From Challenge to control: bridging semiconductor process complexity with science-based AI

As semiconductor manufacturing continues to grow in complexity, conventional physics-based modeling approaches—while essential—struggle to meet the demands of multi-physics, multi-scale process integration. In contrast, purely data-driven AI methods offer speed and flexibility but often lack physical grounding, interpretability, and robustness. This talk explores how science-based AI can help bridge this gap by embedding domain knowledge, physics constraints, and uncertainty quantification into machine learning frameworks.

We will focus on three key challenges in semiconductor process modeling: (1) achieving comprehensive coverage across the nine major process modules, (2) capturing multi-scale and multi-physics interactions, and (3) adapting to hidden or evolving process conditions. Through selected examples—including real-time TCAD surrogates, inductive bias for etching, variability-aware modeling, hidden physics discovery, and layout-aware defect and warpage prediction—we demonstrate how science-based AI approaches can enhance both efficiency and physical consistency.

The talk will close with reflections on emerging directions, including the role of physics-aware AI in enabling digital twins and more integrated, adaptive design–process co-optimization.