Semiconductor Waiting Time Prediction Using Operation-Level ML Models

The patent segments the complex discrete-event simulation into multiple independent machine learning models, each trained on historical data for specific operations or time periods. This segmentation allows parallel execution of multiple simpler models instead of running one complex simulation, significantly reducing execution time while maintaining prediction accuracy through ensemble methods.

The patent creates simplified copies of the production system using historical data to train machine learning models. These models are lightweight copies that capture essential patterns without the computational overhead of full discrete-event simulations, enabling fast predictions while preserving accuracy through proper training on comprehensive historical datasets.

The patent replaces the mechanical discrete-event simulation system with machine learning models that use statistical patterns from historical data. This substitution eliminates the need to maintain complex simulation logic, event queues, and interdependency models, while preserving prediction capabilities through data-driven approaches that are easier to train and maintain.

The patent uses lightweight machine learning models that can be quickly trained and discarded rather than maintaining a single complex simulation system. These models require minimal computational resources and can be retrained efficiently when production patterns change, reducing long-term maintenance complexity compared to discrete-event simulations.