[Paper Review] Learning Atomic Multipoles: Prediction of the Electrostatic Potential with Equivariant Graph Neural Networks

Moritz Thürlemann, Lennard Böselt+1 more|arXiv (Cornell University)|Oct 11, 2021

Machine Learning in Materials Science83 references47 citations

TL;DR

This paper introduces an equivariant graph neural network (GNN) that predicts atomic multipoles up to quadrupole order from molecular geometry, enabling accurate electrostatic potential (ESP) prediction without costly quantum mechanical calculations. The model enforces rotational symmetry by design, achieving ESP accuracy close to DFT reference values while reducing computational cost by orders of magnitude compared to QM methods.

ABSTRACT

The accurate description of electrostatic interactions remains a challenging problem for fitted potential-energy functions. The commonly used fixed partial-charge approximation fails to reproduce the electrostatic potential at short range due to its insensitivity to conformational changes and anisotropic effects. At the same time, possibly more accurate machine-learned (ML) potentials struggle with the long-range behaviour due to their inherent locality ansatz. Employing a multipole expansion offers in principle an exact treatment of the electrostatic potential such that the long-range and short-range electrostatic interaction can be treated simultaneously with high accuracy. However, such an expansion requires the calculation of the electron density using computationally expensive quantum-mechanical (QM) methods. Here, we introduce an equivariant graph neural network (GNN) to address this issue. The proposed model predicts atomic multipoles up to the quadrupole, circumventing the need of expensive QM computations. By using an equivariant architecture, the model enforces the correct symmetry by design without relying on local reference frames. The GNN reproduces the electrostatic potential of various systems with high fidelity. Possible use cases for such an approach include the separate treatment of long-range interactions in ML potentials, the analysis of electrostatic potential surfaces, and for the static multipoles in polarizable force fields.

Motivation & Objective

• To address the inaccuracy of fixed-charge approximations in classical force fields, which fail to capture anisotropic and conformation-dependent electrostatic effects.
• To overcome the long-range accuracy limitations of local machine-learned potentials by enabling explicit treatment of higher-order multipoles.
• To replace expensive quantum mechanical calculations of multipoles with a differentiable, symmetry-preserving machine learning model.
• To develop a model that predicts atomic multipoles (monopole to quadrupole) with equivariance to rotations, avoiding reliance on local reference frames.
• To enable practical use of multipoles in polarizable force fields, ESP surface analysis, and long-range correction in ML potentials.

Proposed method

• Uses an E(n)-equivariant GNN architecture based on Satorras et al. to ensure rotational invariance in multipole predictions.
• Constructs graph representations where nodes are atoms and edges encode interatomic vectors and distances up to a cutoff radius.
• Employs message passing with equivariant updates using vector and tensor products of interatomic vectors as features.
• Predicts atomic multipoles (q, μ, θ) as output by combining message-passed representations with learnable readout layers.
• Trains the model end-to-end using loss functions based on differences between predicted and reference MBIS-derived multipoles and ESP values.
• Uses a sparse, local atomic environment representation based on neighbor vectors, enabling generalization to unseen conformations and molecules.

Experimental results

Research questions

• RQ1 Can a machine learning model predict atomic multipoles up to quadrupole order with high accuracy while preserving rotational symmetry?
• RQ2 Does the predicted ESP from ML-predicted multipoles match the accuracy of DFT-derived reference multipoles?
• RQ3 Can the model generalize to unseen molecular conformations and diverse chemical environments without relying on local reference frames?
• RQ4 How does the model's performance compare to fixed-charge approximations and QM reference calculations in terms of ESP fidelity and computational cost?
• RQ5 Can the model be used to improve long-range electrostatics in machine-learned potentials or polarizable force fields?

Key findings

• The model predicts atomic multipoles (up to quadrupole) with a mean absolute error of 0.013 e for monopoles, 0.022 D for dipoles, and 0.018 e·bohr² for quadrupoles compared to DFT reference values.
• The predicted electrostatic potential (ESP) from multipoles matches the DFT reference ESP with a mean absolute error of 0.0082 a.u. on the van der Waals surface.
• Including predicted dipoles and quadrupoles reduces the ESP error by over 50% compared to monopole-only predictions, even when using ML-predicted multipoles instead of QM-derived ones.
• The model achieves inference speeds of 0.026–0.075 µs/atom for edge feature computation and 3.26–8.21 µs/atom for full prediction, enabling real-time use in molecular simulations.
• The equivariant architecture ensures correct rotational behavior without requiring data augmentation or local reference frames, improving generalization and physical consistency.
• The model generalizes well to unseen conformations and molecules, with consistent performance across diverse chemical space, as validated on a test set of 1,013,949 conformations.