Federated RL Defect Classification for SEM Inspection

The patent replaces rule-based techniques (mechanical/systematic approach) with machine learning models that use reinforcement learning. The ML models learn optimal defect classification strategies through iterative training with reward signals, substituting the rigid rule-based system with an adaptive intelligent system that improves accuracy while reducing manual engineering time for rule creation and maintenance.

The patent changes the fundamental parameters of the classification system by transitioning from fixed rules to dynamic machine learning models. The models adjust their internal parameters (weights and biases) based on training data and reinforcement learning feedback, enabling adaptive classification that reduces misclassification errors and decreases the time required for engineering adjustments.

The patent segments the centralized training process into distributed local training sessions at each client site. Each client trains ML models locally on their own data while contributing to the global model through federated learning. This segmentation improves reliability by leveraging diverse local datasets while managing complexity through modular, distributed architecture that eliminates the need for centralized data aggregation.

The patent introduces a federated learning coordination layer that acts as an intermediary between multiple clients and the global model. This intermediary manages model distribution, collects local model updates, and aggregates them into improved global models, thereby enabling collaborative learning across clients while abstracting the complexity of distributed coordination from individual clients.

The patent applies reinforcement learning where the ML models receive feedback in the form of reward signals based on measurement accuracy and defect classification correctness. This feedback mechanism enables the models to learn optimal measurement and classification strategies that compensate for e-beam noise, improving manufacturing precision by adaptively adjusting to noise conditions rather than being constrained by fixed thresholds.

The patent performs preliminary training of machine learning models on labeled defect data before actual inspection. This preliminary action allows the models to learn noise patterns and measurement optimization strategies in advance, enabling them to make accurate measurements on extremely small structures even when e-beam noise is present during actual inspection.