Tracing radiation-induced degradation in bipolar junction transistors: a novel predictive data-driven framework

Abstract

Bipolar junction transistors’ (BJTs’) dependability in radiation-hardened electronics, nuclear instrumentation, and space systems is adversely affected by total ionizing dose (TID)-induced degradation, which presents a serious obstacle to the longevity and functionality of the device. Proactive maintenance and efficient reliability assessment depend on the accurate prediction of such degradation. This research tackles this issue by creating a thorough data-driven framework that makes use of sophisticated supervised machine learning (ML) models, such as light gradient boosting machine, extreme gradient boosting, and categorical boosting (CatBoost), in addition to ensemble techniques like Stacking and Voting regressors. An 80/20 train-test split and rigorous fivefold cross-validation were used to ensure model robustness, and a carefully selected experimental dataset of 565 data points from different BJT types was used. The metaheuristic pufferfish optimization algorithm (POA) was used to systematically perform hyperparameter tuning, which significantly improved predictive performance. With a test R² of 0.9827, RMSE of 0.0926, and MAE of 0.0628, the POA-Voting model outperformed the rest of the models in terms of accuracy. The models showed accurate and dependable degradation forecasts, continuously keeping mean absolute percentage errors (MAPE) below 2.1%. Comparative studies demonstrated POA’s superior hyperparameter optimization over a genetic algorithm, while SHAP analysis validated the dominant influence of total ionizing dose on degradation. Real-time monitoring, prognostics, and improved device design in crucial radiation-exposed applications are made possible by the resulting ML pipeline, which provides an interpretable and precise tool for predicting radiation-induced transistor degradation.