Bipolar junction transistors (BJTs) are susceptible to total ionizing dose (TID) effects in radiation environments, leading to current gain degradation. Addressing the initial-value sensitivity issue in solving the two-dimensional implicit coupled equations within the BJT transceiver current analytical model based on symmetric double-gate MOSFET surface potential theory, this study proposes a machine learning-based BJT surface potential prediction method. This method integrates a classification model, a sample generation model, and a regression model to achieve accurate prediction of BJT surface potential under TID damage. For the sample generation model, a residual Transformer variational autoencoder is designed to enhance the efficiency of generating effective samples. For the regression model, a dynamic exponential weighted mean squared error function is adopted as the loss function for XGBoost to improve the model’s generalization capability. Simulation results demonstrate that at cumulative total doses of 50 and 100 krad(Si), the average relative error between predicted over-base currents and irradiation test results is as low as 0.1847 and 0.2619, respectively, confirming the accuracy of the proposed prediction method.
Liu et al. (Fri,) studied this question.