Combining Machine Learning and Computational Chemistry for Predictive Insights Into Chemical Systems

Abstract

Machine learning models are poised to make a transformative impact on chemical sciences by dramatically accelerating computational algorithms and amplifying insights available from computational chemistry methods. However, achieving this requires a confluence and coaction of expertise in computer science and physical sciences. This review is written for new and experienced researchers working at the intersection of both fields. We first provide concise tutorials of computational chemistry and machine learning methods, showing how insights involving both can be achieved. We then follow with a critical review of noteworthy applications that demonstrate how computational chemistry and machine learning can be used together to provide insightful (and useful) predictions in molecular and materials modeling, retrosyntheses, catalysis, and drug design.

Orbital Mixer: Using Atomic Orbital Features for Basis Dependent Prediction of Molecular Wavefunctions

“…Machine learning for molecular inference has experienced impressive success in recent years, showcasing spectacular predictive accuracy enabled by large quantities of ab initio data, the incorporation of prior physical and chemical knowledge, and invariant and/or equivariant architectures (Keith et al, 2021). A common paradigm of these works interprets molecules as connected graphs and uses message passing to model interactions as a function of single-particle contributions.…”

“…An explosion of interest has surrounded applying machine learning (ML) methods to quantum chemistry with a plethora of interesting application areas such as learning interatomic potentials (Behler & Parrinello, 2007;Unke et al, 2021c;Bartók et al, 2010;Smith et al, 2017;Chmiela et al, 2017;2018;Schütt et al, 2018;Unke & Meuwly, 2019;Unke et al, 2021b;Batzner et al, 2021;Klicpera et al, 2020;Liu et al, 2021a;Schütt et al, 2021), constructing density functionals (Snyder et al, 2012;Brockherde et al, 2017;Ryczko et al, 2019;Kalita et al, 2021;Li et al, 2021), predicting spectroscopic properties (Gastegger et al, 2017;Westermayr & Marquetand, 2020), optoelectronic properties (Lee et al, 2021;Mazouin et al, 2021;Lu et al, 2020;Gladkikh et al, 2020), activation energies (Lewis-Atwell et al, 2021;Grambow et al, 2020), and a variety of physical properties throughout chemical compound space (Montavon et al, 2013;De et al, 2016;von Lilienfeld et al, 2020;Keith et al, 2021;Liu et al, 2021b;Tielker et al, 2021;Bratholm et al, 2021). Quantum chemistry workflows can obtain such chemical and physical information by modelling the electronic Schrodinger equation in a chosen basis set of localized atomic orbitals that is then used to derive the ground-state molecular wavefunction.…”

Towards Electronic Structure-Based Ab-Initio Molecular Dynamics Simulations with Hundreds of Millions of Atoms

“…In classical simulations they are computed by conventional force fields, or novel neural network and machine learning potentials, which have been parameterized to reproduce experimental or accurate ab-initio data of small model systems [3], [4]. Even though great strides in improving such empirical potentials have been made and often render them surprisingly accurate [5], [6], the transferability to systems or regions of the phase diagram different from the ones to which they have been trained in the first place may be restricted. Ultimately, when assuming a classical model, as ingenious it may be, the access to the quantum mechanical electronic structure is irrevocably lost.…”

Crystal Twins: Self-supervised Learning for Crystalline Material Property Prediction

“…Characterized by their computational efficiency and accuracy, these methods are capable of faster high-throughput screening compared to classical physics models. 1,2 This capability has roots in both novel learning algorithms and improved hardware. Even though ML models can offer faster predictions, the accuracy of these models is highly correlated with the availability of clean labeled data.…”