Physics-informed generative AI for semiconductor manufacturing: Enforcing hard physical constraints in generative models by construction

Researchers propose physics-informed generative AI architectures that enforce hard physical constraints by construction rather than post-hoc filtering, using semiconductor manufacturing as a test case. The work surveys emerging techniques including physics-informed diffusion models, PDE-constrained variational approaches, and conservation-law-respecting networks to ensure generated designs, data, and processes are physically valid rather than merely plausible.

This perspective article addresses a fundamental gap in how generative AI is applied to physically-constrained domains. While diffusion models and large language models excel at producing perceptually plausible outputs, semiconductor manufacturing demands absolute physical validity—invalid mask designs or process recipes cannot simply be discarded as low-quality outputs but render fabrication impossible. The research identifies that current generative approaches treat physical constraints as post-processing filters rather than architectural requirements, a distinction with material consequences in high-stakes manufacturing environments.

The work connects multiple technical threads: differentiable lithography simulators, TCAD (Technology Computer-Aided Design) integration, and process simulation become not just validation tools but active components of model training. By embedding physics directly into generative architectures—through physics-informed diffusion operators, neural operator priors, and conservation-law-respecting networks—researchers can guarantee outputs satisfy fundamental constraints before generation completes. This represents a methodological shift from filtering invalid samples to constructing only valid solutions.

For semiconductor manufacturing, this approach directly impacts yield optimization, defect prediction, and process recipe discovery. Fab operators can trust generated synthetic data and control strategies without downstream verification delays. The framework extends beyond semiconductors to any constrained physical domain where validity is non-negotiable: drug discovery, structural engineering, or materials design. The article proposes developing physics-fidelity benchmarks and multimodal foundation models specifically trained on manufacturing data, suggesting industrial AI applications will increasingly demand constraint-by-construction approaches rather than constraint-by-filtering strategies.

• →Physics-informed generative models enforce hard constraints by architectural design rather than post-hoc filtering, critical for semiconductor manufacturing validity.
• →Integration patterns between generative models and physics-based simulators create differentiable end-to-end pipelines for design and process optimization.
• →Emerging techniques include physics-informed diffusion, PDE-constrained variational models, and conservation-law-respecting networks adapted for manufacturing constraints.
• →Current benchmark standards lack physics-fidelity metrics, necessitating new evaluation frameworks specific to constrained physical domains.
• →This architectural paradigm applies broadly to any domain where physical validity is non-negotiable, including materials design, drug discovery, and structural engineering.