Optimization of passive modular molten salt microreactor geometric perturbations using machine learning

Nuclear reactors are vital to the global economy due to the large amounts of baseload power they can provide with minimal intermittence. However, these reactors are extremely complicated to design and build. Therefore, any improvement in the design process and timeline of nuclear reactor development would have a significant impact on the reactor design community and the world. In recent years, machine learning (ML) algorithms have gained popularity in handling difficult problems that previously required too much data or computational power to effectively solve. These include natural language processing, computer vision, and advanced process optimization. ML algorithms present a potentially powerful tool in the realm of reactor design and have already been implemented in some specific reactor design and operational optimization applications by researchers around the world. While these previous uses of ML algorithms are effective, they are very specific and a more generalized, and easily applied framework could potentially be used to accelerate the design of Gen IV reactors, reducing the workload of reactor designers and researchers. The development of such a versatile framework for reactor optimization using ML, CFD, and neutronic codes, and the validation of such, was the purpose of this work.The baseline reactor to be optimized was the BYU Molten Salt Micro-Reactor (Larsen et al., 2023). To establish this ML optimization, 35 training runs were completed using STARCCM + CFD software, and OpenMC neutronic modeling to generate data. The 3-dimensional cartesian data were extracted for each run, providing up to several million data points per training run. Moderator thickness, salt thickness, and moderator housing (plate) thickness were variables in the geometric optimization, and regressor type, data quantity per run, and training run quantity were variables in the model optimization. The model was first optimized by weighing model variables against train and test accuracy. This optimal model was then given 4096 geometric reactor iterations, and a temperature profile was predicted for each. The lowest-temperature option was chosen as lower temperature results in longer structural material life and less required secondary cooling expense, as well as more favorable neutronics. This result was then validated in CFD. The optimal reactor design surpassed all training runs by at least 220 K in maximum temperature, 85 K in average temperature, and 49 K in standard deviation of temperature. These represent significant improvement over all training designs and the baseline design.