Accelerating Materials Recipe Acquisition via LLM-Mediated Reinforcement Learning

Artificial intelligence tools have empowered advancements in materials research, from accelerating the forward prediction of properties to enabling inverse design of materials with desired properties. The impacts of such prediction methods are broad, for example predicting beam buckling under compression has been demonstrated on both simple homogeneous plastics and more esoteric biopolymer composites - such as bread. Successes in generating material structures possessing desired mechanical properties has also been previously demonstrated with machine learning workflows, for instance hierarchical lattices with specified stiffness or graphene sheets with bespoke fracture paths. However, the broader question of defining the “recipe” to acquire a desired material remains. While the “recipe” for certain material structures may be trivial, for example a macroscopic lattice design can simply be 3D printed, others such as bespoke graphene sheets are considerably more difficult to synthesize. Materials generated as a result of chemical reactions require a precise set of instructions with numerous differing parameters including reactant ratios, mixing procedures, temperatures, and reaction times. Here, we employ recent advances in large language models to organize how differences in materials synthesis “recipes” manifest as differences in property. Then, we use a reinforcement learning approach to learn how to rapidly navigate the space of potential recipes relevant for obtaining a desired resultant property. In doing so, we aim to accelerate the scientific workflow for materials discovery not only in the theoretical identification of a material but also in the practical implementation of how to acquire it. Specifically, we will demonstrate the approach for a class of materials recipes that the public is well acquainted with, yet still has direct parallels to a laboratory setting - recipes for baked goods.