Hybrid mask 3D modeling driven by both physical solution and machine learning

With the increasing adoption of EUV lithography process and its smaller and smaller feature pitch size, the demand for both accuracy and runtime on mask 3D (M3D) modeling has become ever more important. Conventional physics-based modeling provides good stability and prediction on simulating M3D effects, but it can suffer from runtime penalty and accuracy loss for complex and dense patterns. Machine learning (ML) based method is able to achieve high accuracy and rapid computation, however, bearing the risk of potential overfitting. In this work, we will introduce a new hybrid M3D model driven by both physical modeling and machine learning, combining the benefits of both methods. We will first describe the workflow on creating a hybrid M3D model. Then we will present the evaluation results on its accuracy improvements and model apply runtime reduction in both 0.55 NA EUV and 0.33 NA EUV use cases.