LithoDreamer Models Multi-Stage Computational Lithography with Physics-Informed AI.

As semiconductor technology advances to smaller nodes, computational lithography becomes increasingly critical for ensuring the yield and performance of chips. However, existing models often fail to fully capture the continuous physical processes involved, which include mask optimization, optical imaging, resist exposure, and development. LithoDreamer introduces a novel physics-informed World Model (WM) framework designed to address this limitation. It conceptualizes the entire lithography pipeline as a decision-driven, multi-step evolution system. The model captures subtle feature changes between adjacent states, allowing it to model stage-specific physics-informed latent spaces. This enables controlled exploration of process interventions and drives subsequent state transitions. To achieve interpretable intervention optimization without continuous supervision, LithoDreamer employs a contrastive variational optimization paradigm. This method contrasts latent differences between intervention paths with variational evolution constraints, ensuring that the generated evolutions are consistent with real lithography physics. Experiments confirm LithoDreamer's state-of-the-art performance in both forward evolution and inverse planning, with its lithography dataset made publicly available.