Can Advances in Artificial Intelligence Surpass Legacy Algorithms for PWR Core Optimization?

The Pressurized Water Reactor (PWR) loading pattern optimization problem has been a major area of research in the field of nuclear engineering. Recent interest in Deep Reinforcement Learning (DRL) has sparked to solve the PWR loading pattern optimization problem. Our previous work in [2] was to our knowledge the first attempt to leverage DRL, namely Proximal Policy Optimization (PPO) to solve the problem discussed above. Using SIMULATE3 as the reactor physic code to evaluate the cores generated, we performed an analysis to understand the influence of several hyper-parameters to optimize the economy of a Westinghouse-type 4-loop PWR with 7 safety and operational constraints. Nevertheless, we did not compare its performance against classical heuristic method(s). The work presented in this paper attempts to compare DRL against the commonly utilized SA, GA, and a novel parallel TS on the same problem defined in [2] to ensure the soundness of the argument for introducing this new paradigm. Here, we demonstrated that leveraging PPO, a recently developed machine learning-based algorithm, can surpass the classical methods behind the state-of-the-art codes in academia and the industry for the PWR loading pattern optimization. This is just another example which illustrates the necessary shift that the nuclear needs to undertake to adapt to the potent Artificial Intelligence revolution.