Source-linked AI summary

Deep Potential: a general representation of a many-body potential energy surface

Jiequn Han, Linfeng Zhang, Roberto Car, Weinan E

arXiv:1707.01478v2physics.comp-ph

TL;DR

Molecular modeling faces a major accuracy–computational-cost dilemma. The paper develops Deep Potential, an end-to-end neural-network representation of many-atom potential energy surfaces, and reports low errors in comparisons with EAM calculations.

  • Problem

    Molecular science faces a major dilemma between accuracy and computational cost.

  • Method

    Deep Potential is a simple, general, end-to-end deep neural-network representation that uses polar coordinates and imposed rotational and permutational symmetries.

  • Results

    RMSE per atom in comparisons with EAM calculations was 0.10, 0.07, and 0.06 meV/atom, while the representation targets many-atom potential energy surfaces with quantum-mechanics-model accuracy.

  • Takeaways & Limitations

    Deep Potential provides a general representation of many-atom potential energy surfaces for addressing the accuracy–cost challenge in molecular modeling.

  • Takeaways & Limitations

    The current implementation does not explicitly include effects treated with techniques such as Ewald summation.

Abstract

from arXiv · show

We present a simple, yet general, end-to-end deep neural network representation of the potential energy surface for atomic and molecular systems. This methodology, which we call Deep Potential, is "first-principle" based, in the sense that no ad hoc approximations or empirical fitting functions are required. The neural network structure naturally respects the underlying symmetries of the systems. When tested on a wide variety of examples, Deep Potential is able to reproduce the original model, whether empirical or quantum mechanics based, within chemical accuracy. The computational cost of this new model is not substantially larger than that of empirical force fields. In addition, the method has promising scalability properties. This brings us one step closer to being able to carry out molecular simulations with accuracy comparable to that of quantum mechanics models and computational cost comparable to that of empirical potentials.

1 Introduction

Deep Potential addresses the accuracy–cost gap between efficient empirical potentials and accurate but expensive first-principles methods. It uses an end-to-end neural representation that directly processes atomic coordinates while respecting system symmetries.

  • Empirical potentials enable large simulations but have limited accuracy and transferability because their construction relies on fitted, often ad hoc functions.
  • First-principles methods such as DFT and AIMD are more accurate but computationally expensive, restricting simulations to systems of hundreds to thousands of atoms.
  • Deep Potential is introduced as a simple, general, end-to-end deep neural network representation of many-atom potential energy surfaces.
  • The network uses raw atomic coordinates in a proper reference frame and naturally respects the system’s symmetries.
  • Across small molecular isomers and condensed-phase systems, Deep Potential produces promising results and approaches first-principles accuracy at empirical-potential-like computational cost.

2 Results

Deep Potential constructs a symmetry-preserving potential-energy representation from local atomic environments and evaluates it through nearly independent atomic sub-networks. Tests across molecular and condensed-phase systems support finite local cutoffs and accurate energy prediction.

  • Deep Potential framework: The framework represents the potential energy surface directly from atomic configurations while imposing rotational and permutational symmetries through fixed operations.
  • Deep Potential framework: For N atoms, the network uses N nearly independent sub-networks, each corresponding to one atom and contributing a local energy term.
  • Deep Potential framework: Finite cutoffs are sufficient for all extended material systems considered, while sub-networks are coupled only through summation when computing total energy.
  • Deep Potential framework: Each local environment is transformed into an atom-centered Cartesian frame, and atoms inside a finite cutoff radius provide the corresponding sub-network input.
  • Deep Potential framework: The representation (1/r, cos θ, cos φ, sin φ) outperforms Cartesian coordinates (x, y, z) in the tested formulation.
  • Test systems: The method is evaluated on molecular isomers and condensed-phase datasets using MAE and RMSE energy metrics, respectively.

3 Discussion

Deep Potential is presented as a general framework that reproduces empirical and quantum-mechanical reference data across finite and extended systems with chemical accuracy, while retaining favorable computational cost and scalability. Its accuracy, symmetry handling, local-environment design, and tunable network size support broad application, though explicit long-range effects remain future work.

  • Accuracy: Deep Potential reproduces original data within chemical accuracy across all test cases, including empirical-potential and DFT-based systems.Chemical accuracy is approximately 1 kcal/mol or 0.04 eV.
  • Accuracy: 0.09 meV/atom is obtained for Zr with only radial information, compared with 0.21 meV/atom for BPNN.The comparison uses symmetry functions for Zr that were not redesigned specifically for that element.
  • Accuracy: 1.8 meV/H2O is obtained for liquid water with radial and first-shell angular information, one fourth of the distance-only error.Including angular information improves the reported liquid-water prediction relative to distance information alone.
  • Symmetry: Removing permutational symmetry fixing in bulk Cu substantially increases RMSE from 0.2 meV/atom, while removing rotational fixing in C7O2H10 changes MAE from 0.04 to 0.05 eV.These tests assess the contribution of the imposed rotational and permutational symmetry constraints.
  • Computational cost and scalability: Deep Potential has computational cost linear with atom count N, is highly parallelizable, and makes large-scale molecular dynamics simulations computationally feasible.The method requires more size-independent operations than EAM, but both approaches are highly parallelizable.
  • Computational cost and scalability: Reducing the C7O2H10 sub-network yields MAE 0.07 eV instead of 0.04 eV while reducing computational cost by almost two orders of magnitude.The smaller network uses hidden-layer sizes 80-40-20-10 rather than the considerably larger reported network.
  • Computational cost and scalability: Finite local environments support extensivity, but explicit long-range effects are not included in the current implementation and are left for future studies.Long-range Coulombic effects may be included in training data when important, while explicit treatment such as Ewald summation is not implemented here.

4 Methods

The study constructs reference datasets from empirical and first-principles molecular-dynamics trajectories, then trains Deep Potential neural networks using standard deep-learning procedures. Training choices include system-dependent normalization, optimization, batch normalization, learning-rate schedules, and cutoff radii.

  • Reference data sets: 90% of C7O2H10 snapshots are used for training and 10% for testing energy predictions.Each trajectory runs for 1 ps with a variable timestep limiting the largest atomic displacement per step to less than 0.1 Å.
  • Reference data sets: 90,000 of 100,000 randomly selected water snapshots are used for training and 10,000 for testing energy predictions.The path-integral AIMD trajectory is approximately 10 ps long with a timestep of 0.48 fs and uses eight beads.
  • Training procedure: The implementation generates and renormalizes atomic-configuration inputs, then trains fully connected ReLU networks with Adam and batch size 128.Table 2 specifies subnetworks, training epochs, learning-rate schemes, and batch-normalization decay parameters.
  • Training procedure: Batch normalization is used for C7O2H10 and H2O but omitted for Cu and Zr after testing its effectiveness.The adopted cutoff radii are 6.0 Å, 7.0 Å, and 5.8 Å for Cu, Zr, and H2O, respectively.
Loading 1707.01478v2…