PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine learning force fields: Construction, validation, and outlook

Venkatesh Botu, Rohit Batra, James Chapman, Rampi Ramprasad

arXiv:1610.02098v2cond-mat.mtrl-sci

TL;DR

The paper addresses the need for inexpensive, accurate, and versatile force fields beyond the limited scales and domains of conventional approaches. It presents a quantum-mechanics-informed AGNI workflow that learns atomic forces from environment fingerprints, validates the resulting Al model on complex phenomena, and estimates prediction uncertainty. The approach supports simulations of Al surface melting and stress-strain behavior while enabling identification of new environments for continued improvement.

  • Problem

    Quantum-mechanical methods are computationally demanding, while routine simulations remain limited to nanometer and picosecond scales and conventional force fields have restricted applicability.

  • Method

    AGNI constructs an atomic-force model from diverse quantum-mechanical reference environments, environment fingerprints, representative training data, machine learning, testing, and uncertainty estimation.

  • Results

    The constructed Al force field was used to simulate surface melting and stress-strain behavior, extending atomistic simulations to more complex materials phenomena.

  • Takeaways & Limitations

    Uncertainty monitoring can identify new environments and guide training-set augmentation, supporting force fields that are truly adaptable.

  • Takeaways & Limitations

    The force field’s applicability remains tied to the environments represented in its reference and training data, and the energy calculation described is restricted.

Abstract

from arXiv · show

Force fields developed with machine learning methods in tandem with quantum mechanics are beginning to find merit, given their (i) low cost, (ii) accuracy, and (iii) versatility. Recently, we proposed one such approach, wherein, the vectorial force on an atom is computed directly from its environment. Here, we discuss the multi-step workflow required for their construction, which begins with generating diverse reference atomic environments and force data, choosing a numerical representation for the atomic environments, down selecting a representative training set, and lastly the learning method itself, for the case of Al. The constructed force field is then validated by simulating complex materials phenomena such as surface melting and stress-strain behavior - that truly go beyond the realm of $ab\ initio$ methods both in length and time scales. To make such force fields truly versatile an attempt to estimate the uncertainty in force predictions is put forth, allowing one to identify areas of poor performance and paving the way for their continual improvement.

Introduction

Machine-learning force fields combine quantum-mechanical reference data with flexible force prediction, addressing the restricted applicability and cost of conventional approaches. The AGNI workflow constructs, tests, and uncertainty-quantifies an atomic-force model for complex Al simulations.

  • Motivation: Conventional semi-empirical force fields are inexpensive but restricted to the chemistries and environments considered during parameterization.Their systematic, continuous improvement for newer environments is unclear.
  • Motivation: Machine-learning interatomic potentials can surpass conventional potentials in accuracy and versatility while approaching quantum-mechanical accuracy at far lower cost within the training domain.The supported domain remains bounded by the configurational and chemical environments represented in the benchmark dataset.
  • AGNI approach: AGNI learns each atom’s vectorial force directly from a numerical fingerprint of its local environment rather than from a global total-energy representation.Atomic forces are local quantities, and many geometry-optimization and molecular-dynamics simulations require them as the sole necessary input.
  • Workflow: The five-step AGNI workflow generates quantum-mechanical reference environments and forces, fingerprints environments, selects representative training data, learns and tests the force mapping, and estimates prediction uncertainty.Training-set clustering is intended to preserve reference-data diversity while optimizing learning.
  • Uncertainty and adaptation: Uncertainty estimates are used to identify environments outside the force field’s applicability domain and support adding new environments to the training set.This feedback loop is presented as a route toward adaptable force fields.
  • Validation: The Al force field extends prior validation to surface melting and stress-strain behavior while retaining accurate force-based simulation capabilities.The paper also reports that total potential energies can be recovered by integrating atomic forces.

Generating reference data

AGNI force-field construction begins by assembling diverse equilibrium and non-equilibrium aluminum environments, calculating their forces with DFT, and partitioning them into datasets for training, testing, and transferability evaluation.

  • Force calculation: DFT calculations provide the vectorial force components on atoms in the collected equilibrium configurations.These force data are used to learn the underlying potential-energy surface indirectly.
  • Configuration sampling: Reference configurations span bulk, surfaces, defects, clusters, grain boundaries, lattice expansion and compression, and edge dislocations.These environments are intended to mimic the diversity of atomic situations relevant to generalizable force fields.
  • Dataset design: Four datasets, A–D, increase in the complexity and diversity of included environments, with separate training and test sets for each.Training sets construct the force fields, while test sets validate predictive performance.
  • Dataset design: Dataset E contains configurations excluded from force-field construction and is used solely to demonstrate transferability.The dataset has only a test set rather than both training and test sets.

Fingerprinting reference environments

AGNI represents each atom’s local environment with a directional fingerprint designed to map directly to vectorial force components while respecting basic atomic transformations. The fingerprint combines projected neighbor distances, multiple coordination-shell widths, and distance damping within a cutoff.

  • The atomic fingerprint is the critical representation linking an atom’s environment to its directional force components.Its design must reflect force direction and preserve required permutation, translation, and rotation behavior.
  • Neighbor information enters through pairwise distances and scalar projections along a chosen direction, so the fingerprint resolves directional coordination.The projection supplies the directional dependence needed to map individual force components.
  • The representation transforms consistently with force vectors under rotation and remains unchanged under permutation or translation of atoms.The net force magnitude remains unchanged before and after rotation.
  • A spherical mesh of arbitrary directions expands the existing ab initio reference database without additional costly ab initio calculations.This expansion increases diversity and completeness but also introduces extensive redundancy, making later down-sampling necessary.
  • Training directly on millions of reference environments is computationally demanding and may produce misbehaved models, so further down-sampling is essential.The redundancy arises from expanding the reference environments to ensure diversity and completeness.

Clustering reference data

PCA reveals redundancy and coverage in the high-dimensional fingerprint database, enabling representative training environments to be selected by grid sampling rather than biased random sampling. Two principal components capture more than 99% of an 8-dimensional fingerprint’s information, and dataset E lies within dataset D’s domain.

  • PCA projects high-dimensional atomic fingerprints into orthogonal principal components, making their structure and redundancy easier to inspect.The 8-dimensional fingerprint is projected onto a lower-dimensional space for visualization and sampling.
  • More than 99% of the information in the 8-dimensional fingerprint is captured by two principal components, whose scores plot clusters similar environments.Labeled clusters include adatoms, surfaces, and vacancies.
  • Dataset D spans most of the PCA space, indicating broad environmental diversity and making it a strong starting point for training-data sampling.Isolated clusters provide additional locations from which reference data can be sampled.
  • Dataset E lies within dataset D’s domain, suggesting that a model trained on D can predict E’s forces despite never seeing E during training.This expectation follows from the overlap of their transformed fingerprint domains.
  • Random sampling follows the dataset’s probability distribution and can miss sparsely populated regions, whereas PCA grid sampling promotes uniform coverage.The PCA space is divided into uniform sub-grids, with training points sampled from each.
  • Clustering is important because learning and prediction costs scale as O(n^3) and O(n), respectively, with training-set size n.

Learning algorithm

AGNI uses nonlinear kernel ridge regression to interpolate force components from similarities between atomic fingerprints. The framework evaluates predictions with complementary error metrics and tunes kernel parameters through cross-validation and regularization.

  • Nonlinear kernel ridge regression is used as the learning method for mapping atomic fingerprints to force components.Its nonlinear form is suited to the relationship between an atom’s environment and experienced force.
  • KRR compares an atom’s fingerprint with reference fingerprints to make an interpolative prediction of the corresponding force component.
  • The model uses reference-environment fingerprints, with Nt denoting the total number of reference environments considered.
  • Weight coefficients and the length-scale parameter are selected during training using cross-validation and regularization.
  • Performance is assessed with mean absolute error, maximum absolute error, and 2σ to reduce bias from relying on a single metric.

Constructing the force field

The force field is constructed by testing fingerprint resolution and training-set size, then validating models across diverse held-out environments. Eight η values with 1000 diverse environments provide the selected operating point, while broad training diversity supports generalization and AGNI agrees closely with DFT on dataset E.

  • Construction and validation: Models are trained on datasets A–D and validated on test datasets A–E to assess generalizability beyond the training environments.
  • Convergence tests: Increasing fingerprint resolution lowers error below approximately 0.05 eV/Å, while larger training sets help only after adequate resolution is reached.For fingerprints with 2 or 4 η values, training-set size has no effect on predictive capability.
  • Convergence tests: At least 8 η values are needed to distinguish atomic environments sufficiently for the learning algorithm to work effectively.
  • Convergence tests: Eight η values and 1000 training environments are sufficient, beyond which training costs increase without significant error reduction.
  • Computational cost: Each AGNI prediction takes approximately 0.1 ms/atom/core, compared with approximately 1 ks/atom/core for DFT.
  • Training data choice: Diverse training environments produce low MAE, MAX, and 2σ errors and yield the highly generalizable model MD used subsequently.Models trained on narrower domains can perform poorly outside those domains, especially under MAX.
  • Testing out-of-domain configurations: AGNI predictions for grain boundaries, lattice expansion/compression, and edge dislocations in dataset E show excellent agreement with DFT.These configurations were not observed during training, and AGNI is compared with an Al EAM potential in the same tests.

Validating the force field

AGNI force fields were validated on complex aluminum phenomena beyond routine first-principles simulation scales, including surface melting, stress–strain behavior, and energy evolution. The simulations reproduced key reference behavior while exposing implementation boundaries.

  • Surface melting: AGNI simulations captured surface melting in aluminum, with melting beginning near 950 K and reaching the bulk by 1200 K.The onset is close to the experimental value of approximately 933 K.
  • Surface melting: The EAM comparison overestimated the melting temperature at approximately 1100 K.AGNI used high-temperature molecular-dynamics reference data despite not explicitly training on liquid environments.
  • Stress–strain behavior: 107 GPa for AGNI’s C11 elastic coefficient agreed with the 105 GPa ab initio result.An EAM potential produced 106 GPa, while the AGNI force–stress relation also supported quantum-mechanical accuracy for stresses.
  • Scope and limitations: The force-based implementation can describe stress only along non-periodical directions.This is identified as a limitation of the implementation.
  • Energy evolution: AGNI-derived energies were conserved during molecular dynamics and maintained the correct temperature ordering.The difference from ab initio energies was approximately 4 meV/atom.
  • Energy estimation: Reconstructing energy requires a pathway connecting configurations in phase space for accurate force integration.The pathway may be temporal or follow a reaction coordinate.

Uncertainty quantification

The uncertainty model uses the minimum distance from a new atomic environment to reference training environments as a predictor of force-error variance. This identifies poorly represented environments and supports systematic retraining of the force field.

  • Validation and adaptation: Flagged environments can be accumulated for retraining, enabling systematic improvement in force-field accuracy and generalizability.This provides a route toward adaptive AGNI force fields.
  • Distance-based uncertainty: As dmin increases, the variance in force error increases, indicating failures farther from reference training environments.The model treats dmin as a descriptor of closeness to reference cases.
  • Uncertainty model: The uncertainty analysis excludes dmin < 10^-3 because of data scarcity and dmin > 10^-1 because of large predictive errors.These ranges were omitted during binning and statistical fitting.
  • Uncertainty model: A polynomial relation between dmin and the error standard deviation provides an analytical uncertainty estimate.The relation is fitted from standard deviations computed across dmin bins.
  • Uncertainty model: The uncertainty intervals correspond to 68.2% confidence when one standard deviation is used.Higher confidence levels, such as 2s, may also be selected depending on user needs and computational resources.
  • Validation and adaptation: The model flags high-uncertainty atomic environments in grain boundaries, lattice deformation, and dislocations.Figure 13 overlays these uncertainties as error bars on DFT and ML force predictions.

Outlook and Summary

The paper presents AGNI force fields as accurate, computationally inexpensive tools for complex atomistic simulations, combining a systematic construction workflow with uncertainty estimation. It also identifies multi-element systems and rapidly expanding configuration spaces as remaining challenges.

  • Construction and scope: AGNI directly learns forces from local atomic environments using numerical fingerprints, representative sampling, and force learning for elemental aluminum.The workflow systematically generates reference data, represents environments, samples non-redundant data, and learns forces.
  • Adaptive improvement: Uncertainty quantification is proposed to identify the domain of applicability of data-driven force predictions and support adaptive improvement.The approach is presented as a pathway toward continual improvement in accuracy and generalizability.
  • Outlook and limitations: Extending AGNI to multi-element systems remains necessary for exploring diverse chemistries.The authors state that the elemental-system recipe is directly transferable to multi-elemental situations.
  • Outlook and limitations: Exponentially increasing configuration spaces challenge the proposed non-linear regression method and require approaches able to handle large quantities of data.This is identified as a remaining challenge as materials and chemical systems become more complex.
  • Conclusion: AGNI provides high-fidelity force predictions at a fraction of the cost of conventional atomistic simulations.The authors describe its prospects for accelerating materials and chemical simulations.
  • Demonstrated capabilities: The force-field framework was extended to simulate aluminum surface melting and stress–strain behavior.These phenomena broaden the demonstrated scope beyond earlier structural, vibrational, and simpler molecular-dynamics applications.
Loading 1610.02098v2…