Investigating Particle Swarm Optimization as an alternative to Bayesian Optimization in High Entropy Alloy Research

High entropy alloys (HEAs) offer a vast compositional space with a wide array of potential catalyst formulations. However, evaluating all possible HEA catalyst compositions is not possible, necessitating the use of optimization algorithms to efficiently identify the most promising candidates. In this work, we explore particle swarm optimization (PSO) as an alternative to the widely used Bayesian optimization method. While Bayesian optimization has demonstrated reliable convergence to global optima in HEA catalyst research, our findings show that PSO is prone to premature convergence, particularly when encountering strong local optima. We also assess the utility of the data generated during optimization for potential machine learning applications with the aim of comparing experimental and theoretical results. PSO exhibits high exploratory efficiency in the early stages, rapidly mapping the composition landscape. However, unlike Bayesian optimization, PSO ceases to learn once convergence is reached. This limitation suggests that enhancing PSO to mitigate premature convergence is required to improve its performance in HEA research.