Material Design with Human-in-the-Loop Reinforcement Learning and autonomous AI agents

O. Isayev

We proposed a novel ML-guided materials discovery platform that combines synergistic innovations in automated flow synthesis and AI agents. A software-controlled, continuous polymer synthesis platform enables rapid iterative experimental–computational cycles that result in the synthesis of hundreds of unique copolymer compositions within a multi-variable compositional space. The non-intuitive design criteria identified by ML, accomplished by exploring less than 0.9% of overall compositional space, and led to the identification of >10 copolymer compositions that outperformed state-of-the-art materials. Under the RL paradigm, an agent(s) is trained to select actions that maximize the cumulative sum of rewards, which, in the context of chemical discovery, is often consistent with a target property, structural feature, or function. RL agents can learn to suggest synthesis protocols, potential reactants, and experimental conditions by training via value-based or policy-based iterative schemes. These findings demonstrate that machine-guided, human-augmented design is a powerful strategy for accelerating polymer discovery in applications where data is scarce and expensive to acquire, with broad applicability to multi-objective materials optimization.

Post Semiempirical Quantum Chemistry Methods

J. Keith, B. Zulueta, M. Plum, C. Mhatre, M. Marciesky, V. Kam

Accurate quantum chemical predictions remain computationally expensive for navigating large chemical spaces. This talk will report our recent advances in developing post-semiempirical quantum mechanics (Post-SEQM) methods that leverage well-conditioned minimum basis set orbital populations and exploit tight-binding relationships across chemical space. We will start by focusing on the Bond Energy from Bond Orders and Populations (BEBOP) approach, which is rooted in tight-binding theory. BEBOP delivers surprisingly accurate approximate atomization energies that are decomposed into intuitive intramolecular bond energy contributions from a single self-consistent quantum chemistry calculation. We will highlight an application of BEBOP in chemical design of processes for plastic recycling. We will also introduce extension to BEBOP theory that are useful for predicting thermochemically relevant vibrational energy contributions without the computational burden of Hessian calculations. Finally, I will show progress in developing the Pittsburgh Approximate Molecular Orbital (PAMO) method that strives to reproduce chemical bonding from Hartree-Fock theory with a large basis set. These topics will be presented to demonstrate how Post-SEQM methods can accelerate computational screening and rational design across chemical compound space.