Multimodal Current-Convolutional Neural Network Signal Processing System Improves Ion-Sensitive Field-Effect Transistor Sensor

Different from the human-derived electrical parameters’ extraction method to quantify the ion-sensitive field-effect transistor (ISFET) sensor, this work proposes an innovative multimodal current-convolutional neural network (MMC-CNN) signal processing system to achieve robust, accurate, and consensus ISFET analysis for direct pH prediction. The multimodal current includes drain-to-source current (${I}_{s}$), liquid, and back gate leakage current (${I}_{\lg }$,${I}_{\text {bg}}$) in addition to source-to-drain current (${I}_{d}$) from multiple terminals based on a double gate fully depleted silicon-on-insulator (FDSOI) FET. The deep learning algorithm convolutional neural network (CNN) model is jointly trained by (${I}_{d}$,${I}_{s}$,${I}_{\lg }$,${I}_{\text {bg}}$) and can predict pH autonomously without any manual calibration. Eleven divergent current configurations and generalization capability are studied to increase the interpretability and effectiveness of the MMC-CNN system. (${I}_{d}$,${I}_{s}$,${I}_{\lg }$) presents the best statistical performance with a high${R}^{{2}}$score up to 0.9969 (superior to 0.9672 through manual regression), low mean absolute percentage error (MAPE) down to 1.49%, and mean absolute error (MAE) down to 0.0775, whereas the${I}_{\text {bg}}$accession undermines these metrics substantially. The intrinsic relevance between (${I}_{d}$,${I}_{s}$,${I}_{\lg }$,${I}_{\text {bg}}$) and pH is revealed through COMSOL Multiphysics simulation and leakage current measurement, reasonably sustaining the MMC-CNN system. This work can be highly scalable in the multiterminal FET sensor and is anticipated to create a significant innovation in chemical and biological quantitative analysis.