Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Pushing the Limits of Inverse Lithography with Generative Reinforcement Learning

Haoyu Yang, Haoxing Ren

TL;DR

This work rethinks inverse lithography by treating mask synthesis as conditional sampling and training a style-aware generator to propose multiple design-conditioned masks. Through a two-stage process—generative pre-training and reinforcement finetuning with Group Relative Policy Optimization—the sampler learns a posterior over masks that accelerates ILT refinement and helps escape non-convex traps. Empirical results on LithoBench and ICCAD13 show state-of-the-art or competitive EPE with significant throughput improvements, including robust performance at 3 nm and substantial speedups. The framework generalizes to arbitrary sizes, integrates with existing ILT pipelines, and supports flexible multi-objective rewards, offering a scalable path toward practical, high-quality lithography mask synthesis.

Abstract

Inverse lithography (ILT) is critical for modern semiconductor manufacturing but suffers from highly non-convex objectives that often trap optimization in poor local minima. Generative AI has been explored to warm-start ILT, yet most approaches train deterministic image-to-image translators to mimic sub-optimal datasets, providing limited guidance for escaping non-convex traps during refinement. We reformulate mask synthesis as conditional sampling: a generator learns a distribution over masks conditioned on the design and proposes multiple candidates. The generator is first pretrained with WGAN plus a reconstruction loss, then fine-tuned using Group Relative Policy Optimization (GRPO) with an ILT-guided imitation loss. At inference, we sample a small batch of masks, run fast batched ILT refinement, evaluate lithography metrics (e.g., EPE, process window), and select the best candidate. On \texttt{LithoBench} dataset, the proposed hybrid framework reduces EPE violations under a 3\,nm tolerance and roughly doubles throughput versus a strong numerical ILT baseline, while improving final mask quality. We also present over 20\% EPE improvement on \texttt{ICCAD13} contest cases with 3$\times$ speedup over the SOTA numerical ILT solver. By learning to propose ILT-friendly initializations, our approach mitigates non-convexity and advances beyond what traditional solvers or GenAI can achieve.

Pushing the Limits of Inverse Lithography with Generative Reinforcement Learning

TL;DR

This work rethinks inverse lithography by treating mask synthesis as conditional sampling and training a style-aware generator to propose multiple design-conditioned masks. Through a two-stage process—generative pre-training and reinforcement finetuning with Group Relative Policy Optimization—the sampler learns a posterior over masks that accelerates ILT refinement and helps escape non-convex traps. Empirical results on LithoBench and ICCAD13 show state-of-the-art or competitive EPE with significant throughput improvements, including robust performance at 3 nm and substantial speedups. The framework generalizes to arbitrary sizes, integrates with existing ILT pipelines, and supports flexible multi-objective rewards, offering a scalable path toward practical, high-quality lithography mask synthesis.

Abstract

Inverse lithography (ILT) is critical for modern semiconductor manufacturing but suffers from highly non-convex objectives that often trap optimization in poor local minima. Generative AI has been explored to warm-start ILT, yet most approaches train deterministic image-to-image translators to mimic sub-optimal datasets, providing limited guidance for escaping non-convex traps during refinement. We reformulate mask synthesis as conditional sampling: a generator learns a distribution over masks conditioned on the design and proposes multiple candidates. The generator is first pretrained with WGAN plus a reconstruction loss, then fine-tuned using Group Relative Policy Optimization (GRPO) with an ILT-guided imitation loss. At inference, we sample a small batch of masks, run fast batched ILT refinement, evaluate lithography metrics (e.g., EPE, process window), and select the best candidate. On \texttt{LithoBench} dataset, the proposed hybrid framework reduces EPE violations under a 3\,nm tolerance and roughly doubles throughput versus a strong numerical ILT baseline, while improving final mask quality. We also present over 20\% EPE improvement on \texttt{ICCAD13} contest cases with 3 speedup over the SOTA numerical ILT solver. By learning to propose ILT-friendly initializations, our approach mitigates non-convexity and advances beyond what traditional solvers or GenAI can achieve.
Paper Structure (29 sections, 10 equations, 4 figures, 2 tables)

This paper contains 29 sections, 10 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. The Framework
  4. Revisiting Generative AI for ILT
  5. Generative AI as Mask Sampler
  6. Style-Aware Generative Model Architecture
  7. Design insights.
  8. Training Methodology
  9. Generative Pre-Training
  10. Reinforcement Finetuning
  11. Teacher-relative group policy optimization.
  12. Policy loss.
  13. Imitation toward ILT-refined masks.
  14. Overall objective and practicalities.
  15. Discussion
  16. ...and 14 more sections

Figures (4)

  • Figure 1: Comparison of ILT schemes: (a) Pure ILT with iterative updates; (b) GenAI warm-start followed by ILT refinement; (c) Proposed conditional mask sampling that generates multiple candidates, applies batched fast ILT refinement, and selects the best result.
  • Figure 2: Mask optimization metrics. Resketched from curvyilt.
  • Figure 3: Detailed style-aware generator. A pyramid of inputs $\{\boldsymbol{Z}_\ell\}$ is formed (optional head downsample). The coarsest input drives the level 2 generator: stride-2 downsamples, style-modulated residual trunk (AdaIN with $\boldsymbol{w}=f_\phi(\boldsymbol{z})$), then symmetric upsampling and bounded output. Two other levels operate coarse-to-fine by fusing the upsampled output from the previous stage with downsampled features at the current resolution (element-wise addition), refining via local ResBlocks and upsampling. An optional final bicubic interpolation restores the original resolution.
  • Figure 4: Visualization of our methods for an example design (HA_X1__1_0). Left: target. Top row: golden mask (EPE in nm) and five posterior sampled masks from $G$. Bottom row: corresponding nominal images.

Theorems & Definitions (2)

  • Definition 1: EPE curvyiltOPC-ICCAD2013-Banerjee
  • Definition 2: PVB curvyiltOPC-ICCAD2013-Banerjee