A Transferable Machine-Learning Model of the Electron Density

Andrea Grisafi; David M. Wilkins; Benjamin A. R. Meyer; Alberto Fabrizio; Clemence Corminboeuf; Michele Ceriotti

Source-linked AI summary

A Transferable Machine-Learning Model of the Electron Density

Andrea Grisafi, David M. Wilkins, Benjamin A. R. Meyer, Alberto Fabrizio, Clemence Corminboeuf, Michele Ceriotti

arXiv:1809.05349v1Sep 14, 2018physics.chem-ph

TL;DR

Computing electron density is demanding despite its central physical and experimental role. The paper develops an atom-centered, symmetry-adapted model that learns valence density from molecular environments and demonstrates transfer from C4 training molecules to C8 targets. The framework reconstructs density with errors around 1% and predicts octane and octatetraene densities with reported errors of 1.4% and 1.4%, respectively.

Problem
Electron-density computation requires demanding electronic-structure calculations, while prior density-learning models had limited transferability to large and flexible systems.
Method
The framework expands total valence density in optimized atom-centered basis functions and learns its tensorial components with symmetry-adapted regression using a density-based loss.
Results
The model achieves density errors of about 1% for C2 and C4 hydrocarbons and predicts C8 densities after training on corresponding C4 compounds, including 1.4% error for octane.
Takeaways & Limitations
The demonstrated local decomposition provides transferable density predictions for larger molecules from smaller-molecule training data at a cost proportional to the number of environments.
Takeaways & Limitations
Transferability declines when target environments differ substantially from training environments beyond the optimal 3 Å cutoff, although larger training molecules could relax this locality constraint.

Abstract

from arXiv · show

The electronic charge density plays a central role in determining the behavior of matter at the atomic scale, but its computational evaluation requires demanding electronic-structure calculations. We introduce an atom-centered, symmetry-adapted framework to machine-learn the valence charge density based on a small number of reference calculations. The model is highly transferable, meaning it can be trained on electronic-structure data of small molecules and used to predict the charge density of larger compounds with low, linear-scaling cost. Applications are shown for various hydrocarbon molecules of increasing complexity and flexibility, and demonstrate the accuracy of the model when predicting the density on octane and octatetraene after training exclusively on butane and butadiene. This transferable, data-driven model can be used to interpret experiments, initialize electronic structure calculations, and compute electrostatic interactions in molecules and condensed-phase systems.

INTRODUCTION

Electron density is physically fundamental and experimentally informative, but computing it for large systems is demanding. The paper introduces a local, symmetry-adapted machine-learning model designed to overcome limited transferability to large and flexible molecules.

Electron density can be measured directly, analyzed for covalent and non-covalent patterns, and used in principle to access any ground-state property.
Computing electron density requires solving an electronic-structure problem, especially for large systems.
Earlier density-learning work used smoothed nuclear potentials and planewave coefficients but was limited to interpolation among conformers of relatively rigid, small molecules.
The proposed model uses a local basis for valence electron density and symmetry-adapted regression to predict densities for large and flexible systems from smaller compounds.
Training on butadiene and butane enables prediction of electron density for octatetraene and octane by extrapolating information from corresponding local C4 environments.

THEORY AND METHODS

The method represents total electron density with atom-centered basis functions and learns tensorial coefficients using symmetry-adapted regression. It avoids arbitrary atomic charge targets by jointly fitting the density while accounting for basis overlap and regularization.

The density is expanded as a sum of atom-centered basis functions whose coefficients are learned from nuclear coordinates.
Each basis function combines radial functions and spherical harmonics, while coefficients are grouped into spherical multipoles that transform covariantly under rigid rotations.
Symmetry-adapted Gaussian process regression uses rank-2 kernel matrices to encode structural similarity and geometric relationships between target and reference environments.
The model predicts coefficients independently by radial channel, angular momentum, and species, while the non-orthogonal basis couples density components across atoms.
Directly inverting the overlap matrix produced noisy, unstable coefficients because the matrix can be ill-conditioned.
A joint regression and basis decomposition instead selects among nearly equivalent density decompositions by minimizing discrepancy between reference and model densities with regularization.
The practical solution uses regression weights, density projections, symmetry-adapted kernels, and sparse overlap matrices in a non-trivial linear-equation system.

Benchmark Dataset

The benchmark combines 1000 diverse hydrocarbon configurations with DFT/PBE electron-density calculations and an optimized atom-centered basis representation.

The dataset contains 1000 independent ethene, ethane, butadiene, and butane structures generated using replica exchange molecular dynamics.Configurations were extracted from a 300 K replica using farthest-point selection.
Electron densities for the selected configurations were computed at the DFT/PBE level using SBKJC effective core potentials.
The basis uses Gaussian-type orbitals, contracting 12 regularly spaced radial functions into 4 optimized functions.Angular momentum channels through f functions are needed for approximately 1% density decomposition error.
Figure 1 compares density errors from isolated-atomic-density superposition with those from the optimized basis for C2 and C4 molecules.The figure uses red and blue isosurfaces at ±0.005 Bohr−3.

Charge decomposition analysis

The optimized atom-centered expansion resolves charge-density structure through angular-momentum channels, with distinct channels describing isotropic density, bonds, conjugated π-clouds, and anisotropy.

Charge decomposition analysis: The optimized basis decomposition reaches about 1% density error, compared with 16–20% for superposed rigid atomic densities in butane.The decomposition is obtained from density projections and the overlap matrix through c = S−1w.
Charge decomposition analysis: Figure 2 shows angular-momentum-resolved density representations and spectra for C2 and C4 datasets, with statistically zero means for l > 0.The l = 0 contribution represents collective variation relative to the dataset mean.
Charge decomposition analysis: The s channel determines the general density shape, while p functions describe electronegativity gradients near C–H bonds.
Charge decomposition analysis: The d channel captures C–C bond modulation and conjugated π-clouds, while f functions add further non-trivial anisotropy.
Charge decomposition analysis: After subtracting mean pure-s atomic density, hydrogen variability is dominated by l = 1, while d contributions are substantial for alkene carbon atoms.For alkane carbon atoms, f functions are dominant; for hydrogen, they are nearly irrelevant.
Charge decomposition analysis: Density expansion requires higher angular momentum than wavefunction expansion because squaring the wavefunction can generate components up to twice its maximum l.

Density learning with SA-GPR

SA-GPR learns charge densities more easily for rigid molecules than flexible ones, reaching the basis-set limit for density errors but facing larger errors when predicting exchange-correlation energies indirectly.

Density learning with SA-GPR: Butadiene and butane require more than 100 training structures to approach the basis-set limit because their conformational flexibility makes learning more challenging.
Density learning with SA-GPR: Ethene and ethane density errors saturate near the 1% basis-set limit with as few as 10 training points.
Density learning with SA-GPR: The learned density reaches an accuracy nearly 20 times better than superposed rigid atomic densities for applications relying on density representation.Examples include X-ray and transmission electron microscopy modeling and density-based chemical-interaction fingerprints.
Density learning with SA-GPR: With 800 training molecules, exchange-correlation-energy RMSEs are 0.9 kcal/mol for ethene, 1.7 for ethane, 1.9 for butadiene, and 3.5 for butane.These values essentially match the basis-set limit.
Density learning with SA-GPR: Reducing basis-set error is necessary to reach roughly 1 kcal/mol chemical accuracy for exchange-correlation energies.
Density learning with SA-GPR: Direct scalar SOAP regression of exchange-correlation energies substantially outperforms indirect prediction through the learned density while using a simpler model.

Size-extensive extrapolation

The symmetry-adapted local model transfers electron-density predictions from C4 hydrocarbons to larger C8 molecules by learning local environments rather than whole-molecule identities. Training on butadiene and butane yields low errors on challenging octatetraene and octane datasets, while transferability depends on locality.

Size-extensive extrapolation: The model uses representative local environments that need not come from the molecule being predicted, provided they describe similar chemical environments.Predictions for arbitrarily large molecules are obtained by computing kernels between target and reference environments.
Size-extensive extrapolation: Butadiene and butane training configurations are used to predict the valence electron densities of octatetraene and octane.The larger C8 molecules are predicted by extrapolating information learned from corresponding C4 compounds.
Size-extensive extrapolation: 3 Å is the optimal cutoff for extrapolation, compared with 4.5 Å for same-molecule predictions, because larger environments differ between C8 targets and C4 training compounds.Extending the training set to larger molecules could relax this locality constraint.
Size-extensive extrapolation: 1.8% low-density mean absolute error is obtained for octatetraene, versus 1.4% for octane, when learning on the full butadiene and butane datasets.These results use 100 diverse structures selected from the 300 K replica of a long REMD run.
Size-extensive extrapolation: The prediction accurately reproduces the electron-density structure for both octane and octatetraene, despite greater difficulty in octatetraene’s π-cloud middle regions.Those regions lack analogous examples in the butadiene training data.

CONCLUSIONS

The framework addresses transferable charge-density learning by combining optimized atom-centered basis functions, symmetry-adapted regression, and a loss based on total density. It achieves approximately 1% errors for smaller hydrocarbons and predicts larger C8 counterparts from C4 training, while further accuracy improvements remain possible.

CONCLUSIONS: The framework decomposes density into optimized atom-centered basis functions, uses symmetry-adapted regression for geometric covariances, and constrains learning through the total charge density.The atom-centered decomposition supports linear cost with the number of atoms for large molecules or condensed phases.
CONCLUSIONS: Approximately 1% reconstructed-density error is achieved for saturated and unsaturated hydrocarbons with 2 and 4 carbon atoms.The reported accuracy is sufficient for an order-of-magnitude improvement over methods based on superimposed atomic densities in analyses of X-ray and transmission electron microscopy experiments.
CONCLUSIONS: Models trained on C4 compounds predict the electronic charge of their larger C8 counterparts, demonstrating transferability from a symmetry-adapted local decomposition.This is presented as a first example of the scheme’s transferability.
CONCLUSIONS: Further accuracy improvements are likely through basis-set optimization, kernel and density-representation tuning, charge equilibration, and semi-empirical baselines.The work is characterized as a first successful transferable application to molecular properties with richer geometric structure than additive atom-centered values.
CONCLUSIONS: The framework may benefit other properties with complex geometric structure, including Hamiltonians, density matrices, vector fields, and density-response functions.These applications require careful consideration of both structural input and target-property representations.