Utilizing machine learning for developing equivalent circuit-free calibration plots in impedimetric sensors

This is the authors' abstract. We don't add key points for this paper.

• Developed calibration for impedimetric sensor without using equivalent circuit. • Calibration developed using machine learning models trained on sparse datasets. • Machine learning models trained on impedance magnitude showed the best performance. • Validation performed using impedance spectrum simulated with equivalent circuit. • Support vector regression achieved the best results, as measured by R² and MAPE. Electrochemical impedance spectroscopy (EIS) is a powerful technique for developing highly sensitive electrochemical sensors, providing comprehensive characterization by identifying perturbations caused by target analytes. However, analyzing EIS spectra is complex due to the convolution of various relaxation processes. The conventional approach of fitting the EIS spectrum to an equivalent circuit model (ECM) for calibration curves has limitations, including difficulty selecting an accurate ECM, the potential introduction of biases, and the loss of information by focusing solely on charge transfer resistance. This manuscript proposes using machine learning (ML) models to create calibration curves for two impedance datasets (named configuration 1 and configuration 2) obtained from microfluidic electrochemical gas sensors with two electrode configurations. We evaluated the performance of various ML models, including linear regression, support vector regression (SVR), gaussian process regression, decision tree and ensemble-based models. Additionally, we examined the impact of different impedance data types (real and imaginary parts, magnitude, and phase) on model performance. The results show that the SVR models trained on the magnitude of impedance showed the best performance for concentration prediction for both datasets. For configuration 1, the SVR model with radial basis function kernel showed the best performance with an R 2 of 0.95 and mean absolute percentage error ( MAPE) of 11.2 %. For configuration 2, the SVR model with a sigmoid kernel showed the best performance with an R 2 of 0.965 and MAPE of 10 %.