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
Existing machine-learning techniques do not recognize a chemical system’s intrinsic symmetry, motivating a symmetry-aware representation. EANN extends EAM by replacing its scalar embedded density with a richer representation and neural networks, achieving excellent accuracy with a smaller architecture across challenging systems.
Problem
Existing machine-learning techniques do not recognize the intrinsic symmetry of a chemical system, making a suitable machine-learning representation essential.
Method
EANN is inspired by EAM and replaces its scalar embedded density expression with a neural-network-based model using an improved embedded representation.
Results
EANN potentials achieve excellent prediction accuracy for energies and forces, performing very well across challenging systems with a much smaller neural-network architecture.
Takeaways & Limitations
EANN provides a conceptually simple and efficient machine-learning model for representing complex systems.
Takeaways & Limitations
EAM and modified EAM retain intrinsic approximations, while comparing models is difficult because descriptor-computation procedures differ.
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.
Learning with a Physically Inspired Representation
The supplied passages identify the paper’s authors and their institutional location.
- Yaolong Zhang, Ce Hu, and Bin Jiang are listed as the authors.
- Bin Jiang is marked with an asterisk in the author list.
- The listed location is Hefei, Anhui 230026, China.
Abstract
The paper develops EANN, a physically inspired neural-network framework that extends EAM with a GTO-based density vector. It reports accuracy comparable to established models across molecular and periodic systems, while reducing model size, data needs, and computational cost.
- Representation: EANN preserves rotational, translational, and permutational symmetry through local atomic density descriptors.
- Efficiency: EANN implicitly includes three-body information through higher-angular-momentum orbitals without explicitly summing costly angular descriptors.
- Molecular systems: EANN matches GDML and SchNet errors with 1000 data points and uses a smaller neural-network architecture for energy and force prediction.
- Periodic systems: EANN potentials show energy and force RMSEs comparable to or slightly smaller than DPMD and DeepPot-SE, with one or two orders of magnitude fewer weights and biases.
- Results: EANN achieves accuracy comparable to established machine-learning models for large molecular and extended periodic systems, using fewer neural-network parameters and configurations.
Supporting Information
The supporting information compares descriptors and training procedures across relevant machine learning models and illustrates radial distribution functions.
- The supporting information compares descriptors used in relevant machine learning models.
- It also compares the models’ training procedures.
- Radial distribution functions are illustrated alongside these comparisons.
points, respectively. Bold numbers correspond to lowest values.
The section presents prediction-error and computational-time comparisons for machine-learning potentials, including EANN, DPMD, and DeepPot-SE, with results summarized in tables.
- Table 2 reports prediction RMSEs for energies and atomic forces.
- For EANN, DeepPot-SE, and DPMD, training used 15%~20%, 90%, and 90% randomly selected points, respectively.
- Bold numbers correspond to the lowest values.
- Table 3 reports CPU time in seconds per core for computing energies and atomic forces.