Source-linked AI summary
Interatomic potentials for ionic systems with density functional accuracy based on charge densities obtained by a neural network
S. Alireza Ghasemi, Albert Hofstetter, Santanu Saha, Stefan Goedecker
TL;DR
Standard machine-learning interatomic potentials target total energy directly, limiting accurate treatment of long-range charge transfer and ionized systems. This paper instead predicts charge density from local atomic environments, enabling charge redistribution before evaluating energy, and reports density-functional accuracy for neutral and ionized NaCl clusters.
Problem
Directly learning total energy limits accurate interatomic potentials for systems where long-range charge transfer matters, including ionized systems.
Method
A neural network predicts environment-dependent charge density from short-range atomic environments, which is then used to determine total energy.
Results
The approach achieves density-functional accuracy for neutral and ionized NaCl clusters, with a neutral-test RMSE of 0.13 mHa per atom.
Takeaways & Limitations
Allowing charge to redistribute across the whole cluster supports accurate energetics for both neutral and ionized systems.
Takeaways & Limitations
The energy expression is missing kinetic and exchange contributions.
Abstract
from arXiv · showhide
Based on an analysis of the short range chemical environment of each atom in a system, standard machine learning based approaches to the construction of interatomic potentials aim at determining directly the central quantity which is the total energy. This prevents for instance an accurate description of the energetics of systems where long range charge transfer is important as well as of ionized systems. We propose therefore not to target directly with machine learning methods the total energy but an intermediate physical quantity namely the charge density, which then in turn allows to determine the total energy. By allowing the electronic charge to distribute itself in an optimal way over the system, we can describe not only neutral but also ionized systems with unprecedented accuracy. We demonstrate the power of our approach for both neutral and ionized NaCl clusters where charge redistribution plays a decisive role for the energetics. We are able to obtain chemical accuracy, i.e. errors of less than a milli Hartree per atom compared to the reference density functional results. The introduction of physically motivated quantities which are determined by the short range atomic environment via a neural network leads also to an increased stability of the machine learning process and transferability of the potential.
1 Institute for Advanced Studies in Basic Sciences,
The section identifies an institute in Zanjan, Iran.
- The listed affiliation is the Institute for Advanced Studies in Basic Sciences, Zanjan, Iran.
- The affiliation is associated with the address P.O. Box 45195-1159.
- The listed location is Zanjan, Iran.
2 Department of Physics, Universit¨at Basel,
The paper addresses limitations of short-range, total-energy machine-learning potentials for charge-transfer and ionized systems by predicting charge density instead. Applied to NaCl clusters, the method achieves density-functional accuracy, charge conservation, transferability, and physically reasonable low-energy structures.
- Motivation: Standard machine-learning potentials based on short-range atomic environments struggle with long-range charge transfer and ionization.Fixed-charge force fields likewise cannot accurately represent charge transfer between surface and core atoms in NaCl clusters.
- Method: The method predicts environment-dependent electronegativities and uses them to determine an approximate superposition of atomic charge densities.The charge density is optimized through a variational principle and determines the relevant energy differences.
- Method: The approach permits long-distance charge transfer while conserving the system’s total charge and treating ionized systems without reparametrization.
- Limitation: The method’s energy expression omits kinetic and exchange-correlation terms, although physically important energy differences can still be calculated through the Hellmann-Feynman theorem.
- Results: 0.13 mHa per atom is the neutral-test-set RMSE, while the ionized-test-set RMSE is 0.44 mHa per atom for qtot = +1 and qtot = +2.The training and validation RMSE is 0.26 mHa per atom.
- Results: Compression and expansion configurations absent from training and validation demonstrate transferability of the potential.
- Results: The potential describes the entire low-energy configurational space without producing physically unreasonable structures.Minima-hopping runs were used to probe whether the potential failed in parts of this space.