Predicting the damage effect of the total ionizing dose of bipolar junction transistors based on multi-strategy fusion optimization

This study proposes a multi-strategy fusion optimization prediction model that integrates the collaborative optimization of genetic algorithm-optimized conditional tabular generative adversarial networks (GA-CTGAN) and dynamic exponentially weighted squared error-extreme gradient boosting (DEWSE-XGBoost) to address the challenge of low prediction accuracy associated with small-sample data in predicting damage induced by total ionizing dose effects on bipolar junction transistors (BJTs). Initially, Spearman correlation analysis was employed to identify and eliminate input features with low correlations, thereby reducing redundant information. Following this, secondary interpolation preprocessing was applied to the input features using a linear interpolation method, while the output features were optimized through a double exponential smoothing technique to enhance data quality and analytical reliability. Subsequently, genetic algorithms were utilized to globally optimize the hyperparameters of the conditional tabular generative adversarial networks, resulting in the generation of high-quality synthetic data that align with real distributions under equilibrium conditions, thus alleviating the limitations posed by small sample sizes. In addition, a dynamic exponentially weighted squared error loss function was incorporated into the extreme gradient boosting (XGBoost) framework to emphasize learning from challenging samples. Comprehensive validation was performed using irradiation experimental data from 3CK3B-type BJTs as well as publicly available literature datasets. The experimental results indicate that the expanded-sample prediction model achieves a mean absolute error of less than 0.06, reflecting an accuracy improvement of 87% to 96% compared to conventional artificial neural networks and baseline XGBoost models trained on the original small-sample data. This methodology offers both data-driven insights and theoretical foundations for the optimization of radiation-hardened designs in BJTs devices.