Presentation description

The optimization of parameters remains a critical limitation in the pursuit of novel materials. Traditionally, researchers use domain knowledge to tune these parameters but this can be inefficient and costly. Machine learning algorithms such as Bayesian Optimization (BO) have become popular tools to accelerate this process. However, BO relies on the accuracy of the surrogate model provided, commonly a Gaussian Process (GP), whose reliability can be hindered when dealing with small datasets containing outliers. This study investigates the effect of using a Student's T Process (STP) as an alternative surrogate model in BO. The STP offers advantages in BO modeling through its inherent robustness to outliers. By exploring this alternative approach to BO, we expect to increase the effectiveness and efficiency of parameter optimization for materials development.