Integrated Approaches for Data-Driven Materials Discovery in Autonomous Laboratories

With the growing desire to deliver autonomous laboratories, computational materials scientists face new challenges that extend beyond inverse design, such as virtual screening, predictive modeling under various constraints, and optimizing designs in data-scarce domains. In this talk, I will present our integrated materials design framework, which is particularly effective in cases where limited prior data and incomplete physical understanding hinder the predictability of conventional machine learning models. Our approach leverages text mining and natural language processing techniques to extract and organize dispersed information from the literature, enabling the creation of initial materials databases. These databases serve as a foundation for training data-driven models that can be applied to screen vast unknown materials spaces. In some cases, this necessitates the development of novel representations of local structures. This virtual screening efficiently identifies promising materials families for further investigation, narrowing the candidate space for experimental validation. Within these identified materials families, we implement a Bayesian optimization based adaptive discovery workflow to search for compounds with optimal properties. We demonstrate how this approach accelerates the design of materials with critical functionalities, such as electronic metal-insulator transitions, novel ferroelectrics, and future interconnects. Finally, I will address key challenges in achieving materials design, including issues related to data quality and property-performance mismatches, and outline potential solutions synergistic with autonomous laboratories and digital twins.