Data-Driven Device Design Framework for Semiconductor Manufacturing With Dual-Hierarchy DNN Prediction Scheme and PSO

Abstract

Optimizing devices in integrated circuits (ICs) remains challenging due to the complexity of modern manufacturing. Traditional methods like technology computer-aided design (TCAD) require extensive human intervention. Meanwhile, machine learning (ML)-based approaches provide automation but frequently struggle with real-world manufacturing variations across the full fabrication workflow. This work proposes a loop optimization framework composed of three key components: an inverse model for device parameter prediction, a forward model as a TCAD surrogate, and a weighted particle swarm optimization (PSO) algorithm. Trained on real manufacturing data that includes manufacturing variations, the surrogate model better reflects actual fabrication conditions. Compared to TCAD simulations, the trained forward model achieves a computational speedup of nearly 40000 times. Furthermore, by leveraging the inverse model to constrain optimization and a fully adjustable weighting mechanism, the framework enables precise control over optimization intensity and solution reliability, ensuring adjustable tradeoffs. It also supports customizable weight adjustments for different electrical characteristic metrics, allowing users to prioritize specific characteristics as required. Overall, this work provides an efficient and flexible tool for balancing tradeoffs and optimizing semiconductor device design.