Machine Learning Surrogate Models for Nuclear Reactor Design Optimization

The patent applies preliminary action by using machine learning models to pre-screen and eliminate invalid design configurations before submitting them to high-fidelity physics simulations. The system performs rapid ML-based assessments of neutronics, thermal-hydraulic, and stress properties to identify promising designs, then only those passing the ML threshold undergo expensive MCNP Monte Carlo simulations. This preliminary filtering dramatically reduces the number of full physics simulations needed, cutting computation time from months to hours while maintaining simulation fidelity for the designs that matter most.

The patent employs copying by creating simplified machine learning surrogate models that replicate the behavior of complex physics simulation codes. These ML models are trained on a subset of simulation data and then used to rapidly evaluate design configurations without requiring actual physics simulations. The system copies the essential physics relationships in a computationally efficient form, enabling rapid design space exploration with minimal computational resources while maintaining accuracy for decision-making.

The patent merges multiple separate analysis workflows (neutronics, thermal-hydraulic, stress analysis) into a unified machine learning framework. Instead of running separate validated physics codes for each discipline and manually integrating results, the system combines all physics considerations into integrated ML models that simultaneously evaluate multiple design criteria. This unified approach maintains design validity by incorporating physics constraints while eliminating workflow complexity and software integration issues associated with running multiple independent simulation packages.