Source-linked AI summary
Completing density functional theory by machine-learning hidden messages from molecules
Ryo Nagai, Ryosuke Akashi, Osamu Sugino
TL;DR
The paper tackles the difficulty of constructing explicit and transferable exchange-correlation functionals without relying on complicated theoretical assumptions. It learns a flexible neural-network functional from accurate molecular energies and densities, then solves KS-DFT using backpropagated functional derivatives. Trained on only a few molecules, the resulting functionals apply to hundreds of systems with accuracy comparable to standard functionals, while non-local descriptors provide a direct extension.
Problem
The explicit exchange-correlation potential is difficult to derive, while existing machine-learning approaches using whole densities require fixed system sizes and theoretical functionals can trade energy accuracy for density accuracy.
Method
The study trains a flexible feed-forward neural network to map density descriptors to exchange-correlation energy density and obtains its functional derivative by backpropagation within KS-DFT.
Results
Trained on H2O, NH3, and spin-polarized NO, the neural-network functionals are applicable to hundreds of molecular systems with accuracy comparable to or better than existing functionals.
Takeaways & Limitations
Adding descriptors such as the near-region averaged density offers a systematic route to transferable non-local exchange-correlation functionals within the KS-DFT framework.
Takeaways & Limitations
The method was tested using only three reference molecules, and performance can depend on initial densities being sufficiently close to self-consistent solutions.
Abstract
from arXiv · showhide
Kohn-Sham density functional theory is the base of modern computational approaches to electronic structures. Their accuracy vitally relies on the exchange-correlation energy functional, which encapsulates electron-electron interaction beyond the classical one. The functional provides a way to obtain the density and energy without solving the many-body equation and can, in principle, be determined to reproduce the exact ones universally. However, the past approaches are dependent on the theoretical development, which limits the possibility of the functional to human's intuition. Here, we demonstrate a systematic way to machine-learn a functional from a database, without complicated assumptions. The density and energy are related with a flexible feed-forward neural network, which is trained to reproduce accurate dataset, and the KS-DFT is solved by taking the functional derivatives with the back-propagation technique. Surprisingly, a trial functional, trained for just a few molecules, has been shown to be applicable to hundreds of molecular systems with comparable accuracy to the standard functionals. Also, by adding the nodes connected to the hidden layers, a non-local term is straightforwardly included to improve accuracy, which has been hitherto impractically difficult. Utilizing the strategy of rapidly advancing machine learning techniques, this novel approach is expected to enrich the DFT framework by constructing a functional just from a database for materials conventionally difficult to calculate accurately.
Introduction
The paper addresses the difficulty of deriving explicit, transferable exchange-correlation functionals by learning them with flexible neural networks from accurate molecular data. The approach represents local or non-local density information, trains against energies and densities, and differentiates the network within self-consistent KS-DFT.
- The exchange-correlation potential is ideally non-local, but its explicit form remains elusive, motivating machine-learning construction.
- Previous machine-learning approaches using whole density distributions require fixed target-system sizes, limiting application to real materials of varying size.
- The study formulates exchange-correlation energy density with a feed-forward neural network mapping density descriptors to εxc(r).The descriptors can include local quantities such as density, spin polarization, gradient, and kinetic-energy density.
- Backpropagation evaluates the functional derivative δExc/δn(r), allowing the neural-network functional to enter the KS equation through Vxc(r).
- A near region approximation adds an averaged density around r as a non-local descriptor, addressing the difficulty of constructing non-local functionals.The averaged density is incorporated into the descriptor vector alongside meta-GGA variables.
- The network is trained on atomization energies and density distributions of H2O, NH3, and spin-polarized NO using accurate quantum-chemical reference data.Training repeatedly performs KS-DFT calculations and updates network parameters using errors in both quantities.
Results and Discussions
The trained neural-network functionals perform comparably to or better than representative analytic functionals across hundreds of molecular systems and several properties. Adding descriptors systematically improves the representation, while self-consistent KS-DFT supports smooth dissociation behavior and practical non-local extensions.
- Across hundreds of systems containing first- to third-row elements, neural-network functionals are superior or comparable in accuracy to existing functionals at each approximation level.The comparison includes referenced and unreferenced properties, including bond heights and total energies.
- For atomization and total energies, the non-local NRA functional is comparable to hybrid functionals fitted to more than 100 systems.
- The functionals transfer to broader materials and unreferenced quantities despite training on a small molecular dataset.
- As descriptors expand from local to non-local forms, the mapping from descriptors to εxc becomes less multivalued and accuracy tends to improve.The reported sequence is LSDA, GGA, meta-GGA, and NRA.
- The neural-network meta-GGA functional produces bond-dissociation curves for C2H2 and N2, with self-consistent KS-DFT helping suppress unphysical density noise from overfitting.
- The learned enhancement factor resembles representative functionals except near s → +0 and τ → τW + 0, where training data do not cover the limits.The authors state that these limits had little effect on the present test systems but should be constrained for broader applicability.
- The study concludes that flexible neural-network functionals can introduce non-locality with ordinary KS-DFT computational cost.
Methods
The method represents the exchange-correlation energy with a neural network over local or non-local density descriptors, then obtains the exchange-correlation potential through functional differentiation. Training iteratively updates network parameters using KS-DFT errors for atomization energies and density distributions, while numerical non-local integration and practical safeguards support application.
- NN-based functional: The exchange-correlation energy is expressed as an integral of density times a neural-network energy density dependent on descriptor vector g[n](r).The network correction is built with four fully connected layers, and its final layer constrains εxc to be nonpositive.
- Non-local extension: The descriptor vector can combine semi-local variables with a weighted density distribution around each point to represent non-locality.The weighted contribution decays with distance using d(r,r′)=exp(−|r−r′|/σ), with σ fixed to 0.2 bohr; evaluating the potential scales quadratically with system size.
- KS-DFT coupling: Functional derivatives of the neural network are evaluated by backpropagation, allowing the resulting exchange-correlation potential to enter the KS equation.The numerical non-local integrations use the same grid points as exchange-correlation integration.
- Training: Training uses Monte Carlo perturbations of network weights, KS-DFT evaluations, and acceptance or rejection based on a cost function combining energy and density errors.The procedure gradually reduces perturbation size and temperature until the error is sufficiently small.
- Training: The cost function measures atomization-energy deviations from G2 calculations and density errors relative to CCSD densities, with fixed balancing parameters for training.The study sets E0 to 10 hartree and c2/c1 to 1, while density integrations use the KS exchange-correlation grid.
- Computational details: The implementation uses PySCF with the 6-311++G(3df,3pd) basis set, PyTorch for the neural network, and an atomic-density superposition as the initial KS density.The trained parameters are made available with PySCF usage instructions.