Source-linked AI summary
DP-GEN: A concurrent learning platform for the generation of reliable deep learning based potential energy models
Yuzhi Zhang, Haidi Wang, Weijie Chen, Jinzhe Zeng, Linfeng Zhang, Han Wang, Weinan E
TL;DR
Reliable deep learning PES models require representative training data, yet dataset generation has received relatively little attention and is often ad hoc. DP-GEN implements concurrent learning by iterating exploration, labeling, and training while automating workflow management. The platform is designed to generate uniformly accurate PES models with reduced human intervention and computational cost.
Problem
Training datasets for deep learning PES models are often insufficiently representative, while data generation has received less attention than model construction.
Method
DP-GEN implements concurrent learning by generating data during training and iteratively performing exploration, labeling, and model training.
Results
DP-GEN generates uniformly accurate deep learning PES models while minimizing human intervention and computational cost.
Takeaways & Limitations
DP-GEN provides an open-source platform for automating concurrent-learning workflows for deep learning PES generation.
Abstract
from arXiv · showhide
In recent years, promising deep learning based interatomic potential energy surface (PES) models have been proposed that can potentially allow us to perform molecular dynamics simulations for large scale systems with quantum accuracy. However, making these models truly reliable and practically useful is still a very non-trivial task. A key component in this task is the generation of datasets used in model training. In this paper, we introduce the Deep Potential GENerator (DP-GEN), an open-source software platform that implements the recently proposed "on-the-fly" learning procedure [Phys. Rev. Materials 3, 023804] and is capable of generating uniformly accurate deep learning based PES models in a way that minimizes human intervention and the computational cost for data generation and model training. DP-GEN automatically and iteratively performs three steps: exploration, labeling, and training. It supports various popular packages for these three steps: LAMMPS for exploration, Quantum Espresso, VASP, CP2K, etc. for labeling, and DeePMD-kit for training. It also allows automatic job submission and result collection on different types of machines, such as high performance clusters and cloud machines, and is adaptive to different job management tools, including Slurm, PBS, and LSF. As a concrete example, we illustrate the details of the process for generating a general-purpose PES model for Cu using DP-GEN.
I. INTRODUCTION
Deep learning PES models aim to combine the efficiency of empirical force fields with ab initio accuracy, but reliable dataset generation remains underdeveloped and often ad hoc. DP-GEN addresses this gap through concurrent learning that generates data during training and automates the iterative workflow.
- Empirical force fields support larger and longer simulations, but their accuracy and transferability are often problematic.
- Dataset generation has received less attention than model construction, and training datasets are often not sufficiently representative.
- Concurrent learning starts without labeled or unlabeled data and generates data on the fly as training proceeds.
- The iterative process seeks a dataset that is representative enough while remaining as small as possible, unlike sequential learning and active learning.
- The procedure explores configuration space with the current PES, selects high-error snapshots for ab initio labeling, retrains the model, and repeats until convergence.
- DP-GEN is an open-source platform designed to minimize computational cost and human intervention while automating scheduling, task dispatch, and result collection.
II. METHODOLOGY
DP-GEN uses concurrent learning to iteratively explore configurations, label selected structures with reference calculations, and retrain an ensemble of deep potential models. Model-deviation estimates identify configurations requiring labeling while avoiding unnecessary ab initio calculations.
- Workflow: Each DP-GEN iteration comprises exploration, labeling, and training.The workflow repeatedly generates configurations, computes reference labels for selected structures, and retrains models.
- Exploration: The ensemble force-deviation indicator estimates uncertainty, with low variance suggesting adequate training data and high variance signaling configurations for labeling.Force predictions are used because forces are local and sensitive to failures, whereas energy is global and may provide insufficient resolution.
- Exploration: Exploration propagates initial configurations through DP-model simulations using MD, enhanced sampling, MCMC, or genetic algorithms.The sampler can be deterministic or stochastic and is designed to visit configurations of practical interest sufficiently often.
- Exploration: Configurations with σlo ≤ ǫ < σhi are selected for labeling, while configurations below σlo are treated as accurate and those at or above σhi as unreliable.The intermediate trust region supplies candidates for reference calculations; highly deviating structures may be unphysical and are excluded from labeling.
- Labeling and training: Labeling computes ab initio energies and forces for selected configurations, adds them to the training dataset, and retraining produces multiple DP models.Training uses a Deep Potential model whose local-environment construction preserves translation, rotation, and identical-particle permutation invariance.
A. Overview
DP-GEN combines a scheduler with a task dispatcher to coordinate concurrent-learning tasks. The scheduler manages information flow, while the dispatcher submits jobs to computational resources and returns completed results.
- Architecture: DP-GEN has two major modules: a scheduler and a task dispatcher.The scheduler coordinates the concurrent-learning workflow, and the dispatcher communicates with computational resources.
- Scheduler: The scheduler generates exploration, labeling, and training tasks iteratively and exchanges them with the dispatcher.Results collected by the dispatcher are returned to the scheduler so the master process can advance to the next step.
- Task dispatcher: The dispatcher automatically submits tasks, monitors their execution, collects results, and sends those results back to the scheduler.This arrangement separates workflow control from execution on available computational resources.
- Task representation: DP-GEN uses task-specific forward files and shared common files, with completed outputs returned through backward files.Parameter and machine requirements are specified in JSON-format files.
B. DP-GEN scheduler
The DP-GEN scheduler coordinates exploration, labeling, and training while classifying sampled configurations by model deviation. It accumulates valid reference data, retrains model ensembles, and records checkpoints for recovery.
- Exploration: The scheduler prepares exploration tasks with DP models and selects sampled configurations according to their model deviation.LAMMPS interfaced with DeePMD-kit performs DP-based molecular dynamics, and trajectories are stored for analysis.
- Model-deviation screening: Configurations are categorized as accurate below σlo, unreliable at or above σhi, or labeling candidates when σlo ≤ ǫt < σhi.Highly deviating structures may be unphysical, while intermediate-deviation structures are passed to labeling.
- Labeling: Labeling supports VASP, Quantum Espresso, Gaussian, and CP2K to compute first-principles energies, forces, and/or virials.The scheduler adds the resulting labels to the training data and excludes results from calculations that fail expected self-consistent-field convergence.
- Training: DeePMD-kit trains several DP models from accumulated data, varying random seeds while retaining shared network and training settings.The scheduler updates training inputs with newly added data and collects the resulting graph files for subsequent exploration.
- Recovery: The scheduler records process steps in checkpoints that support automatic restart from the latest checkpoint or retrieval of an earlier step.Users can modify parameters and return to a previous step when current results are problematic.
C. Task Dispatcher
The task dispatcher automates the operational work required by DP-GEN at scale. It transfers files, submits and monitors jobs across heterogeneous machines, retries terminated jobs, and avoids repeated computations when restarting.
- Purpose: The dispatcher addresses the tedious and time-consuming manual management of script submission and result collection for enormous task volumes.Its workflow includes upload, submit, check, and download stages.
- Monitoring: Terminated jobs are resubmitted, with up to three retries before user intervention is required for likely problematic settings or unphysical configurations.Finished jobs have their backward files downloaded and passed to the scheduler.
- Execution workflow: It uploads input files, generates adaptive submission scripts, maintains job queues, and collects outputs from finished tasks.The dispatcher checks whether tasks are already submitted before constructing scripts or recovering existing queues.
- Robustness: When restarting the DP-GEN master, the dispatcher recovers existing queues and avoids repeated computations.The paper characterizes DP-GEN as automatic, general, and robust once input files are prepared and execution succeeds.
- File transfer: The dispatcher adapts file transmission to shared filesystems or SSH connections through a uniform interface.It improves I/O efficiency by using symbolic links or moving files rather than copying or removing them.
- Machine support: DP-GEN supports workstations, HPC clusters using Slurm, PBS, or LSF, and cloud machines with launch and termination commands.Machine-specific resource settings are transformed into accepted job-submission scripts.
IV. EXAMPLES
The paper presents Cu as an example for generating and validating a DP model with DP-GEN, including tests summarized in Table I.
- Table I summarizes and describes the tests used to validate the DP model for Cu.
- Cu is used as a concrete example of the DP-GEN model-generation process.
- The generated Cu model is evaluated through a broad set of validation tests.
A. Generation of the model
DP-GEN generates the Cu model through iterative exploration, labeling, and training, using molecular dynamics sampling, uncertainty-based configuration selection, first-principles labeling, and DeePMD-kit training.
- Workflow: LAMMPS, VASP 5.4.4, and DeePMD-kit are used for exploration, labeling, and training, respectively.
- Exploration: The initial exploration starts from relaxed fcc, hcp, and bcc structures and randomly perturbs supercells, atomic positions, and cell vectors.
- Exploration: Later exploration prepares bulk and surface configurations and samples them with canonical molecular dynamics across temperature intervals.
- Selection and labeling: Configurations are selected using trust levels of σlo = 0.05 eV/Å and σhi = 0.2 eV/Å; 7,646 of 25 million sampled configurations are labeled.
- Selection and labeling: VASP labeling uses the PBE generalized gradient approximation, a 650 eV plane-wave cutoff, and a Monkhorst-Pack spacing of 0.1 Å^-1.
- Training: Four smooth DP models with different random seeds are trained using DeePMD-kit and 400,000 gradient-descent steps each.
B. Testing the model
The Cu DP model is tested against DFT, experiment, and MEAM across equilibrium, equation-of-state, phonon, surface, and defect properties, showing broad agreement with reference results.
- Validation scope: The validation covers equilibrium, defect, elastic, modulus, stacking-fault, equation-of-state, phonon, and surface-formation properties of Cu.
- Equilibrium properties: DP predictions agree well with DFT for all properties listed in Table II.
- Equation of state: DP reproduces DFT equation-of-state results for the standard crystalline structures considered, including fcc, hcp, dhcp, bcc, sc, and diamond.
- Equation of state: For sc, diamond, and bcc equation-of-state predictions, MEAM shows large deviations despite performing well near minima for fcc, hcp, and dhcp.
- Phonons: DP phonon-dispersion results agree very well with experiment and are significantly better than MEAM predictions.
- Surface formation: DP predicts bcc surface-formation energies close to DFT and significantly better than MEAM despite no explicitly labeled bcc surface data.
V. CONCLUSION
DP-GEN is presented as a scalable, flexible platform that modularizes exploration, labeling, and training for generating deep-learning PES models. The authors expect it to support diverse molecular and materials applications while reducing effort and improving simulation efficiency and cost.
- DP-GEN was used to generate a general-purpose PES model for Cu, with implementation details reported in the paper.
- DP-GEN separates exploration, training, and labeling under scheduler control, making the workflow highly modularized.The modular design is intended to require minimal effort when incorporating novel functionalities.
- DP-GEN can be extended to different first-principles codes, typically requiring file-conversion and job-submission scripts.
- The platform is expected to support systems including organic molecules, metals, semiconductors, and insulators, and applications beyond general-purpose PES generation.The paper also mentions investigation of specific applications.
- The workflow can be adapted to different software, operating systems, and cloud machines such as AWS.The authors also discuss connecting DP-GEN with general-purpose open-source platforms as future work.
- The authors expect DP-GEN users to obtain an optimal balance of efficiency and cost in atomic and molecular simulations.
Appendix A: Details of the exploration strategy and more numerical results
The appendix documents the copper exploration strategy and reports additional numerical results. It specifies simulation settings and candidate-selection conditions, including pressures used for NPT simulations and standard deviations among four models.
- Table A.1 reports the exploration strategy used for the copper system in each iteration.
- The strategy records the crystal structure, number and length of DPMD simulations, temperatures, statistical ensembles, and labeling-candidate percentages.
- Table A.2 reports predictions of surface formation energies for fcc, hcp, and bcc copper lattices.
- The standard deviations σ(Esf) are evaluated among four models.