Lithography Model Training for Mask Bias Variation Prediction

The patent applies data augmentation techniques to pre-generate transform image data that represents various process variations (such as focus, dose, and mask bias variations) before actual model training. By pre-processing and synthesizing diverse training data through geometric transformations and parameter modifications, the system prepares comprehensive process condition coverage in advance, reducing the need to collect and process extensive experimental data during the learning phase.

The patent creates transformed copies of basic image data through data augmentation methods, generating synthetic training samples that represent different process variations. These copied and transformed images (with modified focus, dose, or mask bias parameters) serve as additional training data without requiring physical experimentation for each condition, thus expanding process coverage while maintaining efficient learning time.

The patent implements a dynamic weighting mechanism that adjusts the importance of different training samples during the learning process. The system dynamically modifies weights based on the specific process variations present in each training iteration, allowing the model to adaptively focus on critical mask bias variation patterns. This dynamic adjustment enhances prediction accuracy for mask bias variations without requiring a fundamentally more complex model architecture.

The patent modifies training parameters by applying data transformations that specifically target mask bias variation characteristics. Through parameter modifications such as focus variation, dose variation, and mask bias variation in the training data, the model learns to predict these variations more accurately. The approach changes the training parameters and data characteristics rather than fundamentally redesigning the model structure, achieving improved accuracy with controlled complexity.