Source-linked AI summary
PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments and Partial Charges
Oliver T. Unke, Markus Meuwly
TL;DR
Computationally solving the electronic Schrödinger equation remains demanding, limiting accurate quantum-chemical calculations to relatively small systems. This paper introduces PhysNet, a physics-informed neural network, and shows state-of-the-art benchmark performance, improved long-range potential-energy descriptions through explicit electrostatics, and accurate generalization from small fragments to Ala10.
Problem
Solving the electronic Schrödinger equation remains computationally demanding and is tractable only for a limited number of atoms.
Method
PhysNet is a physics-informed message-passing neural network that predicts chemical energies, forces, and dipole moments while explicitly incorporating electrostatic interactions.
Results
PhysNet matches or improves state-of-the-art performance across tested benchmarks, correctly describes long-range reaction-PES behavior, and generalizes from small fragments to Ala10 with 0.21 Å RMSD.
Takeaways & Limitations
Explicit physical modeling improves qualitative long-range PES predictions, while systematic training on small reference structures can support generalization to larger molecules with similar structural motifs.
Takeaways & Limitations
Local interactions beyond the cutoff cannot be fully represented without explicitly adding known long-range terms such as electrostatics or dispersion corrections.
Abstract
from arXiv · showhide
In recent years, machine learning (ML) methods have become increasingly popular in computational chemistry. After being trained on appropriate ab initio reference data, these methods allow to accurately predict the properties of chemical systems, circumventing the need for explicitly solving the electronic Schrödinger equation. Because of their computational efficiency and scalability to large datasets, deep neural networks (DNNs) are a particularly promising ML algorithm for chemical applications. This work introduces PhysNet, a DNN architecture designed for predicting energies, forces and dipole moments of chemical systems. PhysNet achieves state-of-the-art performance on the QM9, MD17 and ISO17 benchmarks. Further, two new datasets are generated in order to probe the performance of ML models for describing chemical reactions, long-range interactions, and condensed phase systems. It is shown that explicitly including electrostatics in energy predictions is crucial for a qualitatively correct description of the asymptotic regions of a potential energy surface (PES). PhysNet models trained on a systematically constructed set of small peptide fragments (at most eight heavy atoms) are able to generalize to considerably larger proteins like deca-alanine (Ala$_{10}$): The optimized geometry of helical Ala$_{10}$ predicted by PhysNet is virtually identical to ab initio results (RMSD = 0.21 Å). By running unbiased molecular dynamics (MD) simulations of Ala$_{10}$ on the PhysNet-PES in gas phase, it is found that instead of a helical structure, Ala$_{10}$ folds into a wreath-shaped configuration, which is more stable than the helical form by 0.46 kcal mol$^{-1}$ according to the reference ab initio calculations.
1 Introduction
The introduction motivates machine learning as a way to avoid the computational cost of solving the electronic Schrödinger equation and presents PhysNet as a physics-based message-passing neural network for chemical systems. It highlights PhysNet’s benchmark performance, new datasets for challenging interactions, and generalization from small fragments to larger molecules.
- Motivation: Machine learning methods are increasingly used to estimate molecular properties from reference data without explicitly solving the electronic Schrödinger equation.The electronic Schrödinger equation is computationally demanding even when approximate methods are used.
- PhysNet architecture: PhysNet is introduced as a physics-based message-passing high-dimensional neural network designed to learn potential energy surfaces.Message-passing networks learn chemical-environment representations directly from nuclear charges and Cartesian coordinates.
- Contributions: PhysNet improves upon or matches state-of-the-art performance on the QM9, MD17, and ISO17 benchmark datasets.The introduction also presents two new datasets targeting SN2 reaction potential energy surfaces and protein-like or water-cluster systems.
- Generalization: PhysNet accurately predicts protein sidechain-sidechain and backbone-backbone interaction energies and generalizes from fragments with at most eight heavy atoms to considerably larger molecules.The model is applied to helical deca-alanine geometry optimization using an ensemble of PhysNet models.
- Generalization: 0.21 Å is the RMSD between the PhysNet-optimized helical deca-alanine structure and the structure optimized at the reference DFT level of theory.The two structures are described as almost indistinguishable.
2 Methods
PhysNet combines modular deep-neural-network components that refine atom-centered features using local environments while respecting physical invariances. Its methods also explicitly address long-range electrostatics, charge conservation, computational scaling, and benchmark-data gaps.
- Architecture: PhysNet iteratively refines atom-centered features by coupling each atom’s nuclear-charge representation with neighboring atoms within a cut-off radius.The features encode nuclear charge Z and local atomic environment, and are used to predict atomic contributions to chemical properties.
- Architecture: The architecture uses an embedding layer, repeated modular building blocks, interaction and output blocks, and pre-activation residual blocks.The input nuclear charges are embedded into feature vectors, while residual blocks provide unrestricted gradient flow during training.
- Physical modeling: Energy predictions respect translation, rotation, and permutation invariance, while explicit electrostatics account for interactions beyond the local cut-off.A sufficiently large cut-off would reduce computational efficiency for inverse-distance electrostatic interactions, motivating their explicit treatment.
- Physical modeling: Charge conservation requires correcting predicted atomic partial charges so their sum equals the system’s total charge Q.Neural networks do not guarantee this equality a priori, even when the uncorrected sum is usually close after training.
- Computational considerations: Long-range pairwise interactions make energy evaluation scale quadratically with system size, although Ewald summation and cut-off methods can recover linear scaling.These schemes are identified as applicable without modification.
- Benchmark datasets: Two new benchmark datasets probe chemical reactivity, long-range electrostatics, and many-body intermolecular interactions.The datasets cover SN2 reactions and solvated protein fragments because existing benchmarks did not cover these systems adequately.
3 Results
PhysNet achieves strong performance across QM9, MD17, and ISO17 benchmarks, with ensembles improving predictions further. Explicit long-range electrostatics is important for reaction datasets, while PhysNet-ens5 generalizes to larger protein systems and accurately reproduces reference energies.
- QM9: PhysNet improves upon the state of the art on QM9 energy prediction, and PhysNet-ens5 reduces the error further.PhysNet results were averaged over five independent runs, while PhysNet-ens5 is an ensemble of five models.
- MD17: PhysNet is evaluated on MD17 using energy-and-force prediction, with results averaged over five runs and an ensemble prediction also reported.Model comparisons use different training subsets: some models train on energies only, GDML on forces only, and SchNet and PhysNet are compared under their respective settings.
- ISO17: PhysNet achieves state-of-the-art performance on both energies and forces for known molecules with unknown conformations in ISO17.For unknown molecules with unknown conformations, PhysNet improves upon SchNet only for force predictions and performs slightly worse for energies.
- SN2 reactions: Explicit long-range electrostatics is crucial for the SN2 reaction dataset because ion-dipole interactions extend beyond the 10 Å cutoff.Without long-range augmentation, the model performs significantly worse because ion-dipole interactions decay with the square of distance and affect the overall energy.
- Protein generalization: Ala10 contains 54 heavy atoms, despite training solvated-protein fragments containing at most eight heavy atoms.This tests whether PhysNet predictions generalize from small molecular fragments to a substantially larger protein model system.
- Protein generalization: 0.233 kcal mol−1 atom−1 is the average prediction error for 20 Ala10 trajectory structures, corresponding to a 0.23% relative error.The structures were sampled at 2 ps intervals and evaluated with PhysNet-ens5 against revPBE-D3(BJ)/def2-TZVP reference calculations.
4 Discussion and Conclusion
PhysNet matches or improves state-of-the-art benchmark performance while extending evaluation to chemical reactivity and many-body interactions. Explicit electrostatics improve qualitative potential-energy-surface behavior, although energy errors can differ substantially between Ala10 conformations.
- Benchmark performance: PhysNet matches or improves state-of-the-art performance across all tested quantum-chemical benchmarks, reducing previously published errors by 50–90% in some cases.The passage also notes that kernel-based methods can achieve similar or sometimes better performance on benchmarks such as MD17.
- Datasets and physical knowledge: Two new datasets target chemical reactivity and many-body intermolecular interactions relevant to condensed-phase systems.These datasets address chemical situations not covered by other published datasets.
- Ala10 conformations: The optimized helical Ala10 geometries from PhysNet-PES and ab initio calculations are almost identical, but PhysNet-ens5 has an energy error about an order of magnitude larger for the helix than for the wreath.The passage suggests the helix’s large dipole moment, arising from cumulative effects, as a possible explanation.
- Datasets and physical knowledge: Explicit electrostatic contributions significantly improve the qualitative shape of PhysNet’s SN2 potential-energy surfaces near and beyond the cutoff radius.This result concerns methyl-halide reactions with halide anions.
- Overall conclusion: PhysNet accurately predicts energies and forces across diverse structures, chemical and conformational degrees of freedom, and datasets.The conclusion presents this as a broad capability of the architecture.