Physics-informed machine learning model for bias temperature instability

Abstract

A new machine learning model is presented to predict the dynamic behavior of threshold voltage shifts induced by bias temperature instability (BTI) in CMOS devices. The model is constructed by combining physical theories with machine learning such as an artificial neural network and a Gaussian mixture model (GMM). To enlarge the capture–emission energy (CEE) window and to perform independent estimations of two distinct components of CEE distribution, the GMM with soft clustering is utilized, enabling full lifetime modeling of BTI. By training the CEE map with the consideration of the occupancy probability of traps and then executing the integration along the CEE, the threshold voltage shifts are obtained. This approach forms data-driven modeling that naturally encodes underlying physical theories as prior information. The resulting model exhibits a good performance for predicting the dynamic characteristics of BTI under various stress-recovery conditions.