Inference Model Training With Subset Selection for Metrology Accuracy

The patent creates physical copies of actual product structures as metrology targets. These targets are fabricated using the same lithographic process and contain identical features (e.g., lines, spaces, holes) at the same dimensions as the product structures. By copying the actual product geometry, the targets replicate the same optical distortions, focus variations, and process effects, enabling accurate inference of product feature dimensions while maintaining measurement simplicity.

The metrology targets are prepared in advance during the lithographic fabrication process itself, before the actual product structures are measured. The targets are formed simultaneously with product features using the same exposure and development steps, ensuring they experience identical process conditions. This preliminary creation of measurement references eliminates the need for separate target fabrication and ensures targets represent actual product state.

The patent extracts and removes redundant measurements from the training dataset. By analyzing the dataset, it identifies measurements that do not contribute meaningfully to model learning (such as measurements from non-representative targets or duplicate information) and excludes them. This extraction of essential information reduces dataset size while preserving the key patterns needed for accurate inference, thereby decreasing computational burden and training time.

The patent applies partial action by using only the necessary subset of measurements for training rather than the complete dataset. It determines that a portion of the measurements is sufficient to achieve the required model accuracy, avoiding the excessive computation that would result from processing all available measurements. This selective approach maintains inference quality while improving training efficiency.