On the role of gradients for machine learning of molecular energies and\n forces

2020/07/19 by Anders S. Christensen, O. Anatole von Lilienfeld, Christensen, Anders S. +1 · 13 citations
Materials Science · Computer Science · Chemistry · #Machine Learning in Materials Science #Computational Drug Discovery Methods #Mass Spectrometry Techniques and Applications

paper · pdf · doi:10.48550/arxiv.2007.09593

Abstract

The accuracy of any machine learning potential can only be as good as the\ndata used in the fitting process. The most efficient model therefore selects\nthe training data that will yield the highest accuracy compared to the cost of\nobtaining the training data. We investigate the convergence of prediction\nerrors of quantum machine learning models for organic molecules trained on\nenergy and force labels, two common data types in molecular simulations. When\ntraining and predicting on different geometries corresponding to the same\nsingle molecule, we find that the inclusion of atomic forces in the training\ndata increases the accuracy of the predicted energies and forces 7-fold,\ncompared to models trained on energy only. Surprisingly, for models trained on\nsets of organic molecules of varying size and composition in non-equilibrium\nconformations, inclusion of forces in the training does not improve the\npredicted energies of unseen molecules in new conformations. Predicted forces,\nhowever, also improve about 7-fold. For the systems studied, we find that force\nlabels and energy labels contribute equally per label to the convergence of the\nprediction errors. Choosing to include derivatives such as atomic forces in the\ntraining set or not should thus depend on, not only on the computational cost\nof acquiring the force labels for training, but also on the application domain,\nthe property of interest, and the desirable size of the machine learning model.\nBased on our observations we describe key considerations for the creation of\ndatasets for potential energy surfaces of molecules which maximize the\nefficiency of the resulting machine learning models.\n

Cited by

• Field theoretic atomistics: Learning thermodynamic and variational surrogate to density functional theory
• Magnitude-Modulated Equivariant Adapter for Parameter-Efficient Fine-Tuning of Equivariant Graph Neural Networks
• Fast and Scalable Evaluation of Unbiased Atomic Forces in ab initio Variational Monte Carlo via the Lagrangian Technique
• Learning 3D Anisotropic Noise Distributions Improves Molecular Force Field Modeling
• How Accurate Are DFT Forces? Unexpectedly Large Uncertainties in Molecular Datasets
• Gradients not needed: ML-driven propagation of nonadiabatic molecular dynamics without reference gradients
• ATOM: A Pretrained Neural Operator for Multitask Molecular Dynamics
• Shoot from the HIP: Hessian Interatomic Potentials without derivatives
• Global properties of the energy landscape: a testing and training arena for machine learned potentials
• Multi-head committees enable direct uncertainty prediction for atomistic foundation models
• GEDAN: Learning the Edit Costs for Graph Edit Distance
• A Scalable and Quantum-Accurate Foundation Model for Biomolecular Force Field via Linearly Tensorized Quadrangle Attention
• The Price of Freedom: Exploring Expressivity and Runtime Tradeoffs in Equivariant Tensor Products

Related

• Simple and efficient algorithms for training machine learning potentials to force data
• The dark side of the forces: assessing non-conservative force models for atomistic machine learning
• Machine learning force fields and coarse-grained variables in molecular\n dynamics: application to materials and biological systems
• Ensemble Learning of Machine Learning Force Fields
• Synergistic integration of physical embedding and machine learning enabling precise and reliable force field