In this work, we propose a physics-informed autoencoder, named PIAE, to extract features and reduce the dimensionality of sensor measurement data. This data consists of transistor transfer characteristics, that are drain current vs gate voltage measurements ( \({I_D}\) - \({V_{GS}}\) ) of carbon nanotube field-effect transistors (CNT-FETs) acquired during several gas-sensing experiments. The encoder of the PIAE is trained—via an additional loss term—to output four physically interpretable key features. These features correspond to typical transistor parameters: threshold voltage, subthreshold swing, transconductance, and ON-state current. The decoder module then reconstructs the CNT-FET transfer characteristics from these four key features. We evaluated our method on five different gas measurement runs performed with three distinct CNT-FET devices. The PIAE outperforms both a four-component principal component analysis and a four-parameter compact model in terms of reconstruction accuracy, achieving an average improvement of approximately 50% in the median root mean square reconstruction error. Moreover, the PIAE reaches a reconstruction accuracy comparable to that of a standard autoencoder while providing physical interpretability of the extracted features, opening new opportunities for sensor signal processing and analysis.
