Technical Summary: Physics-informed Generative AI for Semiconductor Manufacturing

Problem Statement

The paper identifies a critical mismatch between current generative AI paradigms and the requirements of physical systems, specifically semiconductor manufacturing. While generative models (e.g., diffusion models, LLMs) excel in perceptual domains where "plausibility" is the primary metric, they fail in physical domains where "validity" is binary. In semiconductor manufacturing, a generated mask, layout, or process recipe that violates hard physical constraints (lithography, transport, reaction, device physics) is not merely low-quality; it is unusable and results in yield loss.

The core problem is that standard generative models trained on historical data learn the appearance of valid designs rather than the governing equations that define them. This leads to three specific failure modes:

1. Plausible but Invalid Outputs: Models generate geometries that look like training data but violate minimum printable dimensions, process windows, or design rules.
2. Loss of Causal Fidelity in Synthetic Data: Synthetic defect data generated by purely data-driven models fails to shift correctly when underlying process physics drift (e.g., tool wear, new consumables), causing downstream classifiers to fail in production.
3. Inverse Design Traps: Optimization over learned manifolds produces solutions optimal for the model but infeasible for physics, as the model has not encoded the true physical Pareto frontier.

The paper argues that post-hoc filtering (generating samples and rejecting invalid ones via simulation) is insufficient because it is computationally expensive, does not improve the generator's internal representation, and fails to scale.

Methodology and Architectural Framework

The paper proposes a shift from "physics-as-penalty" to "physics-informed by construction." This means encoding physical laws directly into the generative architecture so that valid outputs are guaranteed by the model's structure, not just by a loss function or a filter.

The authors define a spectrum of physics coupling (Table 2), distinguishing between:

• Post-hoc Screening: Generator and physics solver are separate; infeasible samples are rejected after generation.
• Physics-Loss Regularization: Physics residuals are added to the training objective, biasing but not guaranteeing validity.
• Simulator-in-the-Loop: Differentiable solvers guide the sampling process (e.g., via gradients), steering the generation trajectory toward feasibility at every step.
• Hard-Constrained Generation: The output parameterization itself is constrained to satisfy physical laws by construction.

The paper surveys four key architectural toolkits that enable this "by construction" approach:

1. Physics-Informed Diffusion Models: Embedding PDE residuals or physics-based scoring functions directly into the denoising trajectory. This ensures every intermediate step is biased toward physical validity, improving convergence and final distribution quality.
2. PDE-Constrained Variational Models: Training Variational Autoencoders (VAEs) with constraint-aware decoders that map latent codes through differentiable physics solvers (e.g., lithography, drift-diffusion) before outputting the final design.
3. Neural-Operator Priors: Using neural operators to learn mappings between function spaces (e.g., mask to aerial image) and using these as priors to constrain the generative model's search space to physically plausible regions.
4. Structure-Preserving Networks: Architectures that encode conservation laws (e.g., continuity, reaction-diffusion) directly into the network parameterization.

Key Integration Patterns

The paper outlines four specific integration patterns where data-driven generative models meet physics-based simulators (Figure 3):

1. Constrained Generative Process Planning: Coupling Large Language Models (LLMs) with process-rule checkers. The LLM handles combinatorial search heuristics, while the solver provides hard feasibility constraints on tool capabilities and thermal budgets.
2. Physics-Governed Synthetic Data Generation: Generating defects by perturbing parameters in a physics-based failure-mode simulator (capturing bridging, voids, etc.) and rendering them through a learned appearance model. This ensures synthetic data shifts correctly with process drift.
3. Simulator-Conditioned Inverse Design: Using differentiable simulators to inject adjoint sensitivity gradients into diffusion samplers. This allows the model to explore the physically feasible design space broadly, avoiding local minima and reducing simulator calls by orders of magnitude compared to standard optimizers.
4. Multimodal Foundation Models: A long-term vision of models pretrained jointly on text, layout/netlist geometry, and simulation outputs (TCAD/lithography) to create shared embeddings across modalities, enabling reasoning that currently requires expert translation.

Results and Evidence

The paper cites recent demonstrations (Table 3) to validate the efficacy of physics-informed architectures over data-only baselines:

• Mask Optimization: Physics-guided diffusion models have achieved order-of-magnitude reductions in simulation calls compared to data-only deep learning while maintaining fabrication-aware geometries.
• Inverse Design: Adjoint gradients injected into denoising steps have reached target figures of merit with roughly threefold fewer simulations than state-of-the-art nonlinear optimizers.
• Validity: Models with hard constraints or simulator-in-the-loop guidance recover physically valid candidates by construction, eliminating the need for post-hoc rejection of infeasible samples.

The paper emphasizes that while post-hoc filtering can raise the validity of surviving samples, only construction-level coupling converts physical law into search efficiency.

Significance and Claims

The paper's central claim is analytical: Physical validity is the operative criterion of success in semiconductor manufacturing, not merely a constraint. Therefore, architectures that enforce validity by construction are fundamentally superior to those that filter for it post-hoc.

The significance of this perspective lies in:

1. Defining a New Design Philosophy: It argues that treating physics as a structural property of the model (rather than an after-the-fact filter) is the only path to making generative AI a "load-bearing component" of the fab stack, rather than just a productivity gadget for ideation.
2. Establishing a Research Agenda: The paper proposes a three-horizon roadmap:
   • Near-term (2026–2027): Establishing physics-fidelity benchmarks that measure the fraction of generated samples passing independent physics solvers, rather than visual quality metrics.
   • Medium-term (2027–2030): Building infrastructure for differentiable fab simulators (open-source reimplementations and high-fidelity surrogates) to enable simulator-in-the-loop training.
   • Long-term (2030+): Developing multimodal foundation models trained on text, layout, and simulation corpora.
3. Addressing Certification: It highlights the critical, often overlooked challenge of qualifying stochastic generative systems for high-reliability production lines (e.g., automotive-grade), arguing that the community must engage with qualification owners directly.

The paper concludes that the choice between "physics-as-structure" and "physics-as-filter" is being made now in every codebase and product roadmap importing generative AI into semiconductor contexts. Only the former will define the next decade of advanced manufacturing.