Materials design with target-oriented Bayesian optimization

Abstract

Materials design using Bayesian optimization (BO) typically focuses on optimizing materials properties by estimating the maxima/minima of unknown functions. However, materials often possess good properties at specific values or show effective response under certain conditions. We propose a target-oriented BO to efficiently suggest materials with target-specific properties. The method samples potential candidates by allowing their properties to approach the target value from either above or below, minimizing experimental iterations. We compare the performance of target-oriented BO with that of other BO methods on synthetic functions and materials databases. The average results from hundreds of repeated trials demonstrate target-oriented BO requires fewer experimental iterations to reach the same target, especially when the training dataset is small. We further employ the method to discover a thermally-responsive shape memory alloy Ti_0.20Ni_0.36Cu_0.12Hf_0.24Zr_0.08 with a transformation temperature difference of only 2.66 °C (0.58% of the range) from the target temperature in 3 experimental iterations. Our method provides a solution tailored for optimizing target-specific properties, facilitating the accelerated development of materials with predefined properties.