Ask for a bandpass filter datasheet and you will get a centre wavelength, a bandwidth, a peak transmission and a blocking range. All four are quoted at 23 degrees C, normal incidence, in air. Then the filter goes into an instrument that operates from minus 20 to plus 60 degrees C, and the channel that was centred at 656.3 nm is now somewhere else. The datasheet was not wrong. It was answering a different question.

Where the drift comes from
Three mechanisms add up. The substrate's refractive index changes with temperature, and its thickness changes through thermal expansion. The thin-film materials do the same, with coefficients that are larger and less well documented than the glass. And the mounting applies stress, which changes both the physical thickness and, through the photoelastic effect, the index.
For a typical 10 nm bandpass filter on B270, the combined shift runs 0.01 to 0.03 nm per kelvin. Narrower filters shift less in absolute terms but the same relative amount, so a 1 nm filter in a fluorescence channel moves a meaningful fraction of its own bandwidth over a 60 K excursion. Long-pass edges behave similarly, and for a steep edge the shift is more damaging than it is for a flat-topped band, because the edge position is the specification.
Why fused silica is not automatically the answer
The instinct is to specify fused silica for thermal stability. Fused silica has very low expansion, which helps, but its refractive index still changes with temperature at roughly ten parts per million per kelvin, and that term dominates. Some borosilicate glasses have a lower net dn/dT because the expansion and index terms partially cancel. The right substrate depends on the coating design and the operating band, not on a general reputation.
What fused silica does deliver reliably is thickness stability and low stress birefringence, which matters for polarisation-sensitive channels. That is a different argument, and it should be made separately.
Athermal design is a coating problem
A coating can be made less sensitive by design. Mixing materials with opposite dn/dT signs in the high-index layers, adjusting the ratio of layer thicknesses that carry the resonance, and shifting the design so that the temperature coefficient of the effective index is reduced all work to some degree. The cost is usually bandwidth or steepness: an athermal design is a compromise, and the compromise should be stated rather than hidden.
For instruments that cannot accept any compromise, the answer is active: a thermoelectric cooler holding the filter at a fixed temperature, or a heater ring around the mount. That converts a stability problem into a power and control problem, which is often the right trade in a laboratory instrument and the wrong one in a field-deployed sensor.
What the mounting contributes
A filter clamped between two metal rings will see stress as the assembly expands. A 25 mm filter in an aluminium mount, heated 40 K, sees the mount grow by about 45 µm; if the filter is constrained, that growth becomes compressive stress in the glass, and the coating's centre wavelength moves. The effect can be larger than the intrinsic thermal drift, and it is entirely avoidable.
Compliant mounting, an elastomer O-ring rather than a rigid clamp, and a seat that allows radial slip remove most of it. For high-reliability instruments, the mount material should be chosen to match the glass expansion rather than the housing.
Measuring drift properly
A chamber with optical ports, a stable reference source and a spectrometer that is itself thermally compensated. Three soak points at minimum, typically minus 40, 25 and 85 degrees C, with two hours at each to reach equilibrium. Measurement at the same point on the filter every time, because centre wavelength varies across the aperture. And a return to 25 degrees C at the end of the run to check hysteresis, which reveals whether the coating or the mount took a permanent set.
The output should be two numbers, not one: the slope in nm per kelvin, and the hysteresis in nm. A filter with a modest slope and no hysteresis is easier to compensate in firmware than one with a small slope and 0.5 nm of hysteresis, because the latter cannot be corrected with a single coefficient.
Putting it in the RFQ
The line that gets results reads: centre wavelength 656.3 nm plus or minus 1 nm at 23 degrees C, thermal coefficient to be reported and not to exceed 0.02 nm/K between minus 20 and plus 60 degrees C, hysteresis below 0.1 nm after one full cycle, measured at the aperture centre with the filter in its final mount.
Suppliers who can answer that line are the ones who have actually put a filter in a chamber. The rest will send the same 23 degrees C datasheet they sent last time, and the instrument will drift in the field for reasons that take a year to diagnose.
Product page: short-pass and band-pass filters — specifications, coating options and lead times.