Source-linked AI summary
Embedded Atom Neural Network Potentials: Efficient and Accurate Machine Learning with a Physically Inspired Representation
Yaolong Zhang, Ce Hu, Bin Jiang
TL;DR
Because standard machine-learning techniques do not recognize chemical-system symmetry, the paper proposes an EAM-inspired neural-network framework. EANN achieves accuracy comparable to established models with a smaller neural-network architecture.
Problem
Standard machine-learning techniques do not recognize chemical-system symmetry, motivating representations that account for it.
Source: Abstract · p. 3Method
EANN is a neural-network framework inspired by EAM that replaces its scalar embedded density with a density representation.
Source: Abstract · pp. 4, 12Results
EANN achieves the same level of prediction accuracy as established models with a much smaller neural-network architecture.
Source: Abstract · pp. 8, 12Takeaways & Limitations
EANN potentials perform very well for the challenging systems evaluated.
Source: Abstract · p. 9Comparing models is difficult because their descriptor-computation procedures differ.
Source: Abstract · p. 10
Abstract
from arXiv · showhide
We propose a simple, but efficient and accurate machine learning (ML) model for developing high-dimensional potential energy surface. This so-called embedded atom neural network (EANN) approach is inspired by the well-known empirical embedded atom method (EAM) model used in condensed phase. It simply replaces the scalar embedded atom density in EAM with a Gaussian-type orbital based density vector, and represents the complex relationship between the embedded density vector and atomic energy by neural networks. We demonstrate that the EANN approach is equally accurate as several established ML models in representing both big molecular and extended periodic systems, yet with much fewer parameters and configurations. It is highly efficient as it implicitly contains the three-body information without an explicit sum of the conventional costly angular descriptors. With high accuracy and efficiency, EANN potentials can vastly accelerate molecular dynamics and spectroscopic simulations in complex systems at ab initio level.
Abstract
The section introduces EANN, an EAM-inspired neural-network potential that uses orbital-based density descriptors to represent atomic energies and implicitly encode three-body information. It describes the model’s accuracy and efficiency across molecular and periodic systems, including its computational speed and practical parameter choices.
- The approach addresses EAM’s limited accuracy and metallic-system scope, while chemical-potential representations must preserve rotational, translational, and permutational symmetry.Source: Abstract · pp. 3, 5
- EANN replaces the scalar embedded density with Gaussian-orbital density components and an atomistic neural network that maps them to atomic energies.The density components are formed from neighboring-atom orbital contributions.Source: Abstract · pp. 5–6
- EANN overcomes EAM’s density approximations, enabling accurate reproduction of ab initio total energies beyond metallic systems.Source: Abstract · p. 6
- The descriptors use Cartesian coordinates without separating two- and three-body terms, while L>0 orbitals implicitly encode three-body information.Source: Abstract · p. 7
Descriptor representation
- EANN descriptors encode angular information through powers of cos(θ_ijk) while avoiding the costly double-loop sum over atom pairs used by ACSFs.The orbital descriptors are reported to converge quickly and can often be truncated at L=2.Source: Abstract · pp. 7, 8, 10
Molecular systems
- With 1,000 training points, EANN yields comparably small errors to GDML and SchNet; using 15,000–20,000 points, it matches the energy-and-force accuracy of DPMD, DeepPot-SE, and SchNet trained on 50,000.Source: Abstract · p. 8
Extended periodic systems
- For extended periodic systems of up to 432 atoms and five atom types, EANN energy and force RMSEs are comparable to or slightly below DPMD and DeepPot-SE.The paper describes these results as evidence that EANN can handle challenging systems with diverse phases and many atomic components.Source: Abstract · p. 9
- EANN has orders of magnitude fewer NN weights and biases than DPMD and DeepPot-SE; about 15%–20% randomly selected points train its unified model per system, versus ~90% for DPMD and DPMD-SE.The authors say satisfactory prediction errors indicate promising transferability.Source: Abstract · p. 9
Computational efficiency
- EANN with ACSFs is more than an order of magnitude faster than AtNN and yields smaller fitting errors.Descriptor evaluation remains the dominant computational bottleneck for large numbers of neighbors, and direct efficiency comparisons with DPMD and SchNet are difficult.Source: Abstract · pp. 9–10
Future directions
- Residual networks could be combined with EANN density descriptors to incorporate deeper learning into the framework.Source: Abstract · p. 10
Parameter choice
- The paper uses an empirical spacing factor β of approximately 0.2, and tests report that EANN performance is not very sensitive to β values from 0.1 to 0.3.A well-behaved EANN representation can be generated by choosing the maximum L and the number of radial positions for each L.Source: Abstract · pp. 11–12
Abstract
- EANN replaces EAM’s scalar embedded density with a GTO-based electron-density vector and its embedding functions with neural networks.The model is described as conceptually and mathematically simple, accurate, and efficient.Source: Abstract · p. 12
- EANN achieves the same prediction accuracy as established ML models for large molecular and periodic systems, using fewer neural-network parameters and configurations.The passages describe this comparison across both big molecular and periodic systems.Source: Abstract · p. 12
- EANN implicitly contains three-body information without summing costly angular descriptors, making it more efficient than the AtNN method using ACSFs.The orbital parameters can be determined with minor human intervention.Source: Abstract · p. 12
- EANN’s accuracy and efficiency make it suitable for molecular-dynamics simulations of extended systems at ab initio level.The passage presents this as an expected use of the model.Source: Abstract · p. 12
Supporting Information
The supporting information covers comparisons of machine-learning descriptors and training procedures, and illustrates radial distribution functions.
Supporting Information
- The supporting information compares descriptors in relevant machine-learning models and describes training procedures, with an illustration of radial distribution functions.Source: Supporting Information · p. 13
points, respectively. Bold numbers correspond to lowest values.
The section compares machine-learning potentials on extended material systems using energy and force prediction RMSEs. It also reports their training-data fractions and identifies tables covering computational cost and the EANN architecture.
points, respectively. Bold numbers correspond to lowest values.
- Table 2 compares energy and atomic-force prediction RMSEs across machine-learning potentials for extended material systems.The reported units are meV/atom for energies and meV/Å for forces; the supplied passages do not include the table’s numerical results.Source: Table 2 · p. 18, points, respectively. Bold numbers… · p. 18
- EANN training used 15%–20% of randomly selected points, while DeepPot-SE and DPMD each used 90%.The passage specifies these as training-point fractions, respectively.Source: points, respectively. Bold numbers… · p. 18
- Table 3 reports CPU time per core, in seconds, for computing energies and atomic forces.Source: Table 3 · p. 19
- Figure 1 presents a schematic architecture of the embedded atom neural network.Source: Figure 1 · p. 20