Accelerated Prediction of Terahertz Performance Metrics in GaN IMPATT Sources via Artificial Neural Networks

This work investigates the application of artificial neural network (ANN)-based regression models to predict the static and dynamic characteristics of GaN impact avalanche transit time (IMPATT) sources in the terahertz (THz) frequency regime. A comprehensive dataset, derived from self-consistent quantum drift-diffusion (SCQDD) simulations of GaN IMPATT structures designed for a wide frequency range from the microwave frequency bands, up to 5 THz, is used to train the ANN models. The models effectively capture the impact of variations in structural, doping, and biasing parameters on device performance. The proposed ANN approach significantly reduces computational time for predicting breakdown characteristics, power output, and conversion efficiency properties of IMPATT sources, achieving similar accuracy to traditional SCQDD simulations while requiring only 7.8–20.1% of the computational time. Mean square errors are observed to be on the order of $10^{-4}$ – $10^{-6}$ , demonstrating the models’ high accuracy. Experimental validation shows strong agreement in terms of breakdown voltage, power output, and efficiency, supporting the potential of machine learning to streamline the design and optimization of high-frequency semiconductor devices.