Customized Acquisition Function Design for Materials Disovery Using Bayesian Algorithm Execution

The development of advanced materials requires precise and efficient search through a vast range of possible material candidates and conditions to find the select few which satisfy highly customized or specific experimental goals. We focus on the area of AI-based sequential decision making where, at each step, the next candidate material is suggested based on previous accumulated data. We develop and extend the recently proposed concept of Bayesian Algorithm Execution to allow users to automatically convert a complex, targeted experimental goal into an adaptive data collection strategy, resulting in substantially improved performance compared to state-of-the-art methods.