Sensitivity Analysis and Performance Tradeoffs in Regression Neural Networks for Magnetic Field Sensing With Rectangular MOS Transistors

This study explores the effectiveness of nonlinear regression (LR) machine learning (ML) models and custom neural networks (NNs) for regression tasks for magnetic field sensing using a rectangular MOS transistor. We focus on sensitivity and average percentage error (APE), comparing various models under controlled conditions with a gate-to-source voltage ( ${V}_{\text {GS}}$ ) of 1.2 V, a drain-to-source voltage ( ${V}_{\text {DS}}$ ) of 1.8 V, and an applied magnetic field of 1.4 mT. The empirical model establishes a baseline sensitivity of 5.0%, but its instability poses a significant challenge to reliable sensor performance. In contrast, K-nearest neighbors (KNNs), random forest (RF), and decision tree (DT) models demonstrate stable sensitivities around 8%. Notably, custom NNs achieve the highest sensitivity, approximately 10%, with stable performance and consistently low APE values around 2%. Key performance metrics such as mean squared error (mse), mean absolute error (MAE), and latency were analyzed. The results show that custom NNs, particularly smaller architectures, offer a compelling alternative to traditional models like KNNs and DT, balancing accuracy, stability, and computational efficiency. This highlights the potential of custom NNs to enhance sensor performance in real-world applications where instability can significantly impact the accuracy and reliability of regression tasks.