Learning-Based Model for Simulated Specimen Imaging

The patent creates a simulated copy of the physical imaging process through a learning-based model. The model is trained on actual optical images, electron beam images, and design data to learn the mapping relationships, then generates simulated images without requiring physical imaging tools. This copying approach eliminates time-consuming physical imaging while preserving mapping accuracy through learned patterns.

The patent performs preliminary training of the learning-based model using actual imaging data before actual use. During this preliminary phase, the model learns the complex mapping relationships between different imaging modalities and design data. Once trained, the model can rapidly generate simulated images without requiring time-consuming physical imaging processes during actual operation.

The patent replaces expensive physical imaging tools with a learned computational model that copies their functionality. The learning-based model is trained on actual images from optical microscopes and electron beam tools, then generates simulated images that maintain the accuracy characteristics of the original tools without requiring their physical presence, significantly reducing costs.

The patent introduces a learning-based model as an intermediary between available data and required imaging output. Instead of directly using expensive physical imaging tools, the model acts as a mediator that translates between different data types (optical images, electron beam images, design data) while preserving the accuracy characteristics that would otherwise require expensive equipment.

The patent creates a dynamic system where the learning-based model can adapt to different input combinations. The model is trained to handle various scenarios: when design information is available, when only optical images are available, when only electron beam images are available, or when multiple data types are combined. This dynamic adaptability allows the system to maintain mapping completeness across different conditions without requiring design information in all cases.

The patent designs a universal learning-based model that can perform multiple mapping functions. The same model architecture handles mapping from optical images to electron beam images, from design data to images, and from images to design data. This multi-functional approach eliminates the need for design information in certain scenarios while maintaining mapping completeness, thereby increasing system versatility.