RAMBO: A Risk-Adjusted Multi-Level Bayesian Optimization for achievable and safe autonomous material discovery

Material discoveries for improved societal, environmental, safety device applications etc. often require optimization of expensive experimental process for materials synthesis, characterization, and learning structure-property relationships over large parameter and function spaces where the active learning method such as Bayesian optimization (BO) have been heavily focused. While the traditional BO assumes confidence in probing at the suggested locations during expensive microscopic measurements, such real measurements often possess some drift. Moreover, the real experiments are often severely non-smooth, thus one cannot guarantee optimality or robustness in a traditional autonomous exploration while ignoring even a minor drift in input spaces. Therefore, we propose a robust autonomous exploration (rAE) via developing a Risk-adjusted multi-level Bayesian optimization (RAMBO). RAMBO is designed with first evaluating the utility-theory inspired risk at the suggested explored locations, which is then fitted to a prediction model in a multi-level structure with scalarizer/function prediction model, and finally propagated to the evaluation of a weighted robust acquisition function. We implemented RAMBO into STM scanned multiple defect data, and Piezoresponse spectroscopy data of PTO thin films. RAMBO demonstrated better performance than BO, and showcased the potential application to real-world experiment, where such risk-adjusted robust exploration can be critical to unlocking achievable and safe discoveries in autonomous research.