Over the last decade, we have witnessed the emergence of ever more machine learning applications in all aspects of the chemical sciences. Here, we highlight specific achievements of machine learning models in the field of computational chemistry by considering selected studies of electronic structure, interatomic potentials, and chemical compound space in chronological order.
Accurate solutions of the Schrödinger equation for the electrons in molecules and materials would vastly enhance our capability for chemical discovery, but computational cost makes this prohibitive. Since Dirac first exhorted us to find suitable approximations to bypass this cost^{1}, much progress has been made, but much remains out of reach for the foreseeable future. The central promise of machine learning (ML) is that, by exploiting statistical learning of the properties of a few cases, we might leapfrog over the worst bottlenecks in this process.
As visible from the publication record in the field (Fig. 1), over the decade since Nature Communications first appeared, machine learning has gained increasing traction in the hard sciences^{2}, and has found many applications in atomistic simulation sciences^{3}. Here, we focus on the progress achieved in the last decade on three interrelated topics (i) electronic structure theory, broadly defined, (ii) universal force field models, as used for vibrational analysis or molecular dynamics applications, and (iii) first principlesbased approaches enabling the exploration of chemical compound space.
Basic challenges
The central challenge of Schrödinger space is to use supervised learning from examples to find patterns that either accelerate or improve upon the existing human algorithms behind these technologies. In density functional theory (KSDFT), this most often means improved approximate functionals; in quantum Monte Carlo (QMC), this is faster ways to find variational wavefunctions; in ab initio quantum chemistry such as coupled cluster considering single, double, and perturbative triple excitations (CCSD(T)), this is learned predictions of wavefunction amplitudes instead of recalculation for every system.
In the condensed phase, molecular dynamics simulations yield a vast amount of useful thermodynamic and kinetic properties. Classical force fields cost little to run, but are often accurate only around the equilibrium. The only firstprinciples alternative is Kohn–Sham density functional theory (DFT), but its computational cost vastly reduces what is practical. A central challenge of configuration space is therefore to produce energies and forces from a classical potential of accuracies comparable to DFT (at least) via training on DFTcalculated samples, possibly for just one element, but with hundreds to thousands of atoms in unique bonding (and bondbreaking) arrangements.
Finally, the challenge of chemical compound space is to explore all useful combinations of distinct atoms. The number of stable combinations is often astronomical. The central aim is to train on quantumchemical examples, and create a ML algorithm that can, given a configuration of atoms, generate the atomization energy without running, e.g., a DFT calculation, in order to scan the vast unknown of unsynthesized molecules for desirable functionalities.
These challenges are hierarchical. Progress in creating better density functionals clearly impacts finding accurate forces for molecular dynamics and accurate searching of chemical compound space. Finding a way to learn molecular energies with fewer examples is useful for chemical compound space, but forces would also be needed to run molecular dynamics, and selfconsistent densities to run orbitalfree DFT. The challenges are also overlapping: improved density functionals may be irrelevant if ML force fields can be trained on CCSD(T) energies and forces.
Progress with machine learning
Schrödinger space
Within DFT, the focus is usually on the everelusive exchangecorrelation (XC) energy^{4}, which is needed as a functional of the spin densities. An ‘easier’ target is orbitalfree (OF) DFT, which tries to find the kinetic energy of Kohn–Sham electrons, to bypass the need to solve the KohnSham equations. A primary question is: can machines find better density functional approximations than those created by people? Two distinct approaches are to improve the accuracy of existing humandesigned approximations or to create entirely new machinelearned approximations that overcome qualitative failures of our present approximations. Often tests are first performed on model systems, and later applied to more realistic firstprinciples Hamiltonians.
In orbitalfree DFT, Snyder et al.^{5} used KernelRidgeRegression (KRR) on a onedimensional model of a molecule an machinelearned functional for OF DFT that breaks bonds correctly, which has been successively built upon^{6}. Brockherde et al.^{7} showed how KRR could be applied by finding densities directly from potentials (the HohenbergKohn map) avoiding functional derivatives. The problem of XC is harder. Nagai et al.^{8} showed that accurate densities of just three small molecules are sufficient to create machinelearned approximations that are comparable to those created by people. In ab initio quantum chemistry, Welborn et al.^{9} have shown how to use features from HartreeFock calculations to accurately predict CCSD energies, while an intriguing alternative is to map to spin problems and use a restricted Boltzmann machine^{10}. In the last year, two new applications for finding wavefunctions within QMC have appeared^{11,12}.
While many avenues are being explored, there is as yet no clearly improved, generalpurpose MLdesigned density functional, MLpowered QMC, or ML approach to ab initio quantum chemistry available to the general user. But for such a complex problem, progress is measured in decades, and we are reasonably confident that such codes could appear over the next five years.
Configuration space
Machine learning models for exploring configurational spaces yield rapid force predictions for extended molecular dynamics simulations. While surrogate models of interatomic potentials using neural networks were firmly established before 2010^{13}, Csanyi, Bartok and coworkers used KRR in their seminal ’GaussianApproximated Potential’ (GAP) method, relying on Gaussian kernel functions and an atom index invariant bispectrum representation^{14}. In 2013, the first flavor of the smooth overlap of atomic positions (SOAP) representation for KRR based potentials was published^{15}. First stepping stones towards universal forcefield, trained ‘onthefly’ or throughout the chemical space of molecules displaced along their normal modes, were established in ref. ^{16,17}. KRR based forcefield models with CCSD(T) accuracy were introduced in 2017^{18}, and based on Behler’s atomcentered symmetry function representations in neural networkbased potentials tremendous progress was made^{19} enabling Smith et al. to train an Accurate Neural network engIne (ANI) on millions of configurations of tens of thousands of organic molecules distorted along aforementioned normal mode displacements^{20}. Impactful applications include KRR potentials used to model challenging processes in ferromagnetic iron^{21}, or Weinan E, Car and coworkers using the Summit supercomputer to simulate 100 million atoms of water with ab initio accuracy using convolutional neural networks^{22}.
Chemical compound space
The idea of using machine learning to mine ab initio materials data bases dates back to 2010 in seminal work by Hautier et al.^{23}. Starting with the Coulombmatrix^{24}, the development of a selection of ever improved machine learning models (due to improved representations and/or regressor architectures) is exemplified^{25} on atomization energies of the Quantum Mechanics results for organic molecules with up to 9 heavy atoms (QM9) data set^{26}, as shown in Fig. 2 “QM9IPAMchallenge”. Such singlepoint energy calculations typically dominate the cost of quantum chemistry compute campaigns, and therefore a vital minimal target for surrogate models.
Examples of improvements of understanding compound space include the discovery of an elpasolite crystal containing aluminum atoms with negative oxidation state^{27}, polarizability models using tensorial learning^{28}, or predicting solvation and acidity in complex mixtures^{29}.
Summary and outlook
Much has happened over the last decade, touching on nearly all aspects of atomistic simulations. Our selection of areas (electronic structure, interatomic potentials, and chemical space) and studies mentioned does not do justice to the overall impact machine learning has had on nearly all branches of the atomistic sciences. Much of the more important work first appeared in rather technical journals such as the Journal of Chemical Physics or Physical Review Letters and is already heavily cited. More recent advances were published in broader journals such as Science, PNAS or Nature and Nature Communications. Some of the outstanding challenges in the field include (i) improved quantum chemistry methods which can reliably cope with reaction barriers, d and felements, magnetic and excited states, as well as redox properties of systems in any aggregation state, (ii) extensive highquality data sets covering many properties over wide swaths of structural and compositional degrees of freedom, and (iii) the removal of hidden and unconscious biases. Extrapolating from the past, the future looks bright: Longstanding problems have been and are being tackled successfully, and new capabilities are always appearing. Likely, the community will soon address challenges that previously were simply considered to be prohibitively complex or demanding, such as automatized experimentation or synthesis of new materials and molecules on demand.
References
Dirac, P. A. M. Quantum mechanics of manyelectron systems. Proc. R. Soc. Lond. Series A, Containing Papers of a Mathematical and Physical Character 123, 714–733 (1929).
Schütt, K. et al. Machine Learning Meets Quantum Physics. Lecture Notes in Physics (Springer International Publishing, 2020).
von Lilienfeld, O. A., Müller, K.R. & Tkatchenko, A. Exploring chemical compound space with quantumbased machine learning. Nat. Rev. Chem. 4, 347–358 (2020).
Kohn, W. & Sham, L. J. Selfconsistent equations including exchange and correlation effects. Phys. Rev. 140, A1133–A1138 (1965).
Snyder, J. C. et al. Orbitalfree bond breaking via machine learning. J. Chem. Phys. 139, 224104 (2013).
Yao, K. & Parkhill, J. Kinetic energy of hydrocarbons as a function of electron density and convolutional neural networks. J. Chem. Theory Comput. 12, 1139–1147 (2016).
Brockherde, F. et al. Bypassing the KohnSham equations with machine learning. Nat. Commun. 8, 872 (2017).
Nagai, R., Akashi, R. & Sugino, O. Completing density functional theory by machine learning hidden messages from molecules. npj Comput. Mater. 6, 43 (2020).
Welborn, M., Cheng, L. & Miller, T. F. Transferability in machine learning for electronic structure via the molecular orbital basis. Journal of Chem. Theory Comput. 14, 4772–4779 (2018).
Choo, K., Mezzacapo, A. & Carleo, G. Fermionic neuralnetwork states for abinitio electronic structure. Nat. Commun. 11, 2368 (2020).
Pfau, D., Spencer, J. S., de G. Matthews, A. G. & Foulkes, W. M. C. Abinitio solution of the manyelectron Schrödinger equation with deep neural networks. Preprint athttp://arXiv.org/abs/1909.02487 (2019).
Hermann, J., Schätzle, Z. & Noé, F. Deep neural network solution of the electronic Schrödinger equation. Preprint at http://arXiv.org/abs/1909.08423 (2019).
Behler, J. & Parrinello, M. Generalized neuralnetwork representation of highdimensional potentialenergy surfaces. Phys. Rev. Lett. 98, 146401 (2007).
Bartók, A. P., Payne, M. C., Kondor, R. & Csányi, G. Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons. Phys. Rev. Lett. 104, 136403 (2010).
Bartók, A. P., Kondor, R. & Csányi, G. On representing chemical environments. Phys. Rev. B 87, 184115 (2013).
Li, Z., Kermode, J. R. & De Vita, A. Molecular dynamics with onthefly machine learning of quantummechanical forces. Phys. Rev. Lett. 114, 096405 (2015).
Rupp, M., Ramakrishnan, R. & von Lilienfeld, O. A. Machine learning for quantum mechanical properties of atoms in molecules. J. Phys. Chem. Lett. 6, 3309 (2015).
Chmiela, S. et al. Machine learning of accurate energyconserving molecular force fields. Sci. Adv. 3, e1603015 (2017).
Behler, J. First principles neural network potentials for reactive simulations of large molecular and condensed systems. Angewandte Chemie Int. Edn. 56, 12828–12840 (2017).
Smith, J. S., Isayev, O. & Roitberg, A. E. Ani1: an extensible neural network potential with dft accuracy at force field computational cost. Chem. Sci. 8, 3192–3203 (2017).
Dragoni, D., Daff, T. D., Csányi, G. & Marzari, N. Achieving dft accuracy with a machinelearning interatomic potential: thermomechanics and defects in bcc ferromagnetic iron. Phys. Rev. Mater. 2, 013808 (2018).
Jia, W. et al. Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning. Preprint at http://arXiv.org/abs/2005.00223 (2020).
Hautier, G., Fischer, C. C., Jain, A., Mueller, T. & Ceder, G. Finding nature’s missing ternary oxide compounds using machine learning and density functional theory. Chem. Mater. 22, 3762 (2010).
Rupp, M., Tkatchenko, A., Müller, K.R. & von Lilienfeld, O. A. Fast and accurate modeling of molecular atomization energies with machine learning. Phys. Rev. Lett. 108, 058301 (2012).
Faber, F. A., Christensen, A. S. & von Lilienfeld, O. A. In Machine Learning Meets Quantum Physics, (eds Schütt, K. T. et al.) 155–169 (Springer, 2020).
Ramakrishnan, R., Dral, P., Rupp, M. & von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data 1, 140022 (2014).
Faber, F. A., Lindmaa, A., von Lilienfeld, O. A. & Armiento, R. Machine learning energies of 2 million elpasolite (ABC_{2}D_{6}) crystals. Phys. Rev. Lett. 117, 135502 (2016).
Wilkins, D. M. et al. Accurate molecular polarizabilities with coupled cluster theory and machine learning. Proc. Natl Acad. Sci. USA 116, 3401–3406 (2019).
Rossi, K. et al. Simulating solvation and acidity in complex mixtures with firstprinciples accuracy: The case of CH_{3}SO_{3}H and H_{2}O_{2} in phenol. J. Chem. Theory Comput. 16, 5139–5149 (2020).
Rupp, M., von Lilienfeld, O. A. & Burke, K. Guest editorial: Special topic on dataenabled theoretical chemistry. J. Chem. Phys. 148, 241401 (2018).
Acknowledgements
O.A.v.L. acknowledges funding from the Swiss National Science foundation (407540_167186 NFP 75 Big Data, 200021_175747, NCCR MARVEL) and from the European Research Council (ERCCoG grant QML). K.B. is supported by NSF CHE 1856165.
Author information
Authors and Affiliations
Contributions
Both authors conceived, discussed, and wrote this article.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
von Lilienfeld, O.A., Burke, K. Retrospective on a decade of machine learning for chemical discovery. Nat Commun 11, 4895 (2020). https://doi.org/10.1038/s41467020185569
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41467020185569
Further reading

Machine learning for exploring small polaron configurational space
npj Computational Materials (2022)

BIGDML—Towards accurate quantum machine learning force fields for materials
Nature Communications (2022)

Big data and machine learning for materials science
Discover Materials (2021)

Ethical Artificial Intelligence in Chemical Research and Development: A Dual Advantage for Sustainability
Science and Engineering Ethics (2021)