Novel End-to-End Production-Ready Machine Learning Flow for Nanolithography Modeling and Correction (2401.02536v1)

Abstract: Optical lithography is the main enabler to semiconductor manufacturing. It requires extensive processing to perform the Resolution Enhancement Techniques (RETs) required to transfer the design data to a working Integrated Circuits (ICs). The processing power and computational runtime for RETs tasks is ever increasing due to the continuous reduction of the feature size and the expansion of the chip area. State-of-the-art research sought Machine Learning (ML) technologies to reduce runtime 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.

• The paper presents TPM-RET, an end-to-end machine learning framework that applies true pixel-based correction to resolve traditional OPC challenges.
• It employs an innovative Inverse Intensity Profile to capture detailed process data, enhancing prediction accuracy and consistency across full-chip scales.
• The framework scales efficiently on CPU infrastructures, achieving over 97% efficiency in 32nm immersion lithography tests and enabling broader semiconductor applications.

An Overview of a Novel End-to-End Machine Learning Flow for Nanolithography Modeling and Correction

This paper introduces an innovative framework, termed TPM-RET (True Pixel-based Machine-learning RET), designed for the modeling and correction of lithography processes utilizing machine learning methodologies. The work attempts to bridge the gap between the theoretical capabilities of machine learning-based resolution enhancement techniques (ML-RET) in optical lithography and practical, production-ready applications. With the relentless push toward nanometer-scale manufacturing, the intricacies of optical proximity correction (OPC) and associated resolution enhancement techniques (RETs) become significantly resource-intensive. The authors propose an end-to-end flow to address these computational challenges, yielding a scalable and efficient solution.

Key Contributions

The paper identifies several critical challenges hampering the adoption of ML-RET techniques, including:

1. Loss of Information: The binary nature of photomask data conceals essential intermediate process information.
2. Full-chip Scale Complexity: Current ML approaches struggle with the scale of processing and introduce inconsistencies due to chip slicing and windowing.
3. Infrastructure and Consistency Challenges: ML models typically require GPU-based platforms, contrasting with the CPU-centric infrastructure already established in fabs.

About TPM-RET Flow

The proposed TPM-RET framework provides a highly scalable architecture that adheres to several design principles aimed at overcoming these challenges:

• True Pixel-Based Correction: By employing pixel-by-pixel correction, TPM-RET avoids inconsistencies related to window-based methods and eliminates boundary conflicts, ensuring coherent output across the chip.
• Inverse Intensity Profile (IIP): The introduction of the IIP concept enables models to capture detailed process information, bridging the gap left by traditional binary photomask representations. This facilitates the model’s ability to infer meaningful relationships, enhancing prediction accuracy.
• Cross-Platform Scalability: The architecture is designed to function efficiently across existing CPU infrastructures, making it more attractive for existing semiconductor manufacturing operations.
• End-to-End Correction: TPM-RET eschews the traditional multi-stage correction approach in favor of a seamless, single-stage process, leveraging the comprehensive data available to machine learning algorithms.

Performance Evaluation

The paper validates TPM-RET using test patterns from a 32nm immersion lithography process, demonstrating substantial alignment between TPM-RET predictions and reference photomasks generated via pxOPC. The TPM-RET framework achieves a remarkable scalability advantage, with performance scaling linearly as the number of CPUs increases, demonstrating over 97% efficiency when deployed on highly parallelized systems.

Implications and Future Directions

From a practical standpoint, the TPM-RET framework holds promise for advancing the full-scale deployment of ML-RET solutions in semiconductor manufacturing environments. Its minimal hardware requirements and alignment with current CPU-based systems make it feasible for integration without significant capital reinvestment. Longer term, further refinements—such as transitioning to a C/C++ implementation—could enhance runtime performance and broaden applicability in real-time lithographic simulations.

The framework's inherent flexibility enables its application beyond OPC to various RET processes, as well as potential applications in predicting electrical stress, broadening its utility across different stages of IC manufacturing.

Overall, while TPM-RET is a step towards realizing production-ready ML-based lithography corrections, continued research in refining the predictive models and exploring broader use cases will be crucial in fully unlocking the potential of this approach within the semiconductor industry.