ML-based classification and information-theoretic metrics of instrument drift in laser-based sensing

Laser-based sensing is commonly used for trace-gas detection and density measurements in environmental, biomedical, and industrial processing applications. Molecular spectroscopy in the mid-infrared region provides sensitive measurements with high selectivity of multiple molecules in a congested spectrum. Therefore, an obvious advantage in detection in the mid-IR region is the fundamental absorption bands of several molecular species. Field instruments with optomechanical multipass cells are prone to drifts emanating from thermal impacts that may limit the measurement sensitivity and precision required for many atmospheric sensing applications. In this paper, we show an integrated information-theoretic and machine learning-based classification method to identify and discriminate the efficacy of instrument thermal drifts. We use the change in the entropy or information as a cumulative measure of the mismatches between the predicted and validation models. We show that entropy is higher when classification mismatches or losses are high, serving as a unique quantitate metric of classification methodology.