Machine Learning Model Training for Lithography Patterning Accuracy

The machine learning model is trained in advance using historical process data and simulation results before actual production. This preliminary training phase allows the model to learn complex patterning process relationships offline, so that during actual manufacturing, the model can quickly predict critical dimensions and provide optimization recommendations without adding significant real-time computational burden.

The machine learning model acts as an intermediary between traditional patterning processes and advanced lithography systems. It bridges the gap by taking simple process parameters as input and providing enhanced prediction capabilities for critical dimensions, thereby improving manufacturing precision without requiring fundamental changes to the underlying lithography equipment or process flow.

The model training process is segmented into distinct phases: initial model construction using simulation data, iterative refinement using measured process data, and final validation. This segmentation allows parallel processing where simulation-based training can occur independently of physical manufacturing, reducing the overall time impact on production while still achieving high prediction accuracy through multiple refinement iterations.

The patent implements iterative model training where the model is progressively refined through multiple training cycles. Rather than requiring complete retraining from scratch, the model undergoes partial updates using new measured data, where each iteration incrementally improves prediction accuracy. This approach balances the need for high precision with acceptable training time by performing targeted updates rather than exhaustive retraining.

Extensive model training and cost function optimization are performed in advance during non-production periods. The model undergoes multiple iterative training cycles using historical data to achieve high prediction accuracy before being deployed for actual patterning processes. Once trained, the model provides rapid predictions during production without requiring repeated iterative optimization, thereby maintaining high manufacturing precision while preserving production throughput.

The machine learning model creates a virtual copy of the complex patterning process physics and chemistry. Instead of performing repeated physical trial-and-error experiments to optimize patterns, the trained model provides rapid virtual predictions of critical dimensions and edge placement. This virtual copying allows extensive optimization to be performed computationally rather than physically, improving prediction accuracy without sacrificing production productivity.