Molecular excited states through a machine learning lens

Abstract

Theoretical simulations of electronic excitations and associated processes in molecules are indispensable for fundamental research and technological innovations. However, such simulations are notoriously challenging to perform with quantum mechanical (QM) methods. Advances in machine learning (ML) open many new avenues for assisting molecular excited-state simulations. In this Review, we track such progress, assess the current state-ofthe-art and highlight the critical issues to solve in the future. We overview a broad range of ML applications in excited-state research, which include the prediction of molecular properties, improvements of QM methods for the calculations of excited-state properties and the search of new materials. ML approaches can help us understand hidden factors that influence photo-This is a post-peer-review, pre-copyedit version of an article published in Nature Reviews Chemistry.

Choosing the right molecular machine learning potential

“…Owing to its generalization capabilities and fast prediction on unseen data, MLPs can be explored to accelerate minimum-energy [10][11][12][13][14][15] and transition-state structure search, 13,16,17 vibrational analysis, 18-21 absorption 22,23 and emission spectra simulation, 24 reaction 13,25,26 and structural transition exploration, 27 and ground- [3][4][5][6][7][8][9] and excitedstate dynamics propagation. 28,29 A blessing and a curse of ML is that it is possible to design, for all practical purposes, an infinite number of MLP models that can describe a molecular PES. These MLP models are usually built from two main components: the ML algorithm and its input X, the molecular descriptor.…”

Choosing the right molecular machine learning potential

“…Owing to its generalization capabilities and fast prediction on unseen data, MLPs can be explored to accelerate minimum-energy [10][11][12][13][14][15] and transition-state structure search, 13,16,17 vibrational analysis, 18-21 absorption 22,23 and emission spectra simulation, 24 reaction 13, 25,26 and structural transition exploration, 27 and ground- [3][4][5][6][7][8][9] and excitedstate dynamics propagation. 28,29 A blessing and a curse of ML is that it is possible to design, for all practical purposes, an infinite number of MLP models that can describe a molecular PES. These MLP models are usually built from two main components: the ML algorithm and its input X, the molecular descriptor.…”

Perspective on integrating machine learning into computational chemistry and materials science

“…3, existing atomistic ML packages such as AMP, 155 sGDML 156 or SchNet-Pack 45,121 could be interfaced with electronic structure packages that heavily expose internal routines (e.g., FHIaims, 157 PSI4, 158 or PySCF 159 ) and be used alongside dynamics packages such as i-Pi 160 and SHARC, 161,162 as well as algebra and electronic structure libraries such as ELSI 2 and ESL. 1 The structure generation, workflow and parser tool Atomic Simulation Environment (ASE) 163 is for example already interfaced with the above examples of AMP and SchNetPack. This could also involve a closer integration with existing data repositories such as NO-MAD, 164,165 the Materials Project, 164,165 the MolSSI QC Archive 166 or the Quantum Machine repository.…”

“…Algorithms for molecular geometry optimization, efficient molecular dynamics simulations, and electronic structure calculations perform highly specialized tasks while being massively scalable and parallelized across a diverse range of hardware architectures. 1,2 Simultaneously, PhD graduates in the field have been trained to be expert users of existing and developers of new simulation workflows. This is the status quo at the time when machine learning (ML) methods enter the stage.…”

“…The application of ML to atomistic simulation and electronic structure theory has been developing rapidly since its earliest works in a modern context. [3][4][5][6][7][8][9][10][11] A number of excellent reviews have recently been written to a) Electronic mail: r.maurer@warwick.ac.uk highlight progress in various contexts including the role of ML in catalyst design, 12,13 in the development of forcefields and interatomic potentials for ground state properties [14][15][16][17][18][19] and excited states, [20][21][22] in quantum chemistry, 23,24 in finding solutions to the Schrödinger equation, 25 and the role of unsupervised learning in atomistic simulation 26 (see Table I for a non-exhaustive list).…”