Multi-objective optimization of molten salt microreactor shielding perturbations employing machine learning

Nuclear reactor shielding is a requirement for all reactors to prevent radiation from escaping the core. Shielding design can be an arduous process that requires significant computational resources to model even a single concept, for either deterministic and stochastic computations. As such, many developed shield concepts are under-optimized and unnecessarily large and/or expensive. Machine learning models have been previously trained and paired with advanced optimizers to assist in this process, providing a low-cost estimate of neutronic codes like MCNP. However, traditional models often require a large quantity of data, usually hundreds or thousands of iterations, to be accurately trained and deployed. Additionally, many neutronic modelers do not have the optimization background to build or deploy an effective shielding optimizer and would benefit from a simple Python framework. A molten salt reactor shielding problem is presented for optimization, utilizing a predictive machine learning model requiring only 21 iterations for training. This model is paired with a closed-loop optimizer using the Gekko Optimization Suite for Python. A baseline shield around the core is presented, and mass and cost constraints are applied at 172 metric tons and $1.4 million, respectively. An optimal shield is generated using the closed-loop optimizer (0.0377mREM/hr), allowing a similarly small dose to that of the baseline, but at a 10.8% reduced mass, a 11.9% reduced material cost, and 61 times computational cost reduction (98.4% improvement). The simple shield optimization framework resulting from this work offers acceleration and simplification of work for shield designers, allowing them to create improved designs in a fraction of the time.