Efficient, Equivariant Predictions of Distributed Charge Models

A machine learning (ML) based equivariant neural network for constructing distributed charge models (DCMs) of arbitrary resolution, DCM-net, is presented. DCMs efficiently and accurately model the anisotropy of the molecular electrostatic potential (ESP) and go beyond the point charge representation used in conventional molecular mechanics (MM) energy functions. This is particularly relevant for capturing the conformational dependence of the ESP (internal polarization) and chemically relevant features such as lone pairs or σ-holes. Across conformational space, the learned charge positions from DCM-net are stable and continuous. Across the QM9 chemical space, two-charge-per-atom models achieve accuracies comparable to fitted atomic dipoles for previously unseen molecules (0.75 (kcal/mol)/e). Three- and four-charge-per-atom models reach accuracies competitive with atomistic multipole expansions up to quadrupole level (0.55 (kcal/mol)/e). Pronounced improvements of the ESP are found around O and F atoms, both of which are known to feature strongly anisotropic fields, and for aromatic systems. Across the QM9 reference data set, molecular dipole moments improve by 0.1 D compared with fitted monopoles. Transfer learning on dipeptides yields a 0.2 (kcal/mol)/e ESP improvement for unseen samples and a two-fold MAE reduction for molecular dipole moments versus fitted monopoles. Overall, DCM-net offers a fast and physically meaningful approach to generating distributed charge models for running pure ML or mixed ML/MM based molecular simulations. level (0.55 (kcal/mol)/e).

💡 Research Summary

This paper introduces DCM‑net, an SO(3)-equivariant graph neural network designed to generate distributed charge models (DCMs) that accurately reproduce the anisotropic molecular electrostatic potential (ESP). Traditional force fields rely on atom‑centered point charges or atomic multipoles, which cannot capture direction‑dependent features such as lone‑pair regions or σ‑holes. DCM‑net learns both the magnitudes and off‑center positions of a small set of charges per atom, preserving rotational equivariance and charge conservation.

The architecture is built on the e3nn library in JAX. Atomic numbers are embedded, and inter‑atomic vectors are expanded using radial Bernstein polynomials and real spherical harmonics. Message‑passing proceeds for several iterations; each step updates scalar (ℓ = 0), vector (ℓ = 1), and higher‑order tensor (ℓ = 2) features via learnable tensor‑product layers that employ Clebsch–Gordan coefficients. This guarantees that any vector‑ or tensor‑valued output rotates consistently with the molecular frame. After the final message‑passing block, two heads decode the features: a scalar head predicts n_DC charges per atom, while a vector head predicts displacement vectors that locate the off‑center charges relative to the atomic nucleus. A hard‑tanh scaling (0.175 Å) limits displacements to ≈0.3 Å, ensuring physically reasonable charge clouds. Charge conservation is enforced post‑hoc by adjusting the predicted charges so that their sum equals the molecular total charge; this constraint is added as a penalty term during training rather than a hard equality, which would impede convergence.

Training minimizes a loss comprising the root‑mean‑square error of the ESP evaluated on a grid (RMSE ESP) and the charge‑conservation penalty. The model is trained on the QM9 dataset (≈130 k molecules) and subsequently fine‑tuned on a small set of 20 dipeptides (≈400 atoms) to test transferability. Results show that a 2‑charge‑per‑atom DCM achieves ESP errors of ~0.75 (kcal mol⁻¹)/e, comparable to fitted atomic dipoles. Adding a third or fourth charge per atom reduces the error to ~0.55 (kcal mol⁻¹)/e, matching atomistic multipole expansions up to quadrupole order. Notably, the ESP around highly electronegative O and F atoms and aromatic π‑systems improves markedly, reflecting the model’s ability to capture localized anisotropy. Molecular dipole moments are also better reproduced, with an average improvement of ~0.1 D over fitted monopoles.

Transfer learning on dipeptides yields an additional 0.2 (kcal mol⁻¹)/e reduction in ESP error for unseen samples and a two‑fold decrease in dipole MAE relative to monopole‑only models. Computationally, DCM‑net inference costs only a few tens of microseconds per molecule, making it suitable for on‑the‑fly charge assignment in molecular dynamics or hybrid ML/MM simulations. The learned charge positions vary smoothly with geometry, avoiding discontinuities that could destabilize dynamics.

In summary, the authors contribute (1) a rigorously equivariant neural network that jointly predicts charge magnitudes and off‑center positions, (2) evidence that a minimal set of distributed charges can faithfully reproduce ESP and dipole moments, (3) a practical scheme for enforcing charge conservation without sacrificing training stability, and (4) demonstration of chemical‑space transferability via fine‑tuning. DCM‑net thus offers a fast, physically interpretable, and extensible route to generate distributed charge models for a wide range of simulation contexts, bridging pure machine‑learning potentials and traditional force‑field approaches.