Session Q43: Emergent Topics in Machine Learning for Molecular Systems and Materials

Ab initio-driven machine-learning models for aqueous systems, interfaces, and molten salts

This presentation focuses on recently developed machine-learning models based on ab initio quantum chemical methods. These allow simulations of aqueous systems, interfaces, and molten salts over time and length scales previously reachable only for classical force fields developed by fitting their macroscopic properties. In particular, we take advantage of the Deep Potential (DeePMD) methodology to capture energies and forces in a given system, as determined from quantum density functional theory. The DeePMD approach has the advantage that a well-developed software framework is available for rapid training of the relevant models and efficient implementation within standard open-source molecular dynamics codes. Example applications are presented here for water’s vapor-liquid and liquid-liquid phase behavior, fluid-phase properties of CO₂, and properties of aqueous electrolyte solutions that are hard to obtain accurately using empirical force fields. In addition, we investigate rates for bubble cavitation and homogeneous ice nucleation [7]. Further illustrations of the power of the DeePMD approach are provided by ongoing work on heterogeneous ice formation on microcline feldspar and the properties of carbonate and hydroxide melts relevant for high-temperature fuel cell operation. Ab initio-based machine learning models are shown to have excellent predictive capabilities for thermodynamic and transport properties of solutions and melts with no need for input from experimental data, they automatically include multibody, polarizability and charge transfer effects, and naturally describe chemical reactions. However, they are limited in accuracy by the underlying quantum chemical methods. They have complex, physically opaque model structures and require additional training for extending to mixtures. Despite these disadvantages, they likely represent the future of molecular modeling at the atomistic level of detail.   

Machine learning for molecular and materials science

We willl present our pastm current and future work on the set of Machine Learning potentials nicknamed ANI, which are able to compute energies and forces from structure, at a cost similar to a classical force field, but with accuracies of high level quantum mechanics. This breaks the old "you can be fast or accurate, but not both" problem in the field of molecular modeling, and allows us to study a number of problem that seemed intractable until a few years ago.

We will present new extensions such as charged systems, scailng to miilions of atoms and hundreds of GPUS, and to the calculations of other properties such as dipoles, NMR chemical shifts, IR spectra, etc.   

Self-assembly of electronic materials and the power of machine learning

There are many problems at the forefront of materials chemistry that are stymied by their inherent complexity. Such problems are characterized by a rich landscape of parameters and processing variables that is combinatorially too large for either an experimental or a computational approach to solve through an exhaustive search. In such cases, the usual approach is an Edisonian trial-and-error approach, which inevitably leaves areas of parameter space largely or wholly unexplored. The problems that we have explored are also characterized by a scarcity of data, since the data are expensive in time and resources to acquire, both experimentally and computationally. This makes it an ideal candidate to solve using a Bayesian optimization (BayesOpt) approach, which provides a strategy for a global optimization of "black box" functions lacking a functional form. For much of a decade, we have used a Bayesian optimization approach to study the solution processing of metal halide perovskites, a promising class of materials for solar cell development. Solution processing offers a low-energy-use and deceptively simple protocol to create electronically active thin films with high solar cell efficiency. In this talk, we will cover our accomplishments, challenges and outlook for what Bayesian optimization might achieve to help us understand, and hence control, these processes. I will end with some ideas of where we are taking BayesOpt in terms of having the ability to model nucleation and growth of metal halide perovskites and the algorithm development we will need to get us there. We will conclude with some observations on where the broader field of 'materials discovery' is headed.   

Exploration of New High-Entropy Materials Enabled by Quantum Computing

High-entropy materials (HEMs) represent a promising category of materials with multiprincipal elements and a wide range of molar ratios, offering novel solutions to critical challenges in energy and the environment ranging from climate change to semiconductor chip shortages. Within this material family, high-entropy catalysts, oxides, semiconductors, superconductors, ceramics, and more have gained prominence. The common challenge among these diverse frontiers lies in the selection of elements and their molar ratios across the extensive compositional space. In this talk, we will explore emerging quantum computing technologies, encompassing quantum simulators and hardware, to effectively address the complex task of elemental design and contribute to the discovery of novel HEMs. Furthermore, we will highlight the potential of quantum machine learning algorithms in expediting the training process. Lastly, we outline several prospective directions for HEM research that can benefit significantly from the transformative capabilities of quantum computing.   

Mary Jo Ondrechen

Machine Learning for protein function prediction and for understanding how enzymes work

Our Machine Learning methodology, Partial Order Optimum Likelihood (POOL) is used to predict biochemically active amino acids in the structures of gene products (proteins). Computed electrostatic and chemical properties of individual amino acids, derived only from the 3D structure of the query protein, serve as input features. Our most recent applications of POOL are described. From predicted local sites of biochemical activity, the biochemical functions of Structural Genomics proteins of unknown function are predicted by local structure matching of predicted spatial arrays of active amino acids with those of proteins of known function. Examples from the Haloacid Dehalogenase superfamily, with experimental testing of function by direct biochemical assay, are featured. POOL analysis also uncovers the types of interactions that enable protein structures to transform the weakly acidic or basic side chains of amino acids into strong acids, strong bases, or nucleophiles in the active site. Finally, the prediction and characterization of druggable sites is presented for SARS-CoV-2 targets.