How robust are modern graph neural network potentials in long and hot molecular dynamics simulations?

Graph neural networks (GNNs) have emerged as a powerful machine learning approach for the prediction of molecular properties. In particular, recently proposed advanced GNN models promise quantum chemical accuracy at a fraction of the computational cost. While the capabilities of such advanced GNNs have been extensively demonstrated on benchmark datasets, there have been few applications in real atomistic simulations. Here, we therefore put the robustness of GNN interatomic potentials to the test, using the recently proposed GemNet architecture as a testbed. Models are trained on the QM7-x database of organic molecules and used to perform extensive molecular dynamics simulations. We find that low test set errors are not sufficient for obtaining stable dynamics and that severe pathologies sometimes only become apparent after hundreds of ps of dynamics. Nonetheless, highly stable and transferable GemNet potentials can be obtained with sufficiently large training sets.

[1]  Johannes T. Margraf,et al.  Kernel charge equilibration: efficient and accurate prediction of molecular dipole moments with a machine-learning enhanced electron density model , 2021, Mach. Learn. Sci. Technol..

[2]  Jonathan P. Mailoa,et al.  E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials , 2021, Nature Communications.

[3]  Volker L. Deringer,et al.  Gaussian Process Regression for Materials and Molecules , 2021, Chemical reviews.

[4]  Johannes T. Margraf,et al.  On the role of long-range electrostatics in machine-learned interatomic potentials for complex battery materials , 2021, ACS Applied Energy Materials.

[5]  Ho Won Jang,et al.  Catalyze Materials Science with Machine Learning , 2021, ACS Materials Letters.

[6]  Jörg Behler,et al.  Machine learning potentials for extended systems: a perspective , 2021, The European Physical Journal B.

[7]  P. Hu,et al.  Accelerating Metadynamics-Based Free-Energy Calculations with Adaptive Machine Learning Potentials. , 2021, Journal of chemical theory and computation.

[8]  Florian Becker,et al.  GemNet: Universal Directional Graph Neural Networks for Molecules , 2021, NeurIPS.

[9]  K. Reuter,et al.  Nano‐Scale Complexions Facilitate Li Dendrite‐Free Operation in LATP Solid‐State Electrolyte , 2021, Advanced Energy Materials.

[10]  Klaus-Robert Müller,et al.  SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects , 2021, Nature Communications.

[11]  A. Tkatchenko,et al.  Combining Machine Learning and Computational Chemistry for Predictive Insights Into Chemical Systems , 2021, Chemical reviews.

[12]  Michael Gastegger,et al.  Equivariant message passing for the prediction of tensorial properties and molecular spectra , 2021, ICML.

[13]  Volker L. Deringer,et al.  Origins of structural and electronic transitions in disordered silicon , 2021, Nature.

[14]  Michael Gastegger,et al.  Machine Learning Force Fields , 2020, Chemical reviews.

[15]  Jonas A. Finkler,et al.  A fourth-generation high-dimensional neural network potential with accurate electrostatics including non-local charge transfer , 2020, Nature Communications.

[16]  A. Tkatchenko,et al.  QM7-X: A comprehensive dataset of quantum-mechanical properties spanning the chemical space of small organic molecules , 2020, 2006.15139.

[17]  Johannes T. Margraf,et al.  Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules , 2020, ArXiv.

[18]  Gábor Csányi,et al.  Machine learning in chemical reaction space , 2020, Nature Communications.

[19]  Weihua Hu,et al.  The Open Catalyst 2020 (OC20) Dataset and Community Challenges , 2020, ACS Catalysis.

[20]  Yu Wang,et al.  IrO_{2} Surface Complexions Identified through Machine Learning and Surface Investigations. , 2020, Physical review letters.

[21]  Mordechai Kornbluth,et al.  Accurate and scalable multi-element graph neural network force field and molecular dynamics with direct force architecture , 2020, 2007.14444.

[22]  O. Anatole von Lilienfeld,et al.  On the role of gradients for machine learning of molecular energies and forces , 2020, Mach. Learn. Sci. Technol..

[23]  Zachary W. Ulissi,et al.  Accelerated discovery of CO2 electrocatalysts using active machine learning , 2020, Nature.

[24]  Stephan Günnemann,et al.  Directional Message Passing for Molecular Graphs , 2020, ICLR.

[25]  Christoph Ortner,et al.  Incompleteness of Atomic Structure Representations. , 2020, Physical review letters.

[26]  M. Ceriotti,et al.  Evidence for supercritical behaviour of high-pressure liquid hydrogen , 2019, Nature.

[27]  Justin S. Smith,et al.  Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning , 2019, Nature Communications.

[28]  Markus Meuwly,et al.  PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments, and Partial Charges. , 2019, Journal of chemical theory and computation.

[29]  Jerzy Leszczynski,et al.  Accurate and transferable multitask prediction of chemical properties with an atoms-in-molecules neural network , 2018, Science Advances.

[30]  K-R Müller,et al.  SchNetPack: A Deep Learning Toolbox For Atomistic Systems. , 2018, Journal of chemical theory and computation.

[31]  K. Müller,et al.  Towards exact molecular dynamics simulations with machine-learned force fields , 2018, Nature Communications.

[32]  Klaus-Robert Müller,et al.  SchNet: A continuous-filter convolutional neural network for modeling quantum interactions , 2017, NIPS.

[33]  Gerbrand Ceder,et al.  Efficient and accurate machine-learning interpolation of atomic energies in compositions with many species , 2017, 1706.06293.

[34]  Klaus-Robert Müller,et al.  Machine learning of accurate energy-conserving molecular force fields , 2016, Science Advances.

[35]  Alexandre Tkatchenko,et al.  Quantum-chemical insights from deep tensor neural networks , 2016, Nature Communications.

[36]  Michele Ceriotti,et al.  Beyond static structures: Putting forth REMD as a tool to solve problems in computational organic chemistry , 2015, J. Comput. Chem..

[37]  Jörg Behler,et al.  Constructing high‐dimensional neural network potentials: A tutorial review , 2015 .

[38]  Pavlo O. Dral,et al.  Quantum chemistry structures and properties of 134 kilo molecules , 2014, Scientific Data.

[39]  R. Kondor,et al.  On representing chemical environments , 2012, 1209.3140.

[40]  Jean-Louis Reymond,et al.  Enumeration of 166 Billion Organic Small Molecules in the Chemical Universe Database GDB-17 , 2012, J. Chem. Inf. Model..

[41]  A. Tkatchenko,et al.  Accurate and efficient method for many-body van der Waals interactions. , 2012, Physical review letters.

[42]  A. Tkatchenko,et al.  Resolution-of-identity approach to Hartree–Fock, hybrid density functionals, RPA, MP2 and GW with numeric atom-centered orbital basis functions , 2012, 1201.0655.

[43]  R. Kondor,et al.  Gaussian approximation potentials: the accuracy of quantum mechanics, without the electrons. , 2009, Physical review letters.

[44]  Matthias Scheffler,et al.  Ab initio molecular simulations with numeric atom-centered orbitals , 2009, Comput. Phys. Commun..

[45]  Michele Parrinello,et al.  Generalized neural-network representation of high-dimensional potential-energy surfaces. , 2007, Physical review letters.

[46]  K. Burke,et al.  Rationale for mixing exact exchange with density functional approximations , 1996 .

[47]  O. Isayev,et al.  ANI-1: an extensible neural network potential with DFT accuracy at force fi eld computational cost † , 2017 .