Deep Learning for Computational Lithography and Mask Optimization

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 220 papers20 conceptsfrozen 2026-10-04

In the world of semiconductor manufacturing, we have reached a physical paradox: we are trying to print features that are significantly smaller than the wavelength of the light used to print them. When light passes through a mask to etch a circuit onto a silicon wafer, it doesn't travel in straight lines; it diffracts, interferes, and blurs.

If you were to use a mask that looked exactly like the intended circuit, the resulting chip would be a useless smear of rounded corners and merged lines. To solve this, engineers use Computational Lithography, a process of mathematically "pre-distorting" the mask so that the final printed image—after all the messy physics of diffraction—matches the original design.

The Computational Bottleneck

Traditionally, this pre-distortion is handled by two main techniques: Optical Proximity Correction (OPC), which moves the edges of shapes and adds tiny "serifs" to corners, and Inverse Lithography Technology (ILT), which treats the mask as a continuous optimization problem.

As Jin Qian and colleagues note in Recent advances in computational lithography technology, these physics-based simulations are becoming prohibitively slow as we push toward sub-7nm nodes. A single chip might require days of compute time on massive server farms just to generate the mask patterns. The research direction across these 17 papers is the transition from these slow, iterative numerical solvers to fast, "learned" neural models that can predict the optimal mask in a single forward pass.

From Shallow Heuristics to Deep Solvers

Early attempts to use machine learning in this space focused on "shallow" models for specific tasks like hotspot detection. However, Haoyu Yang and Bei Yu demonstrated in VLSI Mask Optimization: From Shallow To Deep Learning that deep learning could do more than just find errors—it could drive the optimization itself.

The core intuition is that the relationship between a target layout and its ideal mask is a complex but learnable mapping. Rather than solving Maxwell’s equations from scratch for every new chip, a neural network can learn the "inverse physics" of the lithography system. For instance, Kumara Sastry and Animashree Anandkumar utilized a Convolutional Fourier Neural Operator (CFNO) in Large Scale Mask Optimization Via Convolutional Fourier Neural Operator and Litho-Guided Self Training to achieve speedups of over 13,000x compared to traditional methods. By operating in the frequency domain, the CFNO naturally captures the wave-like nature of light diffraction.

Mimicking the Physics

The most successful recent models don't just treat the problem as a "black box" image-to-image task. Instead, they incorporate the logic of physics solvers into their architecture.

In ILILT: Implicit Learning of Inverse Lithography Technologies, Haoyu Yang and Haoxing Ren introduced implicit layer learning. This allows the network to mimic the iterative refinement of a numerical solver without the computational overhead. Similarly, Wenqian Zhao developed AdaOPC, a framework that intelligently routes "easy" patterns to fast ML solvers while reserving "hard," novel patterns for rigorous numerical optimization, ensuring both speed and reliability.

Ensuring Manufacturability

A recurring challenge in this field is that a neural network might generate a mask that is mathematically perfect but physically impossible to manufacture. Traditional ILT often produces "curvilinear" masks—smooth, organic shapes that are difficult for mask-writing tools to create.

To address this, researchers are moving toward generative models that understand geometric constraints. Yiwen Wu and Xuming He proposed TokMan in TokMan:Tokenize Manhattan Mask Optimization for Inverse Lithography, which tokenizes layouts into "Manhattan-aligned" (rectilinear) segments. By treating mask generation as a discrete sequence task using a Diffusion Transformer, they ensure the output is natively compatible with manufacturing hardware.

Automating the Expert

Finally, the research is moving toward automating the high-level strategy of lithography. Developing an "OPC recipe"—the set of rules and parameters for a specific manufacturing process—usually takes human experts months of trial and error.

Guojin Chen and Bei Yu have pioneered the Intelligent OPC Engineer Assistant (arxiv 2024, aaai 2025). This system uses Reinforcement Learning to explore optimization strategies and Large Language Models (LLMs) to summarize those strategies into human-readable rules. This represents a shift from merely accelerating the math to automating the entire engineering workflow of semiconductor fabrication.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 20 papers
  2. Version 1 · 2026-10-04 · 20 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_optical_proximity_effects_da2ce91f,
  title = {optical proximity effects},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optical-proximity-effects/rdr_da2ce91f43b34835b10177df7b36a22b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_deep_learning_for_computational_lithogra_da2ce91f_v1,
  title = {Deep Learning for Computational Lithography and Mask Optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-learning-for-computational-lithography-and-mask-optimization/rdr_da2ce91f43b34835b10177df7b36a22b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}