Novel End-to-End Production-Ready Machine Learning Flow for Nanolithography Modeling and Correction

Aug 5, 2026

Upload by Aria

Optical lithography is a key enabler for semiconductor manufacturing but relies on resolution enhancement techniques (RETs) whose runtime and computational cost grow with shrinking feature sizes and increasing data volume. While machine learning has been explored to accelerate RET correction and reduce computational load, ML-RET solutions are not yet widely adopted in production. The study analyzes why ML-RET remains production-unready and proposes a highly scalable end-to-end flow for production-ready ML-RET correction.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Traditional RET

State-of-the-art ML-RET and roadblocks

TPM-RET end-to-end modeling and correction

Results

How TPM-RET addresses production difficulties

Future plans and conclusions

arXiv :2401 .02536v 1 [ cs .LG] 4 Jan 2024 Novel End-to-End Production-Ready Machine Learning Flow for Nanolithography Modeling and Correction Mohamed S. E. Habiba (Researcher), Hossam A. H. Fahmya (Researcher) and Mohamed F. Abu-ElYazeeda (Researcher) aEECE, Faculty of Engineering, Cairo University, Egypt | ARTICLE INFO | AB STRACT | | | --- | --- | --- | | Keywords: | Optical lithography is the main enabler to semiconductor manufacturing. It requires extensive | | | Optical Lithography | processing to perform the Resolution Enhancement Techniques (RETs) required to transfer the design | | | Semiconductor Manufacturing | data to a working Integrated Circuits (ICs) . The processing power and computational runtime for RETs | | | Machine Learning | tasks is ever increasing due to the continuous reduction of the feature size and the expansion of the | | | Resolution Enhancement Techniques | chip area. State-of-the-art research sought Machine Learning (ML) technologies to reduce runtime | | | Optical Proximity Correction | and computational power, however they are still not used in production yet. In this study, we analyze the reasons holding back ML computational lithography from being production ready and present a novel highly scalable end-to-end flow that enables production ready ML-RET correction. | | | 1. Introduction
Computational lithography aims to prepare synthesized integrated circuits (ICs) designs for accurate transfer into the semiconductor wafer. It performs elaborate correction flows that adds, subtracts and modifies the original IC patterns with the end goal of minimizing the fabrication process errors. Such correction flows and operations are often referredto as Resolution Enhancement Techniques (RETs) and are known for being computationally expensive and time consuming.
As nanolithography technologies continue to advance towards ever decreasing minimum feature sizes, the same die area is able to host more design data. Henceforth, the data volume required to produce the same wafer area increase substantially, consequently, increasing the computational power and runtime required to perform the RET correction.
Computational lithography industry sought Machine Learning (ML) 1 algorithms to accelerate the RET correction and reduce the computational load. However, ML-RET are not yet adopted for production to the best knowledge of the authors.
In this study, we analyze the reasons holding back ML-RET correction from utilization in RET production flows. Based on this analysis, we introduce a novel flow that mitigates those obstructions and offers an end-to-end production-ready platform for ML-RET with very highscalability. The rest of this paper is organized as follows: in section 2, we discuss the traditional way of performing RET correction. Next, we discuss and analyze the state-ofthe-art of ML-RET and the roadblocks holding it from being
[mohamed1611071@eng1.cu.edu.eg](mohamed1611071@eng1.cu.edu.eg) (M.S.E. Habib); [hossam.fahmy@eng.cu.edu.eg](hossam.fahmy@eng.cu.edu.eg) (H.A.H. Fahmy); [mfathyae@eng.cu.edu.eg](mfathyae@eng.cu.edu.eg)[ ](mfathyae@eng.cu.edu.eg)(M.F. Abu-ElYazeed)
ORCID(s):
1We use the term “machine learning” as generalized term to refer to any advanced AI techniques including deep learning, generative networksand similar technologies. | | a viable production option in section 3. We then introduce TPM-RET, a novel end-to-end production friendly ML flow to model and correct lithography in section 4, and showcase some of its results in section 5. Next, we discuss how TPMRET flow solves the production difficulties in section 6. Finally, we present our future plans and conclusions in sections 7 and 8 respectively.
2. Traditional RET
Optical lithography machinery provided good accuracy early-on for large pattern dimensions. However, its accuracy and fidelity degraded as manufacturing requirements progressed towards smaller dimensions, even though such dimension

FAQs

Why are resolution enhancement techniques (RETs) computationally expensive in nanolithography?

RET corrections require elaborate add/subtract/modify operations to minimize fabrication errors, and shrinking feature sizes increase the data volume and runtime needed for the same wafer area.

What barrier prevents machine learning-based RET (ML-RET) from being production ready?

The paper analyzes multiple reasons holding ML-RET correction back from utilization in RET production flows, noting that state-of-the-art ML-RET has not yet been adopted for production.

What does the proposed TPM-RET flow provide?

It introduces a novel end-to-end production-friendly machine learning flow to model and correct lithography, with very high scalability aimed at resolving production difficulties.

Aug 5, 2026

Upload by Aria