Cross-compensation of FET sensor drift and matrix e effects in the industrial continuous monitoring of ion concentrations

Abstract Field-effect transistor (FET) sensors are attractive potentiometric (bio)chemical measurement devices because of their fast response, low output impedance, and potential for miniaturization in standard integrated circuit manufacturing technologies. Yet the wide adoption of these sensors for real-world applications is still limited mainly due to temporal drift and cross-sensitivities that introduce considerable error in the measurements. In this paper, we demonstrate that these two main non-idealities can be corrected by joint use of an array of FET sensors - selective to target and major interfering ions - with machine learning (ML) methods - including state-of-art deep neural networks (DNNs) - in order to accurately predict ion concentrations continuously and in the field. We studied the predictive performance of the ML models when monitoring pH from combinatorial H+, Na+, and K+ ion-sensitive FET (ISFET) sequences of readings collected over a period of 90 consecutive days in real water quality monitoring conditions. The proposed algorithms were trained against reference online measurements obtained from a commercial pH sensor. Results show a greater capability of DNNs to provide precise pH measurements for longer than a week when fusing H+ with Na+ or K+ ISFET readings, achieving a relative root-mean-square error reduction of 73% over standard two-point pH FET sensor calibration methods.