LithoDreamer world model for multi-stage lithography

LithoDreamer, a physics-informed World Model for multi-stage computational lithography, was submitted to arXiv on 25 Jun 2026 as arXiv:2606.26713. The paper, authored by Yuqi Jiang and nine coauthors with correspondence to Qi Sun, presents a framework that treats the Layout-Mask-Resist Image-After Development Image pipeline as a decision-driven multi-step evolution system and reports state-of-the-art performance in forward evolution and inverse planning.

What is LithoDreamer?

LithoDreamer is presented as the first physics-informed World Model framework for computational lithography, explicitly formulating the Layout-Mask-Resist Image-After Development Image (ADI) pipeline as a multi-step evolution. The authors position the model to capture feature changes between adjacent states and to create stage-specific physics-informed latent spaces that support interpretable intervention optimization without continuous supervision.

The paper lists ten authors: Yuqi Jiang, Yumeng Liu, Zimu Li, Jinyuan Deng, Qian Jin, Yucheng Cui, Yu Li, Xunzhao Yin, Qi Sun, and Cheng Zhuo, and identifies Qi Sun for correspondence. The submission file on arXiv is 13,033 KB and the work appears in the Proceedings of the 43rd International Conference on Machine Learning, Jul. 6-11, 2026.

How does the model work?

LithoDreamer models the pipeline by capturing feature differences between adjacent states and encoding those as stage-specific physics-informed latent spaces, then using those latents to control process interventions and drive subsequent state transitions. The authors introduce a contrastive variational optimization paradigm that contrasts latent differences between intervention paths with variational evolution constraints to guide the model toward evolutions consistent with lithography physics.

Concretely, the framework frames the sequence Layout → Mask → Resist Image → After Development Image (ADI) as a decision-driven evolution: the model learns the mapping of state transitions, represents stage-specific physics in latent form, and uses those latents to explore and optimize interventions that change later states. The paper asserts that this architecture enables interpretable intervention optimization without continuous supervision.

What did the experiments show?

The authors report that LithoDreamer achieves state-of-the-art performance on both forward evolution and inverse planning tasks within their experiments. The submission abstract explicitly states those results and notes that the lithography dataset used for evaluation is publicly available on GitHub (the paper provides a URL placeholder for that repository).

The experimental claims and the public dataset signal that the model was evaluated across the formulated Layout-Mask-Resist Image-ADI pipeline; the paper ties the quantitative claims to the new physics-informed world model and the proposed contrastive variational optimization approach.

Why it matters

Treating the full Layout-Mask-Resist-ADI sequence as a decision-driven evolution gives a single, unified framework for both forward simulation and intervention planning, addressing the paper's observation that lithography is a continuous physical process that existing models fail to capture. If the model and its public dataset reproduce the reported state-of-the-art results, this approach could change how mask optimization and process interventions are explored during computational lithography research and tool development.

What to watch

Check the GitHub dataset and the ICML proceedings record from Jul. 6-11, 2026 for the paper's evaluation details and code links; the arXiv submission is arXiv:2606.26713 (submitted 25 Jun 2026). The next concrete signal will be whether the public repository contains the training and evaluation assets that reproduce the claimed state-of-the-art forward and inverse results.