Deep-Learning Model for Buildup of Ionization Defects in Bipolar Junction Transistors

The ionization defects in bipolar junction transistors (BJTs) are sensitive to ionization dose rate and molecular hydrogen in oxides. The bimolecular reaction model, a physical model with self-coincidence forms, is established in earlier works to describe formations of ionization defects in BJTs. This work proposes a deep-learning (DL) model with an artificial neural network (ANN) framework to calculate the densities of ionization defects. The ANN model can be trained using data prepared by the physical model in conjunction with numerical calculations. Case studies with a dataset of
$\sim 6.0 \times 10^{5}$
points suggest an optimized ANN model which is composed of four hidden layers and 16 exponential linear unit (ELU)-activated neurons per hidden layer. This model provides excellent fits to the data from the physical model for the dose, dose rate, and hydrogen dependence of ionization defects in gate-controlled p-n-p BJTs. Our encouraging results show that the DL modeling methodology may find its effective application in the dynamics of ionization defects.