Diagnosing and correcting three-year drift in metal-oxide chemosensor arrays

Abstract

Metal-oxide chemosensor arrays lose analytical validity over time because their sensing layers age, so that a calibration model built today misclassifies samples measured months later. This work characterises that drift and tests what can be done about it without acquiring new reference standards, using a public archive of 13 910 measurements recorded with a sixteen-element array over 36 months for six volatile analytes between 1.0 and 1000 ppmv, each described by eight descriptors per sensor. A variance decomposition shows that the between-session mean square of a typical descriptor exceeds its within-session residual mean square by a median factor of 123, and that the session-to-session shift is shared by the six analytes (median pairwise correlation 0.52, all 15 pairs positive) and confined to a low-dimensional subspace, two components accounting for 79.7% of its variance. Drift susceptibility is not shared equally: the steady-state resistance change and the slowly smoothed decaying-transient descriptors are the most robust, the baseline-normalised steady-state descriptor and the rising-transient descriptors the most fragile, a 3.0-fold difference. Filtering the descriptors on that ranking does not, however, improve prospective accuracy, so the ratio is a diagnostic rather than a variable-selection criterion. Three signal transforms, eight correction strategies and three classifiers were compared under strictly forward-in-time protocols in which a model never sees a measurement session later than those used to calibrate it. Without correction, six-analyte recognition falls from 99.1% within the calibration session to 41.8% over the following three years. Session-wise autoscaling, which needs neither reference standards nor labels nor a drift model, recovers a large part of that loss; component correction performs comparably but needs a drift model, and orthogonal signal correction is worse than no correction at all.