Source-linked AI summary
Deep Potential Molecular Dynamics: a scalable model with the accuracy of quantum mechanics
Linfeng Zhang, Jiequn Han, Han Wang, Roberto Car, Weinan E
TL;DR
The paper develops a neural-network scheme for molecular simulation trained on ab initio data. DeePMD jointly handles energies, forces, and virials, with the training procedure designed to improve efficiency while keeping these quantities consistent.
Problem
The paper addresses the challenge of making neural-network molecular simulations efficient while training energies, forces, and virials consistently.
Method
DeePMD uses a deep neural-network potential trained from AIMD data to model energies, forces, and virials.
Results
The training scheme makes training significantly more efficient and keeps forces and the virial mutually consistent.
Takeaways & Limitations
Batch Normalization is not necessary within the DeePMD framework.
Abstract
from arXiv · showhide
We introduce a scheme for molecular simulations, the Deep Potential Molecular Dynamics (DeePMD) method, based on a many-body potential and interatomic forces generated by a carefully crafted deep neural network trained with ab initio data. The neural network model preserves all the natural symmetries in the problem. It is "first principle-based" in the sense that there are no ad hoc components aside from the network model. We show that the proposed scheme provides an efficient and accurate protocol in a variety of systems, including bulk materials and molecules. In all these cases, DeePMD gives results that are essentially indistinguishable from the original data, at a cost that scales linearly with system size.
1. Training/Testing Data
The study constructs training and testing datasets from equilibrated AIMD trajectories for liquid water and several ice phases, using randomly selected snapshots and separate remaining data for testing. Simulations use specified time steps, thermostats, variable-cell dynamics, and path-integral bead counts.
- Training and testing data are extracted from AIMD simulation snapshots.
- All simulations use a 0.48 fs time step, while path-integral simulations use eight beads for converged Feynman-path representation.
- Classical AIMD uses a Nosé-Hoover thermostat, and variable-cell dynamics uses the Parrinello-Rahman technique in all cases.
- 95,000 of 105,000 liquid-water snapshots are selected for training.
- The remaining snapshots in each database are reserved for testing.
b. molecules
The molecular dataset covers eight organic molecules, with 95,000 randomly selected snapshots per molecule used for training and the remaining snapshots used for testing.
- The organic-molecule set includes benzene, uracil, naphthalene, aspirin, salicylic acid, malonaldehyde, ethanol, and toluene.
- For each molecule, 95,000 randomly selected database snapshots train the DeePMD model.
- The remaining molecular snapshots are used for testing.
2. Implementation of the method
The implementation combines TensorFlow with custom C++ code to train models and calculate energies, forces, and virials.
- TensorFlow r1.0 is interfaced with C++ codes for data training and property calculation.
- The implementation calculates energy, forces, and the virial.
a. network input data
The network input represents each atom through its cutoff-neighbor environment, standardized descriptors, and a local coordinate frame defined by selected neighboring atoms. Fixed-size inputs accommodate fluctuating neighbor counts, with safeguards for near-zero descriptor variance.
- Each atom’s neighbor list contains atoms within cutoff radius Rc.
- The neighbor list is sorted, and Nc denotes the largest fluctuating neighbor count across snapshots.
- Two selected neighboring axis-atoms define the local frame for each central atom.The general rule chooses the two closest neighbors independently of species.
- For water, the other two atoms in the same molecule are selected, while hydrogen atoms are excluded from axis-atom definitions in organic molecules.
- A rotation matrix transforms neighbor coordinates into the local frame, providing radial and angular information depending on the descriptor choice.
- The fixed-size neural-network input pads unused neighbor nodes with zeros when fewer than Nc neighbors are present.
b. deep neural network for the energy
DeePMD represents each atom’s energy contribution with a species-dependent feedforward neural network, while shared parameters preserve permutation symmetry among atoms of the same species.
- Each atomic contribution to the total energy is computed by a network Nw(i) parameterized by weights that depend on chemical species.
- The network maps descriptors through multiple fully connected hidden layers to the atomic energy Ei.
- Each layer combines a linear transformation with a nonlinear transformation, composed across the feedforward network.
- The hidden-layer activation function used here is the component-wise hyperbolic tangent.
- Atoms of the same chemical species share network parameters, guaranteeing permutational symmetry.
c. forces and virial tensor
DeePMD derives forces and the virial tensor analytically from its local-energy decomposition, using derivatives of the total energy with respect to atomic positions and cell-related quantities.
- The total potential energy is the sum of atomic energies Ei, from which the forces are derived.
- Periodic boundary conditions prevent direct use of absolute coordinates Riα for computing the virial tensor.
- The virial tensor is computed using a formula based on the cell tensor and the local-energy decomposition.
- The virial construction uses vector components between atoms and components of the negative gradient of Ei with respect to xj.
- Position derivatives of the total energy enter both the forces and virial tensor and are evaluated through chain-rule backpropagation.
d. Training Details
Training jointly incorporates energies, forces, and virial tensors with evolving loss prefactors, emphasizing forces early and balancing all quantities later; adding these targets improves regularization and efficiency.
- The network weights are optimized with Adam stochastic gradient descent, using an initial learning rate of 0.001 and exponential decay.
- Each AIMD snapshot supplies 1 energy, 3N force components, and 6 independent virial-tensor elements.
- Because force components are more numerous, matching forces at the beginning makes training efficient.
- Increasing energy and virial prefactors during training produces mutual consistency among energy, forces, and virial.
- Adding forces and/or the virial tensor provides strong network regularization and makes training significantly more efficient.
- Batch Normalization is not necessary within the DeePMD framework.
e. DeePMD details
DeePMD simulations are conducted under conditions matched to corresponding AIMD simulations for water, ice, and organic molecules, including thermodynamic controls, time steps, and trajectory lengths.
- Liquid-water and ice simulations use the same thermodynamic conditions, temperature and pressure controls, and time step as corresponding AIMD simulations.
- All DeePMD trajectories for water and ice are 300 ps long.
- DeePMD and AIMD radial distribution functions are compared for two ice-Ih cases shown in Figures 6 and 7.
- For organic molecules, DeePMD simulations use the same temperature, time step, and trajectory length as corresponding AIMD simulations.
3. Additional Results
The paper extends structural validation beyond liquid water by comparing DeePMD with AIMD for ice and organic-molecule configurations. It also evaluates the oxygen bond-orientation order parameter Q6 and specifies its neighbor-based definition.
- Ice structure: The radial distribution functions of ice Ih are compared between DeePMD and AIMD in Figs. 6–8.The cited passages identify the ice Ih RDF comparisons and their figure locations.
- Orientational order: The probability distribution of the oxygen bond-orientation order parameter Q6 is reported in Fig. 9.Q6 is used to characterize O–O bond orientation order.
- Orientational order: Q6 is defined from spherical harmonics over each oxygen atom’s neighboring oxygens, with a switching function applied to interatomic distances.The neighbor set, spherical harmonic function, and switching function are specified in the definition.
- Orientational order: The Q6 construction includes the central oxygen in the augmented neighbor set used for the order-parameter calculation.The augmented set is written as ˜Nb(i) = Nb(i) ∪ {i}.