Controlled Structure Formation of Sequence-Defined Materials with Reinforcement Learning

Property optimization of sequence-defined materials such as proteins, single-chain nanoparticles, and block co-polymers is simultaneously afforded and hindered by their vast design space. Additionally, the influence of time-dependent processing protocols on material properties introduces further complexity, necessitating efficient exploration strategies. In response, we introduce and test the integration of reinforcement learning (RL) with molecular dynamics (MD) simulations as an approach to navigate this complex space. Using an RL technique known as Q-learning, we dynamically generate processing protocols, initially targeting the radius of gyration in Lennard-Jones chains through temperature adjustments. We then extend our technique to manipulate sequence and chemical properties to achieve desired morphologies in single-chain nanoparticles. This methodology enhances our ability to refine material properties through precise control over sequence, chemistry, and processing protocols while providing deeper insights into their interdependencies. Our results demonstrate the potential of the RL-MD framework as a tool for optimizing properties of sequence-defined materials and elucidating key sequence determinants.