Title: LithoGRPO: Accelerating Inverse Lithography with GRPO-Reinforced Flow Matching

Original: arXiv:2606.00228v1 Announce Type: new Abstract: In semiconductor manufacturing, lithography projects circuit layouts onto silicon wafers through an optical mask. As circuit features shrink below the wavelength of light, optical diffraction causes the printed patterns to deviate from their intended layouts. Inverse Lithography Technology (ILT) addresses this challenge by generating optimized masks that enhance the fidelity of pattern transfer onto wafers. While ILT resembles an image synthesis task, its reliance on explicit physical metrics for mask evaluation limits the applicability of existing generative models. We introduce LithoGRPO, an ILT framework that integrates the flow-matching paradigm with GRPO-based reinforcement learning (RL) fine-tuning, enabling efficient exploration of diverse masks for a given target layout. Unlike purely generative or optimization-based approaches, RL in LithoGRPO exploits the explicitly defined, physics-based reward function of ILT, enabling optimization under complex, process-aware constraints. To the best of our knowledge, this is the first framework that unifies flow matching and RL for mask optimization. To improve RL sampling efficiency, we propose a fast shot-counting algorithm for manufacturability evaluation, achieving over 130x speedup while preserving the mask ranking of the traditional shot-count metric. Extensive experiments demonstrate that LithoGRPO achieves state-of-the-art performance over both optimization-based and learning-based methods, while maintaining efficient mask generation.

Rewrite: Lithography serves as the core mechanism in semiconductor fabrication, projecting circuit designs onto silicon wafers using optical masks. However, as device dimensions fall beneath the light’s wavelength, optical diffraction distorts the printed patterns, causing them to diverge from their target designs. Inverse Lithography Technology (ILT) mitigates this issue by producing optimized masks that improve the accuracy of pattern transfer. Although ILT shares similarities with image synthesis, its dependence on strict physical metrics for evaluation restricts the utility of current generative models.

To overcome these limitations, we present LithoGRPO, a novel ILT framework that combines the flow-matching paradigm with fine-tuning via GRPO-based reinforcement learning (RL). This approach allows for the efficient exploration of a wide variety of masks tailored to specific target layouts. Distinct from traditional optimization or purely generative methods, LithoGRPO leverages the explicit, physics-driven reward functions inherent to ILT. This enables robust optimization under intricate, process-aware constraints. To our knowledge, this represents the inaugural framework to merge flow matching with RL for mask optimization.

Furthermore, to enhance the efficiency of RL sampling, we developed a rapid shot-counting algorithm for assessing manufacturability. This innovation delivers a speed increase of more than 130 times compared to traditional metrics, without compromising the accuracy of mask rankings. Comprehensive experimental results confirm that LithoGRPO outperforms existing optimization-based and learning-based techniques, achieving state-of-the-art results while ensuring rapid mask generation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC