Metrology Fleet Error Compensation Using ML Matching Signals

The patent applies preliminary action by pre-training a machine learning model using composite measurement matching signals from multiple metrology systems. This pre-trained model captures systematic errors and relationships between systems in advance, enabling rapid error compensation during operation without requiring frequent recalibration. The model is trained offline using historical measurement data, and then deployed for real-time correction, thus performing the calibration work beforehand rather than repeatedly during operation.

The patent replaces the traditional mechanical calibration process with a computational machine learning approach. Instead of physically adjusting and recalibrating metrology systems through manual or automated mechanical procedures, the system uses a trained ML model to computationally compensate for systematic errors. This substitution transforms a time-consuming physical process into a rapid computational operation, significantly reducing recalibration time while maintaining measurement precision.

The patent merges multiple calibration procedures into a single unified machine learning model. Instead of applying separate calibration procedures to each metrology system individually, the system combines measurement data from multiple systems to train a comprehensive ML model that captures inter-system relationships. This single model then provides coordinated error compensation across all systems, replacing multiple complex calibration procedures with one unified computational approach.

The patent creates a universal calibration solution through a multi-functional machine learning model that can compensate for systematic errors across different types of metrology systems and various measurement applications. The single trained model serves multiple functions: it corrects errors in different measurement modalities, adapts to various target structures, and maintains consistency across the entire metrology fleet. This universal approach eliminates the need for separate calibration procedures for each system type or application.

The patent performs calibration work in advance by training the machine learning model offline using historical measurement data. Once trained, the model can rapidly compensate for systematic errors during production without interrupting the manufacturing flow. This shifts the time investment from the production phase to the preparation phase, allowing frequent error correction without sacrificing manufacturing throughput during operation.

The patent replaces time-consuming mechanical recalibration operations with rapid computational error compensation using the trained ML model. The computational approach processes measurements and applies corrections in real-time without requiring physical system interruptions or lengthy calibration sequences. This substitution maintains measurement stability while preserving manufacturing throughput by eliminating the productivity loss associated with frequent physical recalibration.