Introduction
As semiconductor manufacturing pushes the boundaries of physics, the lithography process faces significant challenges. When transistors shrink below the wavelength of the light used to print them, diffraction causes the final patterns on silicon to deviate from the intended design. To counteract these distortions, the industry employs Optical Proximity Correction (OPC)—a process of strategically modifying mask geometries by moving edge segments to ensure the printed result matches the target layout.
Traditional OPC tools are computationally intensive, relying on iterative simulations that are slow and expensive. While Machine Learning (ML) and Reinforcement Learning (RL) have been proposed as faster alternatives, they struggle with two primary issues: Spatial Correlation (where moving one segment affects the light intensity of its neighbors) and the Curse of Dimensionality (the exponential growth of the action space as the number of segments increases). This paper introduces CAMO, a correlation-aware framework that leverages Graph Neural Networks (GNNs) and a novel lithography-inspired modulator to overcome these hurdles.
Methodology
CAMO approaches OPC as a sequential decision-making problem, utilizing a specialized architecture designed to capture the physical realities of lithography.
Graph-Based Representation and Feature Extraction
Unlike traditional models that treat layouts as flat images, CAMO represents the layout as a graph. In this structure, nodes represent individual edge segments, and edges represent spatial proximity. To process this data efficiently without the memory overhead of high-resolution pixel grids, the authors utilize "squish patterns" to encode geometry into tensors. A GraphSAGE-based feature extractor then fuses information from neighboring segments, ensuring the model understands the local environment of every edge.
Sequential Decision Making with RNNs
To address spatial correlation, CAMO employs a Recurrent Neural Network (RNN). By making decisions sequentially rather than in parallel, the model can account for the movements of previous segments when deciding the next action. This coordination prevents the "jittery" or unstable mask configurations often seen in independent-agent RL models.
The OPC-Inspired Modulator
The core innovation of CAMO is the Modulated Reinforcement Learning framework. The authors designed a mathematical modulator based on lithography principles to bias the RL agent’s choices. For instance, if a segment exhibits a large "inner" error (where the printed shape is smaller than the target), the modulator increases the probability of an outward move. This injection of domain knowledge constrains the search space, providing "guide rails" that stabilize training and prevent the agent from wandering aimlessly in high-dimensional action spaces.
Two-Phase Training Strategy
CAMO utilizes a hybrid training approach:
- Imitation Learning: The model first learns to mimic an existing OPC engine (such as Calibre) to gain a baseline intuition of mask optimization.
- Reinforcement Learning: The model is then fine-tuned using RL to optimize for specific metrics like Edge Placement Error (EPE), allowing it to eventually surpass the performance of its "teacher."
Experimental Results
The researchers evaluated CAMO on via layer patterns (70nm) and complex metal layer patterns generated via the NanGate 45nm Process Design Kit (PDK). CAMO was compared against state-of-the-art academic models (DAMO, RL-OPC) and the industry-standard tool, Calibre.
Accuracy and Robustness
On via layers, CAMO demonstrated superior precision, reducing the Edge Placement Error (EPE) by 20% compared to Calibre. Furthermore, CAMO achieved a lower Process Variation (PV) band, indicating that the generated masks are more robust against manufacturing fluctuations such as dose or focus variations.
Efficiency and Convergence
CAMO achieved a 1.32x speedup over Calibre on via layers. Notably, in the more complex metal layer tests, vanilla RL models often failed to converge or produced unstable results. In contrast, CAMO’s modulator allowed it to find stable, high-quality solutions where other ML approaches failed.
Conclusion
CAMO demonstrates that treating OPC as a graph-based, sequential optimization problem—rather than a simple image-to-image task—yields significant improvements in both mask quality and computational efficiency. By successfully integrating lithography domain knowledge into the RL policy via a modulator, the framework bridges the gap between purely data-driven models and physical manufacturing requirements.
Future work will focus on scaling this graph-based approach to full-chip layouts and reducing the dependency on commercial tools for the initial imitation learning phase, moving toward a more autonomous, self-supervised optimization system.
Summary updated