Physics-informed generative AI enforces hard constraints in semiconductor design

A new perspective paper from Yaser Mike Banad and Sarah Sharif argues that generative AI for semiconductor manufacturing must enforce hard physical constraints by construction, not through post-hoc filtering. The authors contend that in domains like lithography mask generation, synthetic defect data, and process recipe optimization, physically invalid outputs are not merely low quality but completely unusable. They survey an emerging architectural toolkit including physics-informed diffusion models, PDE-constrained variational models, neural-operator priors, and conservation-law-respecting generative networks. The paper identifies four integration patterns between generative models and physics-based simulators: sequential, parallel, embedded, and iterative feedback. The central claim is analytical: where physical validity is the binding criterion, architectures that enforce constraints by construction will outperform those that filter later.

The research agenda proposes three pillars: physics-fidelity benchmarks to measure constraint satisfaction, differentiable simulator infrastructure that allows end-to-end gradient flow, and multimodal foundation models that jointly handle mask layouts, process parameters, and device physics. The authors see semiconductor fabrication as the sharpest test case for this approach, but suggest the framework generalizes to any scientific domain with hard physical constraints. This work bridges machine learning, computational physics, and hardware engineering, potentially enabling AI to propose designs that are manufacturable by default rather than requiring extensive simulation-based validation. The full paper is available on arXiv (2606.11247).