Surrogate Model Optimization for Nuclear Reactor Core Design

The patent creates a simplified surrogate model (copy) of the complex reactor core physics system. This surrogate model replicates the essential neutronics, thermal, and stress behaviors but computes thousands of times faster than traditional methods. The surrogate model is trained on a dataset generated from high-fidelity simulations, allowing it to predict design outcomes with sufficient accuracy for iterative optimization while reducing computation time from months to hours.

The patent employs inexpensive, computationally lightweight surrogate models that can be rapidly evaluated thousands of times during optimization. These surrogate models are 'disposable' in the sense that they are simple approximations rather than high-fidelity physics solvers, allowing aggressive exploration of design space without the computational burden of traditional methods. The cheap evaluation cost enables parallel processing and extensive design iteration.

The patent performs preliminary action by pre-generating a comprehensive training dataset from high-fidelity physics simulations before the actual optimization process. This upfront investment in creating a robust surrogate model allows subsequent design evaluations to be performed rapidly. The surrogate model captures the complex physics relationships in advance, enabling fast predictions during the optimization iterations without repeatedly running expensive simulations.

The patent introduces a surrogate model as an intermediary between the design parameters and the physics-based validation requirements. Instead of directly running expensive MCNP and thermal-stress simulations for each design evaluation, the surrogate model mediates by providing rapid predictions of key performance metrics. This intermediary layer filters and approximates the complex physics computations, enabling fast design exploration while maintaining reasonable accuracy.

The patent implements feedback through the optimization algorithm that uses surrogate model predictions to guide the search for optimal designs. The algorithm learns from previous evaluations, identifying promising regions of the design space and focusing computational resources there. This feedback mechanism allows the system to converge to high-quality solutions much faster than exhaustive grid searches, adapting the exploration strategy based on accumulated knowledge from prior evaluations.

The patent employs dynamic optimization strategies where the search methodology adapts during the design exploration process. The algorithm dynamically adjusts sampling density, exploration vs. exploitation balance, and parameter ranges based on feedback from surrogate model evaluations. This dynamic approach allows efficient navigation of the design space, concentrating computational effort on regions likely to contain optimal solutions rather than uniformly sampling the entire space.