Manufacturing Recipe Recommendation Using Hybrid AI and Physics Models

The patent merges physics-based-simulation with machine learning models to create a hybrid system. The physics-based-simulation provides fundamental process understanding and predictions, while the machine learning model learns from experimental data including environmental conditions, machine states, and maintenance parameters. These two approaches are combined to overcome the limitations of each individual method, with the machine learning model compensating for factors that physics-based models cannot capture.

The patent creates a composite modeling approach by integrating multiple modeling paradigms. The hybrid model combines the deterministic nature of physics-based-simulation with the data-driven capabilities of machine learning, creating a composite system that leverages the strengths of both approaches to achieve higher reliability and accuracy in recipe optimization.

The patent applies preliminary action by using physics-based-simulation to generate synthetic data before training the machine learning model. This synthetic data is generated based on fundamental physics principles and covers a wide range of conditions that may not be available in experimental data. By preparing this additional training data in advance, the model achieves better accuracy without requiring complex architectures.

The patent creates copies of experimental data through physics-based-simulation. The simulation generates synthetic data that mirrors the structure and characteristics of experimental data but expands the training dataset to include scenarios that would be difficult or expensive to obtain experimentally. This copying approach allows the simple model to learn from a much broader range of conditions.