SWIR wavelength selection by material: silicon, moisture, plastics

Short-wave infrared (SWIR) imaging fails most often at the wavelength step. A team buys an InGaAs camera, points a shelf infrared light at the part and gets a flat grey image, because the band was never matched to the material. Silicon, water and plastics each carry a fixed absorption spectrum, and each picks its own SWIR wavelength: silicon turns transparent above its ~1050 nm band gap, water absorbs hard near 1450 nm, and polymer signatures separate from 1650 nm up. If you are adding SWIR to a wafer, moisture or sorting line, this guide gives you the band for each job, the margin to keep around it, a worked example on a silicon die, and the camera, lens and strobe checks to run before you order.

The band decision at a glance

Three of the four rows need an InGaAs camera; only the 940 nm row runs on the CMOS cameras most plants already own. The lighting splits the same way: shelf spot lights and bar lights cover the visible bands through 940 nm, while a SWIR band is built to order in the special lights category. If the open question is the sensor family rather than the band, the SWIR, NIR and visible comparison walks that choice; this guide fixes the wavelength.

Why the wavelength is a material property

Absorption is a property of the bonds in the target, not a tuning knob. A semiconductor absorbs light above its band gap and passes light below it, which is why silicon turns clear just past 1050 nm. Water absorbs wherever its O-H stretch overtones fall, at 1450, 1950 and 2500 nm. Polymers carry their own bond signatures, so each type absorbs and transmits differently across the SWIR range.

A narrowband LED placed on the feature buys the most contrast per amp of drive: photons that miss the absorption feature do no work. That is the same logic behind single-wavelength monochrome lighting, applied one band further out.

Silicon: 1150-1200 nm for cracks inside wafers and die

Past its ~1050 nm band gap, silicon behaves like dark glass: light crosses the wafer, and anything that interrupts the path — a crack, a residue, a bonding void — lands as a dark feature on a bright transmitted field. That is transillumination, and it reaches features no front-side reflection image can see.

Start at 1150-1200 nm. The 100-150 nm above the band gap keeps transmission fully developed, and the band sits far from the water peak, so humidity films on the surface do not add contrast you cannot explain. Note what the band does to the camera too: silicon sensors respond from 350 to 1000 nm with sensitivity falling fast beyond 900-1000 nm, so any SWIR band returns a black frame on a CMOS camera.

Two hardware notes follow. The optics change with the band, because lens corrections that hold in the visible drift outside it; the SWIR lens basics article lists what moves. And when the target is marks on the wafer surface rather than features inside it, the reflective-wafer OCR guide covers that front-side problem.

Moisture: 1450 nm reads the first 0.4 mm of water

Water absorbs in its O-H stretch overtones near 1450, 1950 and 2500 nm; the first of these is the band that fits both a standard InGaAs window and practical LED construction. Absorption there is roughly sixty times stronger than in the 970 nm band, and the sensor's range is spent within about 0.4 mm of liquid water. Turn that around and the reach of the weak band becomes concrete: the absorption depth that saturates the band in 0.4 mm takes 0.4 mm × 60 = 24 mm of water at 970 nm.

That arithmetic is the selection rule. Wet against dry on a surface — condensation, leaks, wet labels, a fill line read through clear plastic — is a job for the water band: backlight it or flood it with a bright diffuse field, and wet areas read jet black against a bright background. Water deeper inside the part needs the weak band instead: 940-970 nm, where shelf infrared spot and bar lights already run on ordinary CMOS cameras. The near-infrared wavelength selection guide works that trade in detail, and the liquid level lighting guide covers fill-level geometry in transparent containers.

Plastics: 1650 nm and above separates polymer types

White flakes all look the same in visible light because every polymer reflects white the same way. From 1650 nm up, each polymer's bond signature produces its own absorption, so mixed types separate in a single narrowband frame: the polymer that absorbs reads dark, the one that transmits stays bright. That contrast is what recycling-stream sorting is built on, and the same logic answers any plastic-versus-plastic call on a production line.

Band sequencing earns its keep here: fire the polymer band and a second band in step with consecutive frames, then let the logic compare. The sequencing format looks like this:

The LS-SPC-0086 is a 150 mm multi-channel sequencing light source with 8 zones, run from 24 V. One channel per band, each fired against its own frame, is the pattern a two-band polymer sort follows.

Two practical points close the section. At the sort band a standard InGaAs sensor is near its upper edge, so check the camera covers the band, and treat anything well past 1700 nm as extended-InGaAs or MCT territory. Do not fix the band from a table alone either: image real flakes of both polymers at two candidate bands and build on the wider gap. Downstream choices — pixels, resolution, conveyor speed — are the sorting conveyor camera selection article's territory.

Worked example: choosing the band for cracks under a silicon die

Inputs

• Part: a bonded silicon die; the crack to find sits at the die-attach interface, so the light must cross the silicon (typical case for this topic).
• Silicon band gap: ~1050 nm, transmission well established above ~1100 nm (published wafer-inspection practice).
• CMOS sensor response: 350-1000 nm, with sensitivity falling fast above 900-1000 nm.
• InGaAs window: 900-1700 nm, with the highest sensitivity inside that range.
• Candidate bands on the bench list: 940 nm (shelf infrared), 1150 nm, 1200 nm and 1450 nm.

The rule

Pick the lowest band that clears the band gap with margin and sits comfortably inside the sensor window. Keep about 100 nm above the band gap so transmission is fully developed, and drop any band above 1000 nm from a CMOS camera plan, because the sensor stops before the band starts.

Arithmetic

• 940 nm: 940 − 1050 = −110 nm. The band sits below the band gap, silicon still absorbs, and the die stays opaque. Fails at any brightness.
• 1150 nm: 1150 − 1050 = 100 nm of transmission margin; 1700 − 1150 = 550 nm of sensor margin; and 1150 − 1000 = 150 nm past the CMOS cutoff, which confirms the camera must be InGaAs.
• 1200 nm: 1200 − 1050 = 150 nm of transmission margin and 1700 − 1200 = 500 nm of sensor margin. Same camera conclusion, more optical headroom.
• 1450 nm: crosses silicon as well, but it lands on the water absorption peak, so humidity films and surface wetting paint black marks that are not cracks.

Result and decision

Run transillumination at 1150-1200 nm on an InGaAs camera — 1150 nm for a thin die, 1200 nm when the stack is thicker or the source is marginal. The thresholds to carry away: silicon-through work starts above 1100 nm, and any band past 1000 nm commits the project to an InGaAs sensor. A 940 nm shelf light cannot be made to work here: below the band gap, the silicon is simply not transparent.

Camera, lens and strobe drive around the band

The camera comes first: InGaAs covers roughly 900-1700 nm, and bands well past 1700 nm move to extended-InGaAs or MCT sensors. The lens follows, since visible-corrected glass drifts outside its design range. The light is the part most often underestimated, and its problem is heat.

SWIR emitters turn less drive current into light than visible LEDs, so more of the drive leaves as heat. Heat shifts the peak wavelength slightly and raises dark current in the sensor, and both eat the contrast you built the band for. Strobe drive is the answer: pulse the emitter only while the shutter is open, overdrive the pulse for peak brightness on a moving line, and keep average heat low. The lighting controller guide covers the wiring case, strobe mode and LED lifetime shows why pulsing also stretches emitter life, and the strobe rise time guide puts numbers on freezing motion.

A strobed build looks like this on the shelf:

The LS-SPC-0129 is an 84 x 60 x 60 mm high-power strobe light on a 48 V supply. The hidden-strobe technique is the drive pattern a high-power strobe is made for.

Ordering the light: shelf bands and custom builds

The catalog splits the same way the band table does. The 940 nm row is shelf hardware: infrared spot and bar ranges cover 850 and 940 nm wavelengths and run on ordinary CMOS cameras. A SWIR band is a build: wavelength, bandwidth, lit length and drive voltage are specified against the application and manufactured through the custom-built light series. Start at the lights finder or go straight to the lighting step of the configurator; the custom bar format looks like this:

The LS-SPC-0040 is a customized combined bar light in the 200-400 mm size class on a 24 V supply. The same custom series is where each of these SWIR builds lands.

When the usual band is the wrong band

Four cases turn the standard answers around:

• Water deeper than about 0.4 mm. The water band saturates inside the first fraction of a millimetre, so everything underneath reads the same black. Step down to 940-970 nm and trade surface drama for penetration; the CMOS camera you already own keeps working.
• The defect is on the surface, not inside the part. A scratch on a reflective face is a reflection-angle problem, not a band problem; the bright field versus dark field guide settles it faster than any wavelength table.
• The polymer pair separates weakly. Signatures differ by type. Image real flakes of both at the sort band and one band above it, and build on the pair with the wider dark-to-bright gap.
• The line already meets its contrast budget at 940 nm. SWIR buys band-specific contrast, not general brightness. If CMOS hardware answers the call, the SWIR project can wait for a task that needs it.

Checks before ordering

Six checks, one per failure mode:

1. Write the band from the material, in nm. 1150-1200 nm through silicon, 1450 nm for surface water, 1650 nm and up for plastic type — from the absorption table above, not from a camera brochure.
2. Match the sensor window to the band. CMOS stops near 1000 nm, InGaAs reaches 1700 nm, and bands beyond that are extended-InGaAs or MCT decisions.
3. Check every optic in the path. Lens, window and fibre must all pass the band; visible-corrected optics drift outside their design range and lose resolution first.
4. Budget the strobe. Pulse width and duty at line speed decide whether the overdrive peak stays inside the emitter's rating and whether motion freezes.
5. Prove contrast on real samples. Wafer scrap at 1150 and 1200 nm, wet and dry coupons at the water band, flake pairs at the sort band — one image each before the build is ordered.
6. Specify the build in one line. Wavelength, bandwidth, lit length, drive voltage and trigger; that line is what the custom light series quote starts from.

If samples need a bench test first, send the band, the part and the line speed through the contact page. The 940 nm row can start from the spot and bar ranges as they are, and the rest of the illumination decisions — geometry, diffusers, sequencing — live on the lighting topic page.