Predicting Evanescent-Wave Gas Sensor Selectivity from First Principles

On-chip infrared gas sensors guide light along a silicon waveguide. A thin tail of that light, the evanescent field, leaks a fraction of a micron into the surroundings, and molecules sitting in that tail absorb some of it at their characteristic vibration wavelengths. The tail is weak and gas is dilute, so the signal is small. One fix is to coat the waveguide with a porous material that soaks up and concentrates the target gas right where the light is. A 2021 study did this with mesoporous silica nanoparticles. These are silica grains riddled with roughly 2 nm pores whose walls carry silanol (Si–OH) groups. The coating raised ethanol absorbance 15-fold. The authors then ran density functional theory (DFT), a quantum-mechanical method that computes molecular energies from the electron density, to explain afterward why ethanol sticks to silanols more strongly than methane does (Surface Functionalization Utilizing Mesoporous Silica Nanoparticles).

This proposal turns that after-the-fact explanation into a forward prediction. Adsorption energy alone does not determine the optical signal. What the detector sees is a product of three factors: how many molecules the coating holds, how much of the light overlaps the coating, and how strongly each adsorbed molecule absorbs at the laser wavelength. That third factor is not the same as in the gas phase, because hydrogen bonding to a silanol shifts and reshapes the molecule's C–H and O–H stretch bands. DFT can compute both the uptake factor and the absorption factor. A simple mode solver computes the overlap. We propose to combine them into a signal model that ranks analytes, coating chemistries and humidity conditions before anything is fabricated, and to test it against new sensor measurements.

Core Claim

A per-analyte signal model built from DFT adsorption free energies and DFT-computed adsorbed-state infrared cross-sections will predict the relative enhancement factors of coated evanescent sensors. "Relative" here means across analytes, across surface functionalizations, and under competing water vapor. These predictions should beat binding-energy-only ranking, which is the logic used in the 2021 study. In the low-coverage (Henry's law) regime, the enhancement of a coated waveguide over a bare one at laser wavelength $\lambda$ is

$$\eta(\lambda) = \frac{\Gamma_{c}\, K_H\, \sigma_{ads}(\lambda)}{\Gamma_{0}\, \sigma_{gas}(\lambda)}, \qquad K_H \propto \exp\!\left(-\Delta G_{ads}/k_B T\right),$$

The symbols are defined as follows. $\Gamma_c$ and $\Gamma_0$ are the fractions of modal power in the coating and in the bare cladding, both from a mode solver. $K_H$ is the ratio of adsorbed to gas-phase concentration, and it depends on the adsorption free energy $\Delta G_{ads}$ at temperature $T$. $\sigma_{ads}$ and $\sigma_{gas}$ are the absorption cross-sections of the adsorbed and free molecule at $\lambda$.

We will be honest about precision. A 1 kcal/mol error in $\Delta G_{ads}$ moves $K_H$ about fivefold at room temperature, so absolute enhancement predictions will be crude. The claim is about ratios. When two analytes, or an analyte and water, are compared on the same silanol site, errors from the approximate functional partly cancel. The cross-section term can also flip a ranking that binding energy alone would get wrong. A strongly bound molecule whose band red-shifts away from the laser line is one example.

Research Program

The computational arm models silanol and functionalized-silanol sites. The functionalizations are amine, alkyl and thiol silanes, which the 2021 study named as untested. We use cluster and periodic amorphous-silica models with a dispersion-corrected functional, because plain gradient-corrected functionals miss the weak dispersion forces that dominate methane and alkane binding. Each analyte–site pair yields three things: the adsorption energy; harmonic and anharmonically corrected vibrational frequencies and intensities, which give $\sigma_{ads}(\lambda)$ after broadening calibrated on one reference analyte; and vibrational partition functions, which convert energy into $\Delta G_{ads}$. Competitive adsorption with water is handled by computing water on the same sites and using a two-component Langmuir form. The analyte panel will be methane, ethanol, methanol, acetone, toluene, hexane and water.

The optical arm computes $\Gamma_c$ for one to several bilayers of coating. We treat the coating as an effective porous medium. Its scattering loss per bilayer is measured once, not modeled, because DFT has nothing to say about scattering from 159 nm particles. This makes the model predict an optimal bilayer count where enrichment gains meet attenuation, a trade-off the 2021 study left unexplored.

The experimental arm reproduces the 2021 coating protocol (layer-by-layer deposition, calcination) on Si waveguides. It extends that protocol to functionalized particles and multiple bilayers, and measures absorbance with a tunable mid-infrared laser across the analyte panel at controlled relative humidity.

Two other lines of work also inform the design. The crown-ether lead sensor shows that receptor selectivity in photonic sensors is still chosen empirically (Crown ether decorated silicon photonics). It is a natural later transfer target, but aqueous ions add solvation errors that we deliberately avoid in the first study. Machine-learned surrogates such as Deep-Learning DFPT and the equivariant polarizability model of Filling et al. show that response properties can be learned. They matter here only if direct DFT becomes the bottleneck when scaling to many candidate coatings.

Baselines

The naive reference is the bare waveguide, with enhancement fixed at one. Close behind is an analyte-independent geometric model in which enhancement equals $\Gamma_c/\Gamma_0$ times a porosity factor. If this model ranks analytes as well as ours, chemistry adds nothing.

The strongest existing method is the binding-energy-only ranking used in the 2021 study. It reports ethanol absorbance rising from 0.00537 to 0.0802 (15-fold) at 3350 nm, and methane from 0.005 to 0.0331 (6-fold) at 3320 nm. The same summary also mentions a "modest 2-fold" methane increase, and we will remeasure methane to resolve this inconsistency. To keep the comparison fair, we implement this baseline with the same DFT level, the same site models and the same computed $\Delta G_{ads}$ as our model. That makes it the ablation that removes the $\sigma_{ads}/\sigma_{gas}$ term. A second ablation keeps $\sigma$ but drops the water competition term.

An empirical upper reference replaces DFT with measured adsorption isotherms on the same particles plus gas-phase library cross-sections. Our model beats the binding-only baseline if its Spearman rank correlation with measured enhancement, across the roughly 30 analyte–coating combinations, is higher with a bootstrap interval excluding zero difference. Its log-ratio error should also be smaller and should approach the empirical reference.

Why not just…?

Why not just rank receptors by DFT binding energy, as the 2021 study did? Binding strength sets uptake but not how much light each adsorbed molecule removes at a fixed laser line. Hydrogen bonding shifts O–H and C–H bands by tens of wavenumbers, which is comparable to the band offsets that already separate methane's response at 3320 and 3350 nm in that study. The ablation above settles this directly. If removing $\sigma_{ads}$ costs no ranking accuracy, the simpler method wins, and we will report it as such.

Why not just measure isotherms and use tabulated gas-phase spectra? This is a real threat for coatings that already exist. It is our empirical reference, and it may match or beat DFT on known chemistries. The value of DFT lies in screening functionalizations before synthesis. We test that by predicting the amine, alkyl and thiol coatings blind, before fabricating them.

Why not just use a pretrained catalysis potential instead of DFT? Models trained on OC20 and OC22 cover metal and oxide catalyst slabs. They do not cover hydroxylated amorphous silica with organic ligands, and the file notes their errors grow out of distribution. They also give no vibrational intensities. We will run one as a secondary adsorption-energy baseline, but we do not expect it to rank silanol chemistry reliably.

Proposed Experiments

The first check reproduces the published ethanol/methane enhancement ratio from the model without fitting to it. The decisive test is the blind prediction, registered before fabrication, of rank order and magnitude across the seven-analyte panel on bare, silanol and three functionalized coatings. The third check predicts the drop in ethanol response at 30%, 60% and 90% relative humidity from the computed water competition, compared against measurement. The fourth predicts the optimal bilayer count from the measured loss per bilayer and the modal overlap, then fabricates one to four bilayers to confirm it.

Some outcomes would falsify the proposal. One is no rank-correlation gain over binding-only ranking. Another is humidity predictions off by more than an order of magnitude, which would show that the competition model is too crude. In either case the method reduces to empirical screening.