Chemical sensors built from thin semiconducting films work because molecules landing on the surface add or remove electrons from the channel. This "surface-induced charge doping" changes the film's conductance, and every atom in the film is close enough to the surface to feel it. The same surface sensitivity is also why these sensors age. Adsorption sites become poisoned, contaminants accumulate, and slowly charging traps move the baseline. Over months, a classifier trained on fresh-sensor data stops recognizing the gases it once separated cleanly. Current fixes learn the drift from data, and they need labeled samples from later in the sensor's life. Their accuracy still falls as the sensor ages. On the 36-month Vergara gas-drift benchmark, the best context-tracking network reaches 0.764 accuracy in the final batch (Using Context to Adapt to Sensor Drift).
This note proposes a different approach. We take the physics of surface doping seriously enough to write down a small, bounded set of ways a sensor can plausibly drift. Then we train the classifier to win against the worst member of that set. This is worst-case (minimax) optimization: choose the design that does best when an adversary picks the most damaging parameter inside a bounded region. Wireless engineers use it routinely when channel estimates are uncertain. Here the adversary may not perturb the 128 input features arbitrarily. It may only lower each sensor's sensitivity, which models lost adsorption sites, and shift its baseline, within limits that grow with deployment time. Because this set has only about 32 dimensions and a clear physical meaning, the worst case can be computed exactly for linear classifiers. That yields a per-sample certificate: the prediction cannot change under any drift of that size.
Core Claim
A classifier trained once on early-life data, with a minimax objective over a physically structured drift set, will hold a higher worst-batch and late-batch accuracy than (a) nominal training and (b) robust training over generic norm-ball perturbations. This holds without any labels from later batches. The structured set is what keeps the "robustness tax" small. That tax is the loss of nominal accuracy that worst-case designs pay, about 4 dBm of extra transmit power in robust multicell beamforming (Worst-Case SINR Constrained Robust Coordinated Beamforming for Multicell Wireless Systems). The tax is small here because the adversary cannot spend its budget on perturbation directions the physics never produces.
The research file confirms that formal minimax optimization has not been applied to surface-doped chemical sensors. The nearest data-driven work tracks drift with a recurrent "context" network fed by labeled historical batches (Using context to make gas classifiers robust to sensor drift). The nearest optimization work handles linear radio channels with ellipsoidal error sets (Robust Cognitive Beamforming With Partial Channel State Information). We reuse the radio work's method: we turn an infinite "for all perturbations" constraint into a finite convex one through a dual-norm bound. We replace its uncertainty geometry with one derived from adsorption physics, and we target a different problem, classification under aging.
Research Program
The first component is the drift set. Each of the 16 sensors in the Vergara array contributes 8 features, a steady-state resistance change plus rising and decaying moving-average transients. Losing adsorption sites scales the adsorbed coverage, and with it every response feature of that sensor, by a common factor. Trap charging and contamination shift the baseline, which mainly enters through the steady-state and normalized-change features. We therefore model a drifted sample as $x' = Gx + Bm$. Here $G = \mathrm{diag}(g)$ applies one gain $g_s$ to all features of sensor $s$. $B = \mathrm{diag}(b)$ applies one offset $b_s$, and $m$ is a fixed 0/1 mask marking the offset-sensitive features. The drift set at elapsed time $t$ is
$$\mathcal{U}_t = \{(g,b) : 1-\rho(t) \le g_s \le 1,\ |b_s| \le \beta(t),\ s=1,\dots,16\},$$
where $\rho(t)$ and $\beta(t)$ are the largest sensitivity loss and baseline shift allowed after time $t$. The one-sided gain bound ($g_s \le 1$) encodes the physical fact that poisoning removes sites rather than adding them. We fit $\rho$ and $\beta$ as linear-in-time envelopes from drift observed among the training batches only, and extrapolate them forward. This is the standard "bounded rate of change" assumption, made explicit and testable.
The second component is the training objective. For a linear multiclass scorer with weight vectors $w_k$, the worst-case score gap between the true class $k$ and a rival $j$ has a closed form, because the inner maximization over a box is separable by sensor. Write $\Delta = w_k - w_j$, and let $a_s = \sum_{f \in s}\Delta_f x_f$ be the share of the score gap coming from sensor $s$. Then
$$\min_{(g,b)\in\mathcal{U}_t} \Delta^\top x' = \Delta^\top x - \rho(t)\sum_s \max(a_s,0) - \beta(t)\sum_s \Big|\sum_{f\in s}\Delta_f m_f\Big|.$$
This expression is concave in $w$, so the robust hinge loss built on it is convex. Training is then a single convex program, the counterpart of the second-order-cone reformulations in the beamforming papers. Applying the same formula to a test sample gives its certificate: the largest $(\rho,\beta)$ for which its predicted label provably cannot flip. For a small nonlinear network we solve the inner maximization numerically with projected gradient steps over the 32 drift parameters. That inner problem is cheap because it is low-dimensional. The resulting certificates are empirical rather than exact, and we will report them separately.
The third component is a physics check that decides whether the idea deserves to survive. Before any robust training, we use oracle labels to fit the best per-sensor gain and offset mapping batch 1 onto each later batch. We then measure how much of the class-conditional drift this mapping explains. If the gain-and-offset model explains little of the drift in late batches, the structured set is wrong. In that case the proposal reduces to generic robust training, and we would report that negative result rather than continue.
Scope matters. The Vergara sensors are metal-oxide chemiresistors. Their conductance is set by adsorbed species exchanging charge with a surface depletion layer, which is the same surface-doping mechanism discussed for graphene and nanotube transistors (Carbon Nanotube Field-Effect Transistor-Based Chemical and Biological Sensors). They are not low-dimensional channels, though. Extending the approach to competitive binding on nanoscale field-effect receivers, where interferers occupy finite sites (Flexure-FET-Based Receiver with Competitive Binding for Interference Mitigation in Molecular Communication), is a natural follow-up. We do not test it here, because no measured dataset for it exists in the file.
Baselines
The naive reference point is a nominal classifier: a multinomial linear model and a small feedforward network, both trained with ordinary loss on the training batches. Its accuracy shows how much drift costs when nothing is done about it.
The strongest existing method is the context-plus-skill network of Using context to make gas classifiers robust to sensor drift and Using Context to Adapt to Sensor Drift. A recurrent network summarizes recent labeled batches into a drift vector that modulates the classifier. Its reported results are 88.2% mean accuracy across test batches for the feedforward+context model, against 86.9% for a standard network and 76.4% for an SVM ensemble. In batch 10 it reports 0.764, against 0.693 for a neural-network ensemble. We rerun it on the same five gases, with toluene excluded as in those papers, using the same features and the same hyperparameter search budget as our models. We will compare under two protocols. In the first, which those papers use, the model is trained on all batches before the test batch, and this baseline enjoys labeled history. In the second, which is stricter, all models train only on batches 1–2 and are tested on batches 3–10 with no further labels. Beating this baseline means one of two things: matching its mean accuracy under the first protocol while exceeding its worst-batch accuracy, or clearly exceeding it on batches 7–10 under the second.
Two further baselines isolate the key ingredient. The first is the same robust objective with an isotropic $\ell_\infty$ or $\ell_2$ ball on the 128 features, with radius tuned so its nominal accuracy on the last training batch matches our method's. It answers whether the physical structure of the set matters, or whether any robustness helps. The second is drift-structured data augmentation, which samples random $(g,b)$ from $\mathcal{U}_t$ during ordinary training. It answers whether the worst-case optimization matters, or only the physically shaped perturbations.
Why not just…?
Why not just keep the recurrent context network and collect more labels? In implanted or field-deployed sensors, labeled recalibration batches are exactly what is unavailable. Even with labeled history, the context model loses more than ten points between its mean and its batch-10 accuracy (Using Context to Adapt to Sensor Drift). The label-free second protocol tests directly whether our method holds up where the context model has nothing to learn from.
Why not just re-normalize each new batch using its unlabeled statistics? This is the most serious threat. If drift really were per-sensor gain and offset, per-batch standardization would remove much of it at almost no cost. We include batch-wise standardization as a baseline and expect it to be strong. Our method differs in two ways that the experiments can check. It decides on single samples without waiting for a batch. It also does not break when the class proportions change between batches, which shifts batch statistics independently of drift. We will create class-proportion shifts by subsampling test batches. If standardization still matches us under those shifts, we will reframe the contribution as robustness of the residual left after standardization and train our set on that residual.
Why not just augment with random drift instead of solving a minimax problem? Augmentation optimizes average performance over the drift set, which may be enough if the dangerous drifts are common. The ablation above settles this. Minimax should matter most for worst-batch accuracy and certified accuracy. If augmentation matches it on both, the worst-case machinery is not earning its place, and we will say so.
Why not just optimize expected accuracy under a fitted distribution of drift? With roughly ten batches, any fitted drift distribution rests on about ten observations, so its tails are guesses. A bounded envelope needs only a rate estimate. Drift-structured augmentation is the practical version of this alternative and is already in the comparison.
Proposed Experiments
All experiments use the Vergara 36-month gas-drift data with the five-gas protocol of the context papers. The metrics are per-batch accuracy, mean accuracy, worst-batch accuracy, batch-10 accuracy, and certified accuracy: the fraction of test samples correct with a certificate covering the extrapolated $\mathcal{U}_t$. We also report nominal accuracy on the last training batch, which serves as the robustness tax.
The first decisive check is the physics check described above. The second checks coverage: whether the extrapolated envelopes $\rho(t)$ and $\beta(t)$, fitted on batches 1–2, actually contain the oracle-fitted drift of later batches. If they under-cover, the certificates are void. The third is the head-to-head comparison under both protocols against the nominal model, the context network, batch standardization, the isotropic-ball ablation, and the augmentation ablation.
The proposal succeeds if three things hold together. Under the label-free protocol, the structured minimax classifier beats every baseline on worst-batch and batch 7–10 accuracy. Its nominal accuracy stays within a few points of nominal training. And the isotropic ball, at matched nominal accuracy, does not reach the same late-batch accuracy. That combination would show that physics-shaped worst-case sets buy long-life robustness at a small tax, which neither generic robustness nor learned drift tracking currently provides.
References
- Using Context to Adapt to Sensor Drift
- Worst-Case SINR Constrained Robust Coordinated Beamforming for Multicell Wireless Systems
- Using context to make gas classifiers robust to sensor drift
- Robust Cognitive Beamforming with Partial Channel State Information
- Carbon Nanotube Field-Effect Transistor-Based Chemical and Biological Sensors
- Flexure-FET-Based Receiver with Competitive Binding for Interference Mitigation in Molecular Communication