Optical trapping · 16 min
Not the one with the most watts. Wavelength is decided by what the light does to your sample and what your detector can see; noise is decided by how small an error bar you need. Here is how both decisions actually work.
Start here
This surprises people, so it is worth stating plainly. For a bead much larger than the wavelength — which describes the 0.5–2 µm spheres used in almost all biophysics — the trapping efficiency of a single-beam gradient trap is set by geometry and the refractive-index contrast between bead and medium. Wavelength barely enters. Swapping 1064 nm for 830 nm at the same power in the specimen plane changes your maximum force by a few percent, not a factor.
The scaling only bites for particles much smaller than the wavelength. In that Rayleigh limit, stiffness at fixed power and numerical aperture goes roughly as 1/λ⁴, so 830 nm really would be about 2.7 times stiffer than 1064 nm on a 100 nm bead. If you trap nanoparticles, that matters. If you trap micron beads, it does not.
So the wavelength decision is made almost entirely on four other grounds: how badly the light damages a living sample, how much it heats the water around it, how well your position detector responds to it, and how much of it your objective actually transmits. Those four constraints have their optima in four different places.
The four real constraints
Each of these is well characterised in the literature. The difficulty is that no single wavelength wins on more than two of them, which is exactly why 1064 nm — optimal on none — became the standard.
The action spectrum for damage to E. coli varies roughly sevenfold across 790–1064 nm, with minima at 830 and 970 nm and sharp maxima at 870 and 930 nm. Damage scales linearly with intensity — a single-photon, oxygen-mediated process — so you cannot outrun it by shortening exposure at higher power.
Absorption by the solvent, not the bead, dominates heating. At 1064 nm the focus rises about 8 K per watt in water. At 980 nm — right at the photodamage minimum — water absorbs nearly four times more strongly, so the wavelength that is kindest to the cell is one of the harshest on the buffer around it.
A silicon quadrant photodiode is nearly transparent at 1064 nm. Carriers generated outside the depletion layer arrive by diffusion, turning the detector into an unintended low-pass filter with a 3 dB point near 9 kHz. Below 900 nm the effect disappears and tens of kHz of honest bandwidth come back.
High-NA objectives are not achromatic in the near infrared. A standard 100×/1.4 oil objective can transmit 50 % at 830 nm but only 32 % at 1064 nm, while its infrared-corrected sibling holds about 60 % at both. That single specification can outweigh every other power consideration in the system.
Interactive
The 970 nm region is where cells survive best and where water absorbs almost the worst. The 830 nm region is gentle, cool, and detects beautifully — and is a wavelength at which high-power single-frequency sources are comparatively scarce. Drag the slider and the shape of the compromise becomes obvious.
Drag the wavelength. The red curve is how hard your focus heats the water; the dashed violet curve is how hard it damages a cell. Their minima are in different places, and the detector prefers a third answer entirely.
Focal heating
+1.6 K
8.1 K/W at sample
Water absorption
0.12
cm⁻¹, Hale & Querry
Photodamage
2.0×
relative to the 970 nm minimum
Si QPD bandwidth
2.0 kHz
before parasitic filtering
1064 nm
200 mW — roughly 20 pN of maximum trapping force on a micron bead
Verdict at 1064 nm
Heating stays around 1.6 K at this power, and a silicon quadrant detector will quietly low-pass your power spectrum unless you correct for it or change detector.
Water absorption is tabulated data. Focal heating scales that linearly from the 8 K/W measured at 1064 nm, which is a first-order estimate — the real value depends on chamber geometry and how far the trap sits from the coverslip heat sink, and can be two to three times higher deep in a chamber. The photodamage curve is a reconstruction pinned to the published minima at 830 and 970 nm, maxima at 870 and 930 nm, and roughly sevenfold total range; it is cell-type specific in reality. Use the shapes, not the digits.
At a glance
| Wavelength | Focal heating in water | Photodamage to cells | Silicon detection | Verdict |
|---|---|---|---|---|
| 785 nm | ~1.6 K/W | Rising — near the 760 nm danger zone | Excellent | Beads and colloids where you need detection bandwidth. Not first choice for live cells. |
| 830 nm | ~1.9 K/W | Local minimum | Excellent | The classic low-damage, high-bandwidth choice for single-molecule work. |
| 852 nm | ~2.9 K/W | Climbing toward the 870 nm maximum | Excellent | Good when you want a stocked diode line near 830 nm and the sample is not alive. |
| 980 nm | ~29 K/W | Near the global minimum | Poor — a few kHz | Gentlest on cells, harshest on water. Only at low power, and never for fast spectra. |
| 1064 nm | ~8 K/W | Low | Poor — needs correction or InGaAs | The default, and deservedly so: the best compromise plus the deepest component ecosystem. |
| 1310 nm | ~74 K/W | Not characterised here | InGaAs only | Boils the sample. Occasionally used in thin chambers; rarely worth it. |
| 1550 nm | ~790 K/W | Irrelevant — thermal damage first | InGaAs only | Unusable in water. Routine and excellent in vacuum, where there is nothing to absorb. |
Heating is scaled linearly from the ~8 K/W measured at 1064 nm using tabulated water absorption, and is a first-order estimate that rises further when the trap sits far from the coverslip. Photodamage columns follow the published action spectrum for E. coli, measured only between 790 and 1064 nm.
The second decision
“Low noise” is the most abused phrase on a laser datasheet. For a trap it means three specific things, each of which corrupts a different part of your measurement in a way you can calculate before you buy.
κ ∝ P ⟹ ΔF/F = ΔP/P
Trap stiffness is proportional to the power in the specimen plane. A laser that drifts 2 % over a session puts a 2 % systematic error on every force in the dataset. It is not random, so averaging does not touch it, and it propagates directly into any stiffness you calibrate.
Δx = f_obj · θ / M
An angular wobble θ at the laser arrives at the objective pupil demagnified by the beam expander, and the objective turns it into a lateral shift of the focus. 30 µrad through a 4× expander into a 2 mm objective is 15 nm of trap motion — forty-four DNA base pairs, indistinguishable from your molecule moving.
P(f) = kBT / (γπ²(fc² + f²))
Power-spectrum calibration assumes everything in the spectrum is the bead. Laser intensity noise inside the trap bandwidth adds power that the Lorentzian fit misattributes, shifting the corner frequency and therefore both the stiffness and the nm-per-volt conversion.
A very common specification is RMS noise < 0.5 %, 20 Hz–20 MHz. Nearly all of that bandwidth is irrelevant to you: a bead in water rolls off above its corner frequency, typically a few hundred Hz to a few kHz. If the noise were white, 0.5 % RMS over 20 MHz would be about −119 dBc/Hz, and the part landing in DC–10 kHz would be 0.01 % — negligible.
Real laser noise is not white. It is dominated by a 1/f component at low frequency and by a relaxation-oscillation peak in the hundreds of kHz. The wideband figure can be set entirely by that peak while the band you care about is far worse than 0.01 % — or the reverse. Ask for the noise spectral density curve, or an RMS figure over DC–10 kHz. A supplier who can produce one is telling you something real.
In water, the bead is heavily overdamped. Energy cannot accumulate, so intensity noise does not “heat” the particle; it shows up as stiffness error, excess low-frequency position variance, and a biased power-spectrum fit. Drift below 10 Hz is the practical enemy, because that is the band where you cannot distinguish it from biology.
In vacuum, a levitated particle is underdamped and the atom-trap result applies directly: intensity noise at twice the trap frequency parametrically pumps the motion, with an energy e-folding rate π²ν²S(2ν). Holding a 10 kHz trap quiet for 100 seconds demands a fractional intensity noise density below about 3 × 10⁻⁶ /√Hz. This is why levitated optomechanics groups care about RIN more than anyone.
Interactive
Take the pointing and stability figures from the datasheet in front of you, add your objective and beam expander, and see what they cost on the force you are trying to measure. Most people find one term dominates by a wide margin — and it is rarely the one they were negotiating over.
Pointing stability and power stability are the two laser specifications that convert directly into force error. Set them to the values on the quote you are considering.
30 µrad
Typical research DPSS lasers quote ±30–50 µrad long-term.
4×
Expanding the beam demagnifies its angular noise by the same factor.
1.00 %
The 4-hour or 8-hour figure on the datasheet, not the RMS noise figure.
0.050 pN/nm
0.05 pN/nm is a typical single-trap value on a micron bead.
5.0 pN
Kinesin steps at ~5 pN; DNA overstretches near 65 pN.
Trap jitter
15 nm
from pointing alone
Pointing error
0.75 pN
κ × jitter
Power error
0.05 pN
systematic, on every point
Total
15.0 %
0.75 pN of 5.0 pN
Where the error comes from
Pointing dominates here. Buying a quieter power supply will not help — you need better pointing stability, more beam expansion, or an enclosed and thermally stabilised beam path.
For scale: thermal motion
At 0.050 pN/nm the bead itself rattles around by √(kBT/κ) = 9.1 nm RMS, which is smaller than the 15.0 nm of trap jitter above. That comparison is misleading in the way that matters: Brownian motion is broadband and averages down as 1/√N, so a one-second average beats it into the sub-nanometre range. Pointing drift and power drift live at low frequency and in the drift band — they survive averaging, and they are what actually sets the floor on a slow measurement.
This assumes the steering optics are conjugate to the objective back focal plane, so a pointing change arrives as a pure angle at the pupil and translates the trap. If your relay is not conjugated, the same pointing noise instead walks the beam across the pupil and modulates both stiffness and trap position at once, which is worse and harder to model. Thermal drift of the whole optical path adds to everything shown here and is usually the larger term over hours.
The specification sheet
Corrupts: The stiffness calibration, and therefore every force you report
Trap stiffness is proportional to power in the specimen plane. A 2 % drift is a 2 % systematic error on the whole dataset that no averaging removes.
Corrupts: The power spectrum you fit for the corner frequency
Excess intensity noise inside the trap bandwidth adds power that the Lorentzian fit attributes to the bead, biasing both stiffness and the detector calibration.
Corrupts: The position of the trap itself
An angular wobble at the objective pupil translates the focus in the sample plane. It is indistinguishable from the molecule you are watching having moved.
Corrupts: How well the focus matches diffraction limit, hence axial trapping
A poor mode fills the pupil unevenly and produces a focus with a shallower axial gradient, which is where a single-beam trap is weakest to begin with.
Corrupts: Everything, intermittently
A multi-mode DPSS laser can hop between modes, producing sudden power steps and small pointing shifts. It looks exactly like a real event in your data.
Corrupts: Power downstream of every polarising element
Traps are usually steered and split with polarising optics. Drifting polarisation becomes drifting power at the sample even if the laser output is perfectly stable.
One specification that never appears on a datasheet matters as much as any of these: back-reflection sensitivity. A coverslip reflects a few percent of your trapping beam straight back down the optical path, and both DPSS and diode lasers become unstable when that light re-enters the cavity. A Faraday isolator immediately after the laser is not optional on a trapping bench — it is part of the noise budget.
The trap you cannot see does not exist
The parasitic-filtering problem is the single most common way a well-built 1064 nm trap produces quietly wrong numbers. An ordinary silicon quadrant photodiode low-passes the signal with a 3 dB frequency near 9 kHz, so the measured power spectrum rolls off early, the fitted corner frequency comes out low, and the stiffness is underestimated — with no obvious sign that anything is wrong.
There are four honest fixes: use an InGaAs detector, use a position-sensitive detector specified for the near infrared, heavily reverse-bias the silicon, or fit the published filter model alongside the Lorentzian. Choosing a wavelength below 900 nm sidesteps the problem entirely, which is a legitimate reason to consider 785 or 830 nm when your samples are not alive.
Our recommendations
Ordered by what you are building rather than by price. Every figure below is from the model’s own specification sheet; where a number is not published we say so rather than estimate it.
A single-frequency 1064 nm fibre laser, and the quietest source in our catalogue by an order of magnitude. The RIN figure is specified as a spectral band rather than a single wideband number, which is what you want to see. Choose this for microrheology, single-molecule force spectroscopy, holographic multi-trap work, or anything where the force value is the result.
The lowest-noise free-space single-frequency DPSS option at 1064 nm. 300 mW at the laser is roughly 60–100 mW at the specimen plane after a realistic optical train, which is a genuinely strong single trap. A sensible default for a research instrument that does not need watts.
The same single-frequency architecture with more than twice the power, so you can run the laser well below maximum current — which is itself a noise and lifetime argument. Useful when you split the beam for a second trap or a detection arm.
Watt-class single-frequency output for setups that divide power across many traps, or that lose most of it in a spatial light modulator. Water-cooled, so plan the bench and the acoustic environment accordingly.
A low-noise TEM₀₀ 1064 nm head without the single-frequency premium. Right for a teaching trap, an alignment-stage instrument, or a build where you want a working trap before committing to the final source.
Single-mode fibre delivery moves the laser and its heat off the optical table and hands the microscope a clean, repeatable mode. The trade is that fibre adds its own pointing and polarisation drift, so specify PM fibre and mount the launch rigidly.
When you need the full bandwidth of a silicon quadrant detector — fast power spectra, high-frequency microrheology, kilohertz feedback — a 785 nm source removes the parasitic filtering problem entirely. Water absorbs five times less here than at 1064 nm too. The caveat is photodamage: this is a beads-and-colloids wavelength, not a live-cell one.
The closest stocked line to the 830 nm photodamage minimum, with silicon detection still fully intact and low solvent heating. Note that 852 nm is on the rising edge toward the 870 nm damage maximum, so for live cells a true 830 nm source is worth specifying instead. Ask us for the noise and stability figures on the specific configuration.
Sits essentially at the global photodamage minimum, which makes it attractive for long holds on living cells. The cost is heating: water absorbs roughly 3.5 times more strongly here than at 1064 nm, so keep the power at the sample low and calibrate at the temperature you actually run at.
None of these is a trapping subsystem on its own. A laser is roughly a fifth of the parts list and rather less than a fifth of the problem — everything below sits between it and a number you can publish.
The rest of the instrument
Grouped by subsystem, with the reason each one belongs in a trapping budget rather than a generic description. If you are costing a build, this is close to the full parts list.
This is where a trap becomes a force transducer. A camera finds beads; only a fast photodiode chain gives you thermal spectra, kilohertz feedback, and active microrheology.
Low-noise transimpedance amplifier with X, Y, and sum outputs. The sum channel is what lets you normalise away common-mode intensity noise.
Digitising a QPD at 8 bits throws away the resolution you paid the laser for. 20 bits at 100 ksps oversamples a 10 kHz corner frequency comfortably.
Supply ripple on the detector rails appears as position noise. It is the cheapest noise source to remove and the most often ignored.
Closes the loop when the QPD error signal has to drive trap steering or sample stabilisation rather than just being recorded.
One honest warning: 250 kHz of amplifier bandwidth does nothing if the photodiode in front of it is the bottleneck. At 1064 nm an ordinary silicon quadrant detector rolls off near 9 kHz regardless of how good the electronics are. Specify the detector material and the wavelength together, or move below 900 nm.
Everything between the laser aperture and the objective pupil. This chain sets how much power reaches the sample, how stable the trap is, and whether you get one trap or many.
Not optional. The coverslip returns a few percent of the beam into the cavity, and both DPSS and diode sources destabilise when it does.
Fills and slightly overfills the objective pupil, which is what produces a steep axial gradient. It also demagnifies the laser’s pointing noise by the same factor.
The standard way to control power without touching laser current, to gate the trap, and to close an intensity-stabilisation loop against a pickoff photodiode.
Lets you run the laser at its quiet operating point and throw power away, rather than turning the current down into the noisy regime.
Splits one source into two independently steerable traps, or into a trapping arm and a detection arm with adjustable ratio.
Holographic multi-trap generation. Budget for it in the power calculation — an SLM path can cost you more than half the beam.
Dichroics and blocking filters to separate the trapping wavelength from the imaging path. At 1064 nm the separation is easy; at 785 nm it competes with your fluorophores.
The noise budget from the lab above assumes the optics hold still. Over the hours a real measurement takes, mechanical drift is usually larger than anything the laser contributes.
Nanometre sample positioning, and the standard route to a viscous-drag stiffness calibration by driving the stage at known velocity.
Two-axis steering, ideally conjugate to the objective back focal plane so the trap translates without walking the beam across the pupil.
Coarse XYZ for loading samples and finding the chamber, underneath the fine piezo motion.
Rigid, pre-aligned mounting for the expander and detection relay. Fewer degrees of freedom means less to drift.
The floor under your entire error budget. A 15 nm pointing specification is meaningless on a bench that moves by more than that.
You still need to see the sample, and you need to survive an invisible kilowatt-per-square-centimetre beam on an open bench.
For finding beads, video tracking, and fluorescence. Match sampling to the diffraction limit rather than buying pixels.
Filter the catalogue by sensor, cooling, frame rate, and interface against what your trap actually needs.
1064 nm is invisible and focuses to the retina efficiently. Wavelength-specific eyewear from a documented hazard analysis, not generic tinted glasses.
Enclose the beam path and the sample reflections. It also cuts air currents and acoustic pickup, so it pays for itself in stability.
If what you need is measurements rather than an instrument-building project, our MicroRheo platform arrives as a calibrated trapping and microrheology system with the laser, detection chain, mechanics, and analysis already specified against each other. Everything in this guide has already been decided once, defensibly.
Common questions
It sits inside the near-infrared window where water absorbs weakly and biological molecules absorb little, it produces low photodamage, and it is the wavelength where high-power, low-noise, single-mode sources are cheapest and most mature. Nd:YAG, Nd:YVO₄, and ytterbium fibre lasers all land there, and objectives, dichroics, and isolators are all readily available for it. It is not the gentlest wavelength or the best-detected one — it is the best overall compromise with the deepest component ecosystem.
Only for particles much smaller than the wavelength. In the Rayleigh regime, stiffness at fixed power and NA scales steeply with 1/λ⁴, so 830 nm would be roughly 2.7 times stiffer than 1064 nm on a 100 nm bead. For the 0.5–2 µm beads used in most biophysics, trapping efficiency is set by geometry and refractive-index contrast, and wavelength almost drops out. That is why wavelength is essentially never chosen for force — it is chosen for what the light does to the sample and what your detector can see.
A useful rule is about 1 pN of maximum trapping force per 10 mW delivered to the specimen plane on a micron bead. The catch is the word "delivered". A non-IR-corrected 100×/1.4 oil objective can transmit as little as 32 % at 1064 nm, and the isolator, expander, steering optics, and dichroic take their own cut. Budget for a factor of three to five between laser output and specimen plane, then add headroom so you are never running the laser at full current.
Silicon is nearly transparent at 1064 nm, so photons are absorbed deep in the substrate rather than in the depletion layer. Those carriers reach the junction by slow thermal diffusion, which acts as an unintended low-pass filter with a 3 dB frequency around 8–10 kHz. Ninety percent of the signal power is lost at 80 kHz. The fixes are an InGaAs detector, a position-sensitive detector built for the near infrared, heavy reverse biasing, correcting the measured spectrum with the published filter model, or moving to a wavelength below about 900 nm where the effect vanishes.
If you are holding, sorting, or positioning and never converting displacement into force, no — a stable multi-mode source with decent beam quality is fine, and the money is better spent on the objective. Low noise earns its cost the moment you quote a number in piconewtons, fit a power spectrum, or run a measurement long enough that drift competes with your signal. Microrheology, single-molecule force spectroscopy, and any active-feedback trap all fall in the second category.
Less than it appears. Almost all of that 20 MHz bandwidth lies far above anything a trapped bead in water can respond to, where the trap rolls off past its corner frequency of typically a few hundred Hz to a few kHz. A wideband figure can hide a large 1/f component in exactly the band you care about, or it can be dominated by a relaxation-oscillation peak that is harmless to you. Ask for the noise spectral density curve, or at minimum for an RMS figure over DC–10 kHz.
Not for coherence reasons — trapping does not care about linewidth, and back-focal-plane detection works with a few nanometres of bandwidth. You buy single-frequency operation to eliminate mode hopping and mode-partition noise, which show up as step changes in power and small pointing shifts that are indistinguishable from real events in a force trace. If you are building a dual-trap or interferometric instrument, or fibre-delivering the beam, the argument becomes stronger.
In vacuum, yes, and the standard result applies: intensity noise at twice the trap frequency parametrically drives the particle, with an energy e-folding rate of π²ν²S(2ν) where S is the fractional intensity noise spectral density. Keeping a 10 kHz trap warm-free for 100 seconds needs roughly 3 × 10⁻⁶ /√Hz. In water the particle is heavily overdamped, energy does not accumulate, and the same noise instead shows up as stiffness error and excess low-frequency position variance. Different mechanism, same conclusion about the laser.
Particle type and size, medium, objective, force range, and required bandwidth are enough for us to work backwards to a wavelength, a power at the aperture, and a defensible noise specification — including the optics that stand between the two.