Source-linked AI summary
Active Learning of Uniformly Accurate Inter-atomic Potentials for Materials Simulation
Linfeng Zhang, De-Ye Lin, Han Wang, Roberto Car, Weinan E
TL;DR
Materials simulation needs PES models that retain ab initio accuracy and efficiency while remaining reliable beyond their training configurations. The paper proposes DP-GEN, an active-learning loop that explores configurations, labels selected cases with accurate reference calculations, and retrains Deep Potential models. Applied to Al, Mg, and Al-Mg alloys, DP-GEN produces uniformly accurate predictions for diverse properties, including structures not explicitly used for training.
Problem
PES models must be accurate and efficient, yet machine-learning predictions lack guaranteed quality far from training data and large ab initio datasets are costly.
Method
DP-GEN iteratively explores configurations, selects uncertain cases using model deviation, generates DFT reference data, and retrains an ensemble-based Deep Potential model.
Results
DP-GEN produces uniformly accurate PES models for Al, Mg, and Al-Mg alloys, including crystals, defects, and surfaces absent from the training data.
Takeaways & Limitations
The procedure constructs broadly applicable PES models while using a minimal set of accurately labeled reference configurations.
Takeaways & Limitations
DP-GEN may fail when high-dimensional configuration spaces or large free-energy barriers impede sampling, and a large model deviation is sufficient but not necessary for poor performance.
Abstract
from arXiv · showhide
An active learning procedure called Deep Potential Generator (DP-GEN) is proposed for the construction of accurate and transferable machine learning-based models of the potential energy surface (PES) for the molecular modeling of materials. This procedure consists of three main components: exploration, generation of accurate reference data, and training. Application to the sample systems of Al, Mg and Al-Mg alloys demonstrates that DP-GEN can produce uniformly accurate PES models with a minimal number of reference data.
INTRODUCTION
Accurate and efficient PES models are needed because ab initio methods are accurate but inefficient, while empirical force fields offer limited accuracy guarantees. The paper adapts active learning to generate PES data on the fly and demonstrates uniformly accurate models for Al, Mg, and Al-Mg alloys.
- Motivation: Ab initio methods provide accuracy but are highly inefficient, whereas empirical force fields are efficient but offer limited accuracy guarantees.
- Motivation: Machine-learning PES models can combine ab initio accuracy with efficiency, but their predictions are not guaranteed for configurations far from training data.
- Active-learning strategy: Active learning seeks general-purpose models with a minimal number of training data when labeling is expensive.
- Active-learning strategy: For PES modeling, active learning consists of exploration, labeling selected configurations with reference calculations, and training an updated model.
- Paper objectives: DP-GEN uses Deep Potential models, DPMD sampling, model deviation, and DFT labeling to construct PES models for Al, Mg, and Al-Mg alloys.
- Paper objectives: The resulting model achieves uniform accuracy for crystals, defects, and surfaces that were not explicitly included in training data.
- Active-learning strategy: Unlike conventional active learning over an existing unlabeled dataset, this procedure generates its dataset on the fly through sampling.
METHODOLOGY
DP-GEN iteratively explores configurations, labels selected structures with electronic-structure calculations, and retrains Deep Potential models. Its sampler combines varied initial structures with volume-temperature exploration, while an ensemble-force deviation indicator identifies uncertain configurations.
- Scheme: DP-GEN initializes a small labeled dataset, trains preliminary DP models, and repeats exploration, labeling, and training iterations.
- Model: The DP model represents total potential energy as a sum of atomic energies depending on local neighbor environments within a cutoff radius.
- Model: A deep neural network parameterizes each species-dependent atomic energy, capturing nonlinear and many-body interatomic interactions.
- Sampler: The sampler explores configuration space by selecting diverse initial crystal structures and varying volume-temperature conditions.
- Model: Feature encoding preserves translational, rotational, and permutational symmetries, while position derivatives provide forces for training against DFT energies and forces.
- Indicator: The indicator uses prediction deviations among DP models trained from different parameter initializations to identify configurations needing labeling.
- Indicator: Maximum force standard deviation is preferred over energy deviation and average atomic deviation because it better signals local prediction failures.
RESULTS
DP-GEN yields DP models that agree satisfactorily with DFT and experiment across diverse Al, Mg, and Al-Mg alloy properties, while using a small labeled subset of explored configurations. The models also show accuracy beyond explicitly sampled structures and training-volume ranges.
- Active-learning efficiency: 650 million configurations were explored, but only 0.0044% were selected for ab initio labeling.The resulting labeled data were used to train a DP model for the Al-Mg system.
- Pure-element properties: DP predictions agree well with DFT for Al equations of state across FCC, HCP, DHCP, BCC, SC, and diamond structures.DP also reproduces small FCC–DHCP and DHCP–HCP energy differences of 12 meV/atom and 19 meV/atom, respectively.
- Pure-element properties: DP phonon dispersions agree very well with experiment, whereas MEAM poorly reproduces short-wavelength, especially zone-boundary, phonons.For liquid Al, MEAM is overstructured and its diffusion coefficient is one order of magnitude smaller than experiment or DP.
- Simulation scale: DP retains ab initio-level property accuracy while enabling simulations of large systems, including liquid calculations with 4000 atoms using modest computational resources.The paper presents this efficiency as an opportunity to extend the power of ab initio methods.
- Surface properties: DP predicts Al and Mg surface formation energies close to DFT, while MEAM performs worse across the investigated surfaces.The study examines non-equivalent surfaces with Miller indices smaller than 4 for Al and smaller than 3 for Mg.
- Alloy properties: Across almost all tested Al-Mg alloy cases, DP agrees satisfactorily with DFT and is significantly more accurate than MEAM.The comparison covers formation energies, equilibrium volumes, elastic constants, vacancy formation energies, interstitial pathways, and surface energies for structures not explicitly included in training.
SUMMARY AND OUTLOOK
DP-GEN is presented as a general, practical, and modular scheme that progressively improves models through efficient exploration and targeted DFT labeling. Its applicability extends beyond the demonstrated metallic systems, but sampling and indicator failures remain important limitations requiring validation.
- Summary and outlook: DP-GEN builds Al-Mg models without an existing DFT database or exhaustive structure enumeration, using DPMD exploration and DFT labeling only for high-deviation configurations.The MP database was used only for testing, while selected configurations were progressively labeled and added to improve the model.
- Summary and outlook: The scheme modularizes training, exploration, and labeling, allowing enhanced sampling or genetic algorithms to be incorporated into exploration.The accumulated model and data can also support restarting exploration when new species are added.
- Summary and outlook: DP-GEN may extend to ceramics and polymers, provided the model, indicator, and sampler remain applicable to their bonding and configuration spaces.The authors frame this as a possibility rather than a demonstrated result.
- Summary and outlook: High-dimensional configuration spaces, large free-energy barriers, and misleadingly small indicator variance can cause DP-GEN to miss poorly described physics.The authors state that rigorous mathematical theory for the indicator is missing and recommend systematic validation before exploring new physics.
Simulation protocol
The simulation protocol uses a smooth deep potential with a 9 Å cutoff, deviation-based configuration selection during exploration, and DFT labeling with specified VASP settings.
- Simulation protocol: The smooth Deep Potential uses a 9 Å cutoff, cosine switching from 2 Å to 9 Å, embedding-net sizes {25, 50, 100}, and fitting-net sizes {240, 240, 240}.A skip connection connects neighboring layers, giving the architecture a ResNet-like structure.
- Simulation protocol: Configurations are selected for labeling when model deviation lies in [0.05, 0.15] eV/Å for Al and Al-Mg, or [0.03, 0.13] eV/Å for Mg.The exploration strategy and explored and labeled counts are reported separately for the three systems.
Additional simulation results
Supplementary results compare DP and MEAM with DFT and experiment for Mg equation-of-state and liquid-structure calculations. The Mg EOS comparison excludes high-volume HCP and DHCP points where DFT relaxations fail.
- Additional simulation results: Mg equation-of-state results compare DP solid lines and MEAM dashed lines against DFT cross points, after shifting DP and MEAM energies to align MEAM FCC with DFT.DFT relaxations fail for some HCP and DHCP structures above 44 Å^3, so those EOS portions are omitted.
- Additional simulation results: Liquid Mg RDFs at P = 1 bar and T = 953 K compare DP and MEAM predictions with experimental data, including a 3.5 Å ≤ r ≤ 7 Å zoom-in.The figure passage identifies the experimental comparison and the inserted zoomed range.