Physics-informed dual-objective optimization of high-entropy-alloy nanozymes by a robotic AI chemist

Engineering artificial nanozymes as substitutes for natural enzymes presents a significant scientific challenge. High entropy alloys (HEAs) have emerged as promising candidates for mimicking peroxidase (POD) activity thanks to their unique properties and versatility. However, designing or discovering HEAs that surpass the catalytic efficiency of natural horseradish peroxidase involves complex challenges, often hindered by the multidimensional nature of HEAs’ compositional variability and the intricate interplay of enzymatic behaviours. Therefore, an intelligent and efficient approach to accelerate this discovery is crucial. In this study, we address these challenges by deploying a robotic artificial-intelligence chemist equipped with theoretical calculations, machine learning, Bayesian optimization, and on-the-fly data analysis by a large language model (LLM). Our approach centres on a physics-informed, multi-objective optimization framework that simultaneously optimizes multiple desirable properties of nanozymes, including maximum reaction rate and substrate affinity, ultimately optimizing catalytic efficiency. By integrating an auxiliary knowledge model based on physical insights and collaborative decision-making enabled by LLM-in-the-loop into Bayesian optimization, we enhanced the data-driven discovery workflow. Our physics-informed approach, with instant LLM-in-the-loop feedback, significantly outperformed both random sampling and standard Bayesian optimization. Consequently, we efficiently explored a vast chemical space and identified HEAs with enzymatic properties that significantly exceed those of the most effective catalysts based on HEAs or single atoms reported in the literature, as well as the natural enzyme.