Raman spectroscopy · 14 min
532, 638, or 785 nm is the first decision you make about a Raman system and the hardest one to reverse. Three physical rules decide it, and they pull in opposite directions.
The governing trade-off
Almost every disappointing Raman measurement traces back to one of three relationships. None of them is complicated on its own. The difficulty is that the wavelength which optimises one actively degrades another.
I_Raman ∝ ν⁴ ∝ 1 / λ⁴
Scattering strength rises with the fourth power of frequency. Moving from 785 to 532 nm multiplies the Raman signal by (785/532)⁴ ≈ 4.7 — the difference between a five-second and a twenty-five-second acquisition.
SNR ∝ I_Raman / √I_background
Fewer molecules have electronic transitions reachable at 785 nm than at 532 nm. Because the background contributes shot noise, a fluorescence hump ten times the Raman signal costs far more than the ν⁴ law gives back.
λ_Stokes = 1 / (1/λ_ex − ν̃)
Raman shifts land at longer wavelengths than the laser. From 785 nm, the C–H stretch at 2900 cm⁻¹ arrives near 1016 nm, where silicon quantum efficiency has already collapsed.
Interactive
The simulation below applies the ν⁴ law, a fluorescence background, and a real silicon QE curve to the same set of Raman bands. Raise the fluorescence to a biological sample and watch 532 nm — the strongest scatterer — become the worst choice on the list.
Signal strength, fluorescence background, and detector reach all move at once. Change the wavelength and watch which of the three becomes your limiting factor.
Raman shift (cm⁻¹)
Raman signal
4.7×
ν⁴ law, vs 785 nm
Fluorescence
10.0×
background, vs 785 nm
Peak contrast
1.4×
signal ÷ √background
Lateral resolution
361 nm
0.61 λ / NA 0.90
Typical polymer or pharmaceutical
NA 0.90
Detector reach
Silicon still responds across the full fingerprint and C–H stretch region (to ~3400 cm⁻¹).
The ν⁴ scaling, Stokes-shift arithmetic, and silicon QE curve are quantitative. The fluorescence model is a schematic exponential — real fluorescence is wildly sample-specific and can swing by orders of magnitude between two samples of nominally the same material. Use this to understand the shape of the trade-off, never as a prediction for your sample.
At a glance
| Wavelength | Relative signal | Fluorescence risk | Lateral resolution | Typically chosen for |
|---|---|---|---|---|
| 405 nm | ~14× vs 785 nm | Severe on almost any organic | ~275 nm at NA 0.9 | Wide-bandgap semiconductors, resonance Raman, diamond |
| 532 nm | ~4.7× vs 785 nm | High — the usual reason it fails | ~360 nm at NA 0.9 | Carbon and 2D materials, minerals, inorganics, semiconductors |
| 638 nm | ~2.3× vs 785 nm | Moderate | ~432 nm at NA 0.9 | General-purpose work, mixed sample sets, some biological samples |
| 785 nm | 1× (reference) | Low — the standard escape route | ~532 nm at NA 0.9 | Polymers, pharmaceuticals, biological tissue, forensics |
| 1064 nm | ~0.24× vs 785 nm | Essentially none | ~721 nm at NA 0.9 | Strongly fluorescent samples where nothing else works |
Relative signal is the ν⁴ ratio against 785 nm before detector response is applied. Lateral resolution is 0.61 λ / NA at NA 0.9.
A defensible order
The order matters more than any single specification. Optimising signal first is the most common way to end up with an instrument that cannot measure the samples it was bought for.
Fluorescence is the only failure mode that returns no spectrum at all. A weak spectrum can be integrated for longer; a spectrum buried under a fluorescence hump cannot be recovered. Ask what your sample does under green light before optimising anything else.
Once fluorescence is survivable, move as far to the blue as you can tolerate. Every step shorter buys signal at the fourth power — the difference between 785 and 532 nm is roughly a factor of five in integration time for the same signal-to-noise.
Confirm the highest Raman shift you care about still lands where your detector responds. If you need the C–H stretch region near 2900 cm⁻¹ on a silicon array, 785 nm is already marginal and 1064 nm is out of the question.
Lateral resolution scales as 0.61 λ / NA. At NA 0.9, moving from 532 to 785 nm costs roughly 170 nm of resolution. That is irrelevant for bulk measurement and decisive for mapping small inclusions or single particles.
Open source
Wavelength choice determines what reaches your detector. What happens next — calibration, baseline removal, peak identification — decides whether it becomes an answer. We develop and maintain Raman Studio, an open-source desktop application for acquiring, processing, calibrating, and identifying Raman spectra, built around OpenRAMAN spectrometers.
Live camera feed with hardware ROI management for FLIR/PointGrey sensors.
Baseline removal, Savitzky–Golay smoothing, median filtering, blank subtraction, and peak detection.
Wavelength calibration against known emission lines, with automatic peak-to-line assignment.
Reference-library scoring by cosine similarity (HQI) with spectral overlay, plus structure-based prediction from SMILES or compound names via Mol2Raman.
Common questions
785 nm if your samples are organic, biological, or unknown, because fluorescence is the failure mode that produces no usable spectrum at all. 532 nm if you work with inorganics, carbon materials, or semiconductors, where fluorescence is rarely a problem and the ν⁴ signal advantage is decisive. 638 nm is the sensible single choice for a mixed sample set.
Raman scattering intensity scales with the fourth power of the scattered frequency. Because 532 nm has a higher frequency than 785 nm, the ratio is (785/532)⁴ ≈ 4.7. All else being equal, 532 nm excitation returns roughly five times more Raman signal from the same sample.
A Raman shift of 2900 cm⁻¹ from 785 nm excitation lands at about 1016 nm, and 3200 cm⁻¹ lands beyond 1030 nm. Silicon detector quantum efficiency has collapsed to a few percent there and reaches zero near 1100 nm. The fingerprint region is unaffected, but high-wavenumber bands become difficult or impossible without an InGaAs detector.
Not inherently — the opposite, usually. For a fixed spectrograph, a given interval in cm⁻¹ spans more nanometres at longer excitation wavelengths, so each detector pixel covers fewer wavenumbers. The practical catch is that the full Stokes range then needs more detector real estate and runs into the silicon cutoff.
Baseline correction removes the shape of a fluorescence background, but not the shot noise it carries. If the background is much larger than the Raman bands, that noise buries weak peaks and no algorithm recovers them. Photobleaching, shifted-excitation difference methods, and time-gating all help, but choosing the right excitation wavelength is the only approach that prevents the problem instead of treating it.
It can. Shorter-wavelength photons carry more energy, and at the power densities of a focused confocal spot this matters for biological samples, pigments, and thermally sensitive materials. Sample burning during 532 nm measurements is a common and often unrecognised failure. Reduce power and lengthen integration before assuming a sample is simply a weak scatterer.
The honest answer usually requires measuring one representative sample at two wavelengths. Our confocal Raman system is configurable across 532, 638, and 785 nm for exactly this reason.