Abstract
This paper proposes a physics-constrained neural network (PCNN) framework to model displacement damage-induced degradation in AlGaN/GaN HEMTs. The architecture combines feed-forward neural networks (FFNN) with Monotonic Block to enforce the physical correlations between bias voltages, radiation fluence, and drain-source current degradation. A hybrid optimization strategy integrates data-driven learning with differentiable physics regularizations, incorporating analytical constraints based on displacement damage mechanisms to guide training. This ensures both accurate interpolation within the experimental fluence range and physically consistent extrapolation beyond it. Smoothing operators are introduced to enhance the continuity of predicted device characteristics. Experimental validation using 7.6 MeV carbon-ion and 30 MeV fluorine-ion irradiation confirms that the model achieves high accuracy (RMSE < 0.055%) and reliably predicts key physical parameters within the training range (RMSE < 6.5%). Notably, the model maintains reasonable physical trends even at untrained fluence levels. By integrating machine learning with device physics, the proposed framework provides a compact, generalizable approach for modeling radiation effects, offering critical insights for radiation-hardened design and lifetime prediction of GaN-based MMICs in space applications.

This publication has 50 references indexed in Scilit: