GAN Process Parameter Prediction for Semiconductor Fabrication

The patent creates a virtual copy of the semiconductor fabrication process through a trained neural network model. Instead of physically testing parameters on real wafers, the system copies the fabrication process in silico using the trained model, which has learned the complex relationships between process parameters and outcomes from historical data. This virtual copying eliminates the need for time-consuming physical trial-and-error while maintaining parameter determination accuracy

The patent performs preliminary training of the neural network model using historical fabrication data before actual parameter determination is needed. This preliminary action involves feeding the model extensive historical process data and outcomes, allowing it to learn optimal parameter relationships in advance. When actual parameter determination is required, the pre-trained model can quickly predict optimal parameters without needing to perform physical experiments

The patent replaces physical material consumption with virtual simulation. The trained neural network model creates a digital twin of the fabrication process, allowing parameter optimization to be performed in the virtual domain. This copying approach eliminates the need to consume actual semiconductor materials, chemicals, and equipment time during the parameter determination phase, while still achieving accurate parameter identification

The system uses historical fabrication data that has already been collected during normal production to train the neural network model. This self-service approach means the system learns from its own past operations without requiring additional external resources or materials. The historical data serves dual purposes: it represents past material consumption that already occurred, and it provides the training fuel for the AI model that will eliminate future material waste

The patent replaces the mechanical system of physical experimentation with an information-based system. Instead of mechanically adjusting parameters and observing physical outcomes, the system uses information processing through the neural network model. The model processes historical data patterns and predicts optimal parameters computationally, substituting the slow mechanical trial-and-error process with rapid information analysis and prediction

The system fundamentally changes the approach from continuous physical parameter adjustment to discrete computational parameter prediction. The neural network model has learned the optimal parameter values or relationships from historical data, allowing it to directly predict suitable parameters for new fabrication scenarios without needing to physically explore the parameter space through experimentation