physics-informed neural network-assisted compact modeling of utb-soi and nanowire mosfets for ultra-low power edge-ai applications

Physics-informed neural networks (pinns) offer a promising route to bridge device-level simulations and compact circuit models. in this work, we present a hybrid modeling framework that integrates tcad datasets with a baseline compact model and applies a pinn correction to capture stress-condition effects with high fidelity. the proposed approach achieves ≤ 2% route mean square error (rmse) across more than 2,000 bias points, maintaining stable predictions under temperature (273–373 k) and radiation (0–100 krad) variations. extracted berkeley short-channel igfet model (bsim) parameters enable direct spice simulation, ensuring compatibility with standard circuit design workflows. for deployment, the trained pinn is exported as a quantized onnx model, achieving sub-millisecond inference and ultra-low energy consumption (0.25 pj/op) on a cortex-m55 platform. this dual pathway supports both high-accuracy circuit simulation and real-time edge inference, making it suitable for embedded applications under constrained conditions. comparative analysis with recent ann-based models confirms that our physics-informed approach offers superior interpretability, spice readiness, and deployment efficiency. all datasets, code, and models are released to support reproducibility, benchmarking, and further research in compact modeling and edge-ai integration.