from working.
01. Why this works
Silicon is a window, just not to your eye
Silicon has an indirect band gap of about 1.12 eV, which puts the absorption edge at 1107 nm. Photons carrying less energy than the band gap cannot promote an electron across it, so they pass through unabsorbed. Everything else in this guide follows from that one number.
What makes the effect useful rather than merely interesting is how sharp it is. Absorption in silicon does not taper off gently across the near infrared; it collapses. Between 1000 nm and 1200 nm the absorption coefficient falls by more than three orders of magnitude, which means the difference between an opaque wafer and a transparent one is a couple of hundred nanometres of wavelength.
How far light gets into silicon
Change the wavelength850 nm: silicon is opaque
Deep in the silicon absorption band. The beam is gone within about 19 microns, which is why a conventional silicon camera sees a wafer as a solid object. Nothing below the surface is accessible at this wavelength, at any exposure.
1000 nm: the edge of useful
Absorption has dropped by an order of magnitude but a full thickness wafer still passes well under one percent. Thinned die become workable here, which is why some inspection is done with silicon sensors on thinned parts and then stops working the moment a full wafer appears.
1050 nm: transition
The band edge is close and transmission climbs fast. A 300 micron wafer now passes a usable fraction, but the result is strongly thickness dependent, so anything that varies in thickness across the field varies in brightness too.
1100 nm: past the band gap
The indirect band gap of silicon sits at 1.12 eV, which corresponds to 1107 nm. Above it, absorption falls off a cliff. A full 775 micron wafer transmits more than three quarters of the light that gets inside it, and buried structure becomes visible.
1200 nm: effectively transparent
Intrinsic silicon has stopped absorbing in any way that matters. Whatever you lose from here on is lost at the surfaces, not in the bulk, and the image no longer changes with wafer thickness. This is the comfortable operating point for through-silicon work.
1550 nm: transparent, and convenient
Fully transparent, well inside standard InGaAs response, and served by an enormous telecom ecosystem of sources, fiber and optics. The practical reason to come this far is component availability rather than any further gain in transmission.
Absorption coefficients are the Green and Keevers values for intrinsic silicon at 300 K. Internal transmission is exp(-αd). The second figure adds the two uncoated silicon faces, which reflect roughly 31 percent each and cap total transmission near 52 percent no matter how transparent the bulk becomes.
The bulk stops mattering somewhere around 1200 nm. Past that point, choosing a longer wavelength buys you component availability rather than transmission, and on heavily doped material it actively costs you.
02. Optical layout
Transmission, reflection and emission
Three arrangements cover nearly all semiconductor SWIR work. The choice is usually forced by how much of the part you can physically reach, and it decides both what defects you can find and how much light you will have to work with.
Three ways to point a SWIR camera at silicon
Watch the layout changeTransmission
Source on one side, camera on the other, wafer in between. Anything that blocks or scatters SWIR shows up as a shadow: a void in a bond line, a crack, a filled via, a metal feature. It is the highest contrast geometry available and the simplest to set up, and it only works when you can get to both faces.
- Wafer bond voids and interface defects
- Cracks and micro-cracks through the bulk
- Through silicon via fill and depth
Reflection
Source and camera on the same side. Light enters the silicon, reflects off a buried interface or a metal layer, and comes back out. You lose the contrast that a clean shadow gives you, and you gain the ability to inspect a part that is mounted, bonded or otherwise accessible from one side only.
- Backside alignment marks and overlay
- Die attach and underfill through the die
- Bumps, pillars and interconnect from behind
Emission
No source at all. The device emits the light itself, either as band edge photoluminescence near 1150 nm under optical excitation, or as electroluminescence when the junction is biased. Signals are faint and exposures are long, which is where cooled scientific InGaAs earns its cost.
- Solar cell electroluminescence and cracks
- Photoluminescence lifetime and defect mapping
- Hot spot and latch-up localization in ICs
Transmission gives the best contrast and needs both faces. Reflection tolerates a mounted or bonded part and gives up contrast in exchange. Emission needs no source at all, produces the faintest signal of the three, and is the only one that reports on the electrical behavior of the device rather than its geometry.
03. Applications
What SWIR actually finds in a semiconductor
Defects visible in the SWIR are the ones that change refractive index, block light, or emit it. That covers a surprising amount of what goes wrong in wafer level packaging and 3D integration, and it explicitly excludes anything hidden behind metal.
What you are looking for, and how to look for it
Pick a defect typeWafer bond voids
Trapped gas or incomplete contact at a bonded interface scatters and reflects SWIR, so a void reads as a bright or dark blob against a uniform field. Because the two bonded wafers together can exceed 1.5 mm of silicon, this is the application that most rewards going well past the band edge.
What drives the camera choice: Uncooled InGaAs is normally enough. Field of view matters more than sensitivity, since you want the whole bonded pair in as few tiles as possible.
Cracks and micro-cracks
A crack is a refractive index discontinuity, so it scatters light out of the collection cone and appears as a dark line. Sub-surface cracks that no surface inspection can reach show up clearly, which is why this is standard practice on sawn and ground wafers.
What drives the camera choice: High resolution helps more than sensitivity. Fine cracks are narrow, not dim, so sampling is the limiting factor.
Through silicon vias
A filled metal via is opaque against transparent silicon, giving very high contrast for presence, absence and alignment checks. Partial fill and voiding inside the via read as intensity variation rather than a clean shadow.
What drives the camera choice: Magnification is the constraint. Via pitch drives the sampling requirement, and that in turn drives the tile count.
Die attach and underfill voids
Looking through the back of a mounted die at the attach layer below it. Voids in solder, epoxy or underfill change reflectivity at that interface, and they matter because they become thermal and mechanical failures later.
What drives the camera choice: One-sided access is the whole reason for this geometry. Uniform, well-baffled illumination matters more than the camera.
Backside alignment and overlay
Reading front-side alignment marks through the wafer from behind, so that backside processing can be registered to front-side features. This is routine in MEMS, power devices and 3D integration.
What drives the camera choice: Repeatability and low distortion beat raw sensitivity. Calibrate the optical path, not just the camera.
Solar cell defects
Electroluminescence under forward bias, or photoluminescence under optical excitation, both peaking near 1150 nm. Cracks, finger interruptions, shunts and low lifetime regions appear as dark areas in an otherwise uniform glow.
What drives the camera choice: This is the one that genuinely needs a cooled scientific sensor. Signals are weak and exposures run into seconds.
04. Detectors
Choosing between the InGaAs families
Axiom Optics supplies SWIR cameras across the full range, from uncooled Sony SenSWIR sensors through cooled scientific InGaAs to extended InGaAs past 1700 nm. The choice is driven less by wavelength than by exposure time and tile count.
Which SWIR camera family fits
Pick a technologyUncooled Sony SenSWIR Volume inspection
The IMX990, IMX991, IMX992 and IMX993 family brought InGaAs into a price bracket that production inspection can absorb. The IMX992 and IMX993 in particular drop the pixel to 3.45 microns, which roughly doubles the sampling of the earlier 5 micron parts at the same magnification.
- Room temperature, no cooling overhead
- Visible plus SWIR response in one sensor
- 3.45 micron pixels on the newer parts
At Axiom: Alvium VSWIR, Goldeye Pro 5GigE, Alvium 1800 VSWIR cores
TE-cooled scientific InGaAs Faint and long
Thermoelectric cooling drops dark current far enough that exposures of seconds stay usable. This is what photoluminescence and electroluminescence work needs, because the sample emits a small number of photons and no amount of gain recovers a signal buried in dark noise.
- Long exposures without dark current swamping the signal
- Low read noise for faint emission work
- Stabilized operating temperature for repeatability
At Axiom: C-RED 2, C-RED 2 Lite, Goldeye TEC series
High-speed low-noise Throughput
When the limiting factor is how many fields you can acquire per shift rather than how faint the signal is. Fast readout with low noise keeps the tile count from turning into a scheduling problem, which as the calculator below shows happens sooner than most people expect.
- High frame rates at full resolution
- Low noise without deep cooling
- Suited to scanning a wafer tile by tile
At Axiom: SCION SWIR 640 and 1280, C-RED 3, FXO Sony SenSWIR
Extended InGaAs Past 1700 nm
Standard InGaAs stops near 1700 nm. Extended compositions reach 1.9 or 2.2 microns, which is rarely necessary for silicon itself but becomes relevant for encapsulants, adhesives and compound semiconductors that absorb inside the standard window.
- Cut-off at 1.9 or 2.2 microns
- Deep cooling required, higher dark current
- For materials that standard InGaAs cannot see through
At Axiom: Goldeye XSWIR, C-RED 2 ER 1.9 and 2.2 micron
The most common specification mistake is buying sensitivity for a job that needs speed, or speed for a job that needs sensitivity. Transmission imaging under good illumination is a bright, fast measurement. Photoluminescence is a dark, slow one. The same camera is rarely right for both.
05. The arithmetic nobody does first
Resolution, field of view and how long the scan takes
Magnification is usually chosen from the smallest feature that must be resolved, and the consequence for cycle time is discovered later. It is worth doing that arithmetic before the optics are ordered rather than after.
Sampling versus how long the scan takes
Interactive1x: whole-part screening
Wide fields, few tiles, and a whole 300 mm wafer surveyed in minutes. Good enough to find bond voids, large cracks and gross defects, and not close to good enough to judge a via.
2x: general inspection
The usual compromise. Sampling is fine enough for most interface and crack work while the tile count stays inside a sensible cycle time.
5x: feature level
Now you are resolving individual structures, and the tile count has become the dominant cost of the measurement. Worth doing on a sampled area rather than a whole wafer.
10x: detail only
Reserved for failure analysis on a known location. Covering a whole wafer at this magnification is not an inspection strategy, it is an overnight job.
Worked for a 1280 × 1024 array on a 3.45 micron pixel, the pitch used by the newer Sony SenSWIR parts. Smallest sampled feature is two pixels, the Nyquist limit, and real systems need more than that for a reliable call. Times count acquisition only; stage motion, settling and autofocus typically double or triple them.
The jump between 2x and 5x is the one that catches people. Sampling improves by a factor of two and a half, and the number of fields needed to cover a 300 mm wafer goes up by more than six. That is why full wafer inspection at high magnification is normally replaced by screening at low magnification followed by targeted review of whatever the screen flags.
06. Diagnosis
Why the first setup does not work
Through-silicon imaging fails in a small number of characteristic ways, and almost none of them are the camera. The symptoms below cover most of what turns up on a first attempt.
Why a through-silicon setup fails
Pick a symptomHeavily doped substrate
What you see: The wafer is darker than physics predicts
Free carriers absorb in the SWIR, and the effect grows with wavelength rather than shrinking. A lightly doped wafer that is transparent at 1550 nm can have a heavily doped counterpart that is visibly attenuated at the same wavelength, because free carrier absorption scales roughly with the square of wavelength.
- Ask for the substrate resistivity before specifying a wavelength
- On heavily doped material, test near 1200 nm as well as 1550 nm
- Expect epitaxial layers on a doped handle to behave like the handle
Rough or ground backside
What you see: The image is a uniform gray haze
As-sawn and ground surfaces scatter strongly. The silicon is transparent, but light leaves the surface in every direction and very little of it reaches the collection cone, so contrast collapses even though transmission looks fine on paper.
- Polished or etched surfaces transmit far better than ground ones
- An index matching fluid or film can recover much of the loss
- Test on a real part surface, never on a polished witness wafer
Interference fringes
What you see: Concentric rings across the field
A wafer with two parallel polished faces is an etalon. With a narrowband source, the 31 percent reflection at each face produces visible fringes that move with thickness and temperature, and they are easy to mistake for a defect map.
- Broaden the source bandwidth to wash the fringes out
- Tilt slightly off normal incidence where the geometry allows
- Suspect fringes first when a pattern tracks thickness variation
Metallization
What you see: Whole regions are simply black
SWIR passes through silicon, not through metal. Power and ground planes, thick redistribution layers and dense back end of line stacks are opaque, so any structure underneath them is unreachable from that direction no matter what wavelength you choose.
- Map which regions are optically accessible before committing
- Consider inspecting from the other side where the metal is thinner
- Accept that some structure needs X-ray or acoustic methods instead
Illumination that is not uniform
What you see: Real defects hide in the shading
Transmission imaging is a ratio measurement pretending to be a picture. Any non-uniformity in the source appears as apparent absorption, and low contrast defects such as partial bond voids are exactly what gets lost in it.
- Take and apply a flat field with the sample removed
- Diffuse the source rather than relying on a bare emitter
- Re-run the flat field whenever the source or optics are touched
Wrong sensor for the signal
What you see: Emission work never gets off the ground
Uncooled InGaAs is excellent for transmission imaging under bright illumination and unsuitable for photoluminescence at long exposure. Dark current accumulates with time, so the same camera that gives a clean image in 2 ms gives noise in 2 s.
- Match cooling to exposure time, not to the wavelength
- For emission work, budget for a TE-cooled scientific sensor
- Take a dark frame at the same exposure and temperature, every time
07. Specifying a system
Six questions to settle before you buy
1. How thick is the silicon, and how doped?
Total path length sets the wavelength floor and substrate resistivity sets the ceiling. A heavily doped 775 micron wafer is a very different optical problem from a lightly doped thinned die.
2. Can you reach both faces?
If yes, transmission gives you the best contrast available. If not, plan for reflection and budget more attention to illumination uniformity than you were expecting.
3. What is the smallest feature that matters?
Two pixels across it is the absolute floor and more is safer. That number chooses your magnification, and your magnification chooses your cycle time.
4. How long is a single exposure?
Milliseconds and an uncooled sensor is fine. Seconds and you need thermoelectric cooling. This question decides more of the budget than the wavelength does.
5. What state are the surfaces in?
Ground and as-sawn surfaces scatter badly. Test on real parts with real surface finish, because a polished witness wafer will tell you the setup works when it does not.
6. Have you run a feasibility test?
SWIR contrast on a real part is hard to predict from material data alone. Axiom Optics runs sample evaluations in its own lab, which is usually the fastest way from an idea to a specification worth quoting.
08. Frequently asked
Common questions about SWIR wafer inspection
Why can a SWIR camera see through a silicon wafer at all?
Silicon has an indirect band gap of about 1.12 eV, which corresponds to 1107 nm. Photons with less energy than the band gap cannot promote an electron across it, so they are not absorbed and pass straight through. Below that wavelength silicon absorbs strongly, which is why a normal camera sees a wafer as a solid gray object and a SWIR camera sees the structure inside it.
What wavelength should I use for through-silicon inspection?
Somewhere between 1200 nm and 1550 nm for most work. By 1200 nm intrinsic silicon has effectively stopped absorbing, so nothing further is gained in the bulk by going longer. The reason to choose 1550 nm anyway is practical: sources, fiber and optics at telecom wavelengths are cheap and abundant. Heavily doped substrates are the exception, since free carrier absorption gets worse at longer wavelengths rather than better.
How much light actually gets through a wafer?
Less than the absorption figures suggest, because the bulk is not the only loss. Silicon has a refractive index near 3.5, so each bare face reflects about 31 percent. Two uncoated faces cap total transmission near 52 percent however transparent the material becomes. If your budget assumed 100 percent, you are about a stop short before anything else goes wrong.
Do I need a cooled camera?
It depends on exposure time, not on wavelength. Transmission imaging under decent illumination runs in milliseconds and an uncooled Sony SenSWIR sensor handles it comfortably. Photoluminescence and electroluminescence run for seconds, and over that period dark current accumulates until it swamps a faint signal. That is the case where thermoelectric cooling stops being a luxury.
Can SWIR see through metal layers?
No. SWIR transparency is a property of silicon, not a general ability to see through solid objects. Power planes, thick redistribution layers and dense interconnect are opaque at every SWIR wavelength. Map which regions are optically accessible before designing the inspection, and expect that some structure will need X-ray or acoustic microscopy instead.
Why does my through-wafer image look hazy with no visible defects?
Usually surface scatter. A ground or as-sawn backside sends light in every direction, so very little reaches the collection cone even though bulk transmission is high. Polished or etched surfaces behave completely differently, and an index matching fluid recovers much of the loss. Always run the feasibility test on a real part surface rather than a polished witness wafer.
In summary
A window that opens at 1100 nm
SWIR inspection of silicon rests on one property of the material and a handful of practical consequences. Past the band gap edge, the wafer stops being an obstacle and becomes something you look through, which makes buried interfaces, sub-surface cracks, filled vias and die attach layers accessible to a camera rather than only to destructive analysis or acoustics.
What separates a working system from a frustrating one is rarely the sensor. It is knowing that surfaces reflect a third of the light at each face, that ground backsides scatter, that heavily doped substrates absorb more at longer wavelengths rather than less, and that metal is opaque no matter what. None of that is difficult once it is expected.
Test it on your own wafers
Axiom Optics supplies SWIR and VSWIR cameras, lenses and hyperspectral systems across North America, from uncooled Sony SenSWIR sensors to cooled scientific and extended InGaAs. Tell us the stack, the defect and the cycle time, and we will help you work out what fits.