Machine learning accelerated computational design of materials and processes

T.P.M. Goumans, M. Hellström, P.S.N. Onofrio, N. Aguirre, R. Rüger
Software for Chemistry & Materials,

Keywords: computational chemistry, materials modeling, machine learning, multi-scale modeling


Machine learning (ML) can accelerate materials research & discovery in several ways, often in complimentary approaches. We will discuss a few directions where we can leverage ML in the context of atomistic modeling for designing new materials and their integrated use in devices and chemical reactors. The central framework in the Amsterdam Modeling Suite (AMS) enables the exploration of potential energy surfaces (PESs), mechanical, and electronic properties at several levels of theory. The central AMS driver supports advanced PES explorations, molecular dynamics (MD) and Grand Canonical Monte Carlo (GCMC). On-the-fly machine learned potentials such as FLARE[1] and universal graph neural network potentials such as M3GNet[2] can immediately be used for simulations such as chemical vapor deposition with the molecule gun in AMS. The ParAMS module furthermore provides a comprehensive framework to build training data and optimize machine learned potentials, as well as ReaxFF and DFTB parameters. With different levels of electronic structure methods available in AMS, we are exploring ML methods to predict properties more efficiently for molecular materials. Examples for OLED applications include training DFTB transfer integrals on DFT data, TDDFT(B) luminescence and excitonic properties on accurate qsGW+BSE calculations, and predicting novel molecules that have the desired optical and electronic properties yielding the best OLED device performance in multiscale simulations.[3] For catalysis, we can accelerate the multiscale workflow[4] through employing machine learning for reaction exploration, and by building surrogate models for a faster integration between kinetic Monte Carlo and Computational Fluid Dynamics. We will briefly discuss future directions relevant to battery materials and polymers, where we can harness the powerful combination of atomistic modeling with ML. [1] J. Vandermause et al. npj Computational Materials 6, 20 (2020) [2] C. Chen, S. Ong, Arxiv (2022) [3] [4]