From ML to physics informed neural networks: scalable digital twin for stress-aware design in photonic structures

The semiconductor industry increasingly relies on accurate stress prediction for device optimization, particularly as advanced nodes incorporate strain engineering for performance enhancement. Traditional Technology Computer- Aided Design (TCAD) simulations, while accurate, face computational bottlenecks when applied to large-scale design space exploration. This paper presents a comprehensive study of machine learning-based simulation surrogates for predicting mechanical stress in semiconductor front-end processes. We evaluate three model architectures-XGBoost, Gradient Boosting Regressor (GBR), and Residual Multi-Layer Perceptron (ResMLP)-on both 2D planar and 3D FinFET structures simulated using FLOOPS [7] (Florida Object-Oriented Process Simulator). Our results demonstrate that the ResMLP architecture achieves superior performance with a mean absolute error (MAE) of 2.97 MPa and R2 of 0.9999 for interpolation tasks, while providing the fastest inference time of 0.7 ms. We further analyze model behavior under extrapolation conditions, revealing that neural network-based approaches degrade more gracefully than tree-based methods when predicting outside training bounds. The methodology establishes a framework for developing digital twins that can accelerate stress-aware semiconductor design workflows by orders of magnitude while maintaining prediction accuracy suitable for first-pass design optimization.