LithoDreamer: A Physics-Informed World Model for Multi-Stage Computational Lithography

Computational lithography is critical for modern chip manufacturing, but it involves a complex, continuous physical pipeline—from mask optimization to final development—that traditional static models struggle to capture. The authors ask whether a World Model (WM) framework can better simulate this multi-stage evolution by modeling the process as a series of decision-driven transitions rather than simple point-to-point mappings.

LithoDreamer introduces a physics-informed world model that decomposes the lithography pipeline into stage-specific latent spaces. The framework consists of three main components:

1. Space Prior Approximation (SPA): Uses eigenvalue decomposition to identify stage-specific physical latent spaces, ensuring that process interventions remain physically feasible.
2. Policy Model: A stochastic policy that plans latent process interventions (such as dose or focus adjustments) within these constrained spaces.
3. Transition Model: A Diffusion Transformer (DiT) that predicts how these interventions drive the state evolution across the Layout-Mask-Resist-ADI pipeline.

To train the model without requiring intermediate supervision, the authors propose a contrastive variational optimization paradigm. This method uses terminal-stage observations to guide the model, while contrastive learning ensures that interventions remain consistent with the specific physical dynamics of each stage.

LithoDreamer achieves state-of-the-art performance in both forward evolution (predicting the final image from a layout) and inverse planning (optimizing a mask to achieve a target image). By modeling lithography as a causal, multi-step process, the framework effectively handles the continuous adjustments required in industrial-grade manufacturing. Experiments on a dataset of 280,000 samples demonstrate that the model significantly outperforms existing learning-based methods in both in-domain and out-of-domain scenarios.