Source-linked AI summary
Machine-learning interatomic potentials for materials science
Y. Mishin
TL;DR
Large-scale atomistic materials simulations need efficient interatomic potentials, but traditional models and newer ML models embody different trade-offs in physical grounding, accuracy, and transferability. The paper reviews these classes, compares their strengths and weaknesses, and introduces physically informed ML potentials that combine ML regression with physics-based potentials. Within their validity boundaries, ML potentials can approach DFT accuracy at much lower computational cost, while the physically informed approach is presented as improving transferability to unknown configurations.
Problem
Large-scale materials simulations require efficient energy and force predictions, while traditional potentials have limited accuracy and ML potentials lack physics-based transferability to unknown structures.
Method
The paper reviews traditional and ML interatomic potentials and introduces physically informed ML potentials that integrate ML regression with a physics-based interatomic potential.
Results
ML potentials can provide nearly DFT-accurate energy and force predictions at orders-of-magnitude lower computational cost within their validity boundaries, while physically informed models improve transferability compared with purely mathematical models.
Takeaways & Limitations
ML potentials extend DFT-level calculations to larger systems and longer simulations, and physically informed designs support the development of more general-purpose ML potentials.
Takeaways & Limitations
The major limitation of ML potentials is their lack of physics-based transferability, making predictions outside the interpolation domain potentially uncontrollable and physically meaningless.
Abstract
from arXiv · showhide
Large-scale atomistic computer simulations of materials rely on interatomic potentials providing computationally efficient predictions of energy and Newtonian forces. Traditional potentials have served in this capacity for over three decades. Recently, a new class of potentials has emerged, which is based on a radically different philosophy. The new potentials are constructed using machine-learning (ML) methods and a massive reference database generated by quantum-mechanical calculations. While the traditional potentials are derived from physical insights into the nature of chemical bonding, the ML potentials utilize a high-dimensional mathematical regression to interpolate between the reference energies. We review the current status of the interatomic potential field, comparing the strengths and weaknesses of the traditional and ML potentials. A third class of potentials is introduced, in which an ML model is coupled with a physics-based potential to improve the transferability to unknown atomic environments. The discussion is focused on potentials intended for materials science applications. Possible future directions in this field are outlined.
1 Introduction
This introduction establishes interatomic potentials as the computational foundation of large-scale atomistic materials simulations and frames the field's shift toward machine-learning approaches. The article reviews traditional, ML, and physically informed potentials, emphasizing their complementary strengths and weaknesses for materials applications.
- Motivation: Interatomic potentials predict atomic energies and forces, enabling large-scale molecular-dynamics and Monte Carlo simulations whose accuracy depends strongly on potential quality.Classical-potential computations scale linearly with atom number and provide access to broad length and time scales.
- Traditional potentials: Traditional potentials have a long history of successful materials applications, but their quality varies widely and development remains heavily dependent on human expertise.Some potentials achieve high accuracy and widespread use, whereas many others remain poor quality or unused beyond their original publication.
- Machine-learning potentials: Machine-learning potentials emerged as a major research direction by constructing interatomic models through machine-learning methods and large quantum-mechanical reference databases.The article focuses mainly on supervised, high-dimensional regression models for materials science rather than unsupervised learning approaches.
- Article scope and organization: The review focuses on materials-science applications including microstructure, defects, mechanical and thermal properties, and alloy thermodynamics and kinetics.Chemical applications, molecular matter, molecule-surface interactions, and many unsupervised-learning topics are outside the article's scope.
- Article scope and organization: The article's leading theme is a comparison of traditional and ML potentials, whose strengths and weaknesses are described as partly complementary.A summary comparison is presented in Table 1, followed by more detailed analysis.
2 The traditional interatomic potentials
Traditional interatomic potentials enable large-scale atomistic simulations by mapping local atomic environments to total energies and forces through physically motivated, computationally efficient functions. Their main trade-off is limited accuracy versus often useful transferability beyond the training data, while construction remains expertise-intensive.
- 2.1 What are the potentials, and why do we need them?: DFT calculations provide highly accurate energies and forces but scale as N^3 or slower, limiting static systems to a few hundred atoms.Classical potentials address materials processes requiring many atoms and extensive statistical sampling.
- 2.1 What are the potentials, and why do we need them?: A potential represents the potential energy surface as a function of atomic positions, computes local atomic energies, and sums them to obtain total energy.Forces follow from the gradient of the total energy with respect to atomic positions.
- 2.2 The physical basis of traditional potentials: Traditional potentials use physically motivated functional forms, with examples including EAM, MEAM, ADP, Tersoff, and Stillinger-Weber models for different bonding types.Their forms incorporate assumptions about interatomic bonding and preserve invariance under rotations, translations, and atom permutations.
- 2.3 Training of traditional potentials: Training typically optimizes a small set of global parameters against experimental properties and relatively few DFT energies or forces.Developers then test additional properties and iteratively adjust weighted loss terms using expert judgment.
- 2.4 Accuracy and transferability of traditional potentials: Traditional potentials offer fast, parallelizable calculations for very large systems, but their approximate physics and few parameters limit accuracy for subtle structures and effects.Nevertheless, their physical basis often supports meaningful predictions outside the training dataset when the assumed bonding remains applicable.
- 2.5 Classification of interatomic potentials: Potential applications include general-purpose models spanning broad properties, special-purpose models targeting one simulation type, and artificial models varying selected physical properties.The same functional form can support any of these application categories.
3 Machine-learning potentials
ML interatomic potentials represent local atomic environments with invariant structural descriptors, regress atomic energies, and sum them to reconstruct the system’s potential-energy surface. Their flexibility yields high interpolation accuracy and broad material applicability, but extrapolation and property transferability remain limited, motivating special-purpose and physically informed approaches.
- Core idea: ML potentials map the system’s configurational space onto a potential-energy surface by numerically interpolating DFT reference energies with many adjustable parameters.They are trained to approximate the potential-energy surface rather than particular physical properties.
- Local representation: Local environments are converted into fixed-length, invariant structural descriptors, which a regression model maps to atomic energies before summation yields the total energy.The descriptors preserve translation, rotation, and permutation invariance while replacing variable-size neighbor coordinates with a K-dimensional feature vector.
- Training: Training compares summed atomic energies with supercell reference energies, reflecting the absence of one-to-one correspondence between local feature vectors and supercell energies.The potential predicts individual atomic energies, which must be summed before comparison with the DFT supercell energy.
- Accuracy and transferability: ML potentials readily reach training accuracies of several meV/atom, but predictions for configurations outside the training database require extrapolation that can be unpredictable or physically meaningless.This limitation follows from the flexibility of high-dimensional numerical regression and is not resolved by low training error alone.
- Accuracy and transferability: Even potentials fit to similar PES errors can predict substantially different physical properties, so selecting a reliable potential requires testing multiple versions and adding targeted reference structures.Training the PES does not automatically guarantee accurate property predictions.
- Database strategies: Active learning can generate economical databases within a chosen configuration domain, but behavior outside that domain remains uncontrollable and full automation of general-purpose potentials is not expected.The covered domain depends on the simulation protocol or structure-generation strategy, while special-purpose automation remains plausible.
- Applications: One practical strategy is to develop special-purpose potentials for particular tasks, retaining high accuracy and computational efficiency while keeping simulations in the interpolation regime.The approach avoids claiming broader applicability than the training domain supports.
4 Physically-informed machine-learning potentials
Physically-informed ML potentials combine physics-based interatomic models with ML regressions to retain high training accuracy while improving transferability to unfamiliar atomic environments.
- Energy calculation: The ML regression predicts local potential parameters rather than atomic energies directly, after which a physics-based potential computes each atom’s energy.For atom i, the regression outputs parameters p_i suited to its local environment, and Φ(R_i, p_i) produces E_i.
- Transferability: Physics-guided extrapolation uses the embodied interaction model to make predictions for new environments rather than relying solely on mathematical interpolation.This is expected to expand the reliability domain relative to purely mathematical models, although physically-informed potentials can still fail far outside familiar environments.
- PINN implementation: PINN combines neural-network regression with an analytical bond-order potential whose form is general enough for metals and nonmetals.The improved formulation fits global parameters and adds small local perturbations that vary with atomic environments.
- Reported applications: The improved PINN scheme significantly improves transferability without compromising accuracy or increasing computational cost relative to the original formulation.General-purpose PINN potentials for Al and Ta accurately describe broad mechanical and thermal property spectra and other materials-science properties.
- Core concept: Physically-informed ML potentials form a distinct class that combines explicit physical interaction models with ML descriptor-regressor training on DFT data.They can represent many-body interactions, bond-order effects, and bond screening while learning from reference calculations.
5 Summary and outlook
ML potentials offer DFT-like accuracy at much lower computational cost, but their extrapolation to unknown structures remains physically unreliable. The paper therefore advocates physics-informed ML potentials and anticipates a shift toward materials discovery and explanation.
- Current status: ML potentials can predict energies and atomic forces with nearly DFT accuracy while scaling linearly with atom number and running orders of magnitude faster than DFT.Their accuracy can be improved by expanding the DFT reference database and continuing training.
- Limitations: Their major limitation is weak physics-based transferability: outside the interpolation domain, predictions may be uncontrollable and physically meaningless.Monitoring trajectories or using on-the-fly training can mitigate the risk when simulations remain similar to the training data.
- Future directions: Physics-informed ML potentials couple ML regression with physics-based interatomic potentials to retain training accuracy while improving transferability without significant computational overhead.The paper presents PINN potentials as a promising route toward general-purpose ML potentials.
- Limitations: For multicomponent materials, traditional potentials are often inheritable, whereas most ML potentials require new constructions when chemical elements are added.Inheritance is especially desirable for broadly applicable potentials intended for multipurpose use.
- Outlook: The field is expected to move from methodology development toward discovering materials phenomena, explaining known behavior, and predicting otherwise inaccessible properties.The authors envision tighter integration between flexible ML models and physical guidance.
Appendix A: Table of abbreviations
The appendix defines abbreviations used throughout the paper for interatomic-potential methods, simulation techniques, computational tools, and related concepts.
- Abbreviations: The abbreviation list covers methods such as ACE, BOP, GAP, MTP, NN, PINN, SNAP, and SOAP, alongside DFT, MD, MC, and LAMMPS.It also expands AIMD, FCC, PES, and VA.
Appendix B: The bond-order potential in PINN
The PINN model uses a bond-order potential whose terms represent short-range repulsion, long-range bonding, angular bond order, screening, and on-site energy contributions. Eight local parameters are adjusted by the neural network, while cutoff parameters remain global.
- Potential structure: The BOP energy includes smoothly truncated neighbor interactions, with the cutoff distance rc controlling the interaction range and d controlling truncation width.The summation runs over neighbors separated from atom i by distance r_ij.
- Bonding terms: Exponential terms describe short-range repulsion and long-range attraction, while bond-order and angular terms represent bonding dependence on local geometry.The bond count is weighted by bond angles θ_ijk.
- Screening: Bond screening is computed as a product of partial factors S_ijk, with nearby atoms contributing stronger screening to bond i-j.An atom located directly on the bond nearly breaks it through strong screening.
- On-site energy: The on-site energy represents promotion energy for covalent bonding and embedding energy in metals.This term connects the BOP formulation to the embedded-atom method for metallic systems.
- PINN parameterization: The BOP has ten parameters, with eight locally adjusted by the neural network and d and rc treated as global parameters.The current PINN formulation therefore uses an output layer with m = 8 nodes, with species-dependent parameters in multicomponent systems.
Appendix C: Local structural descriptors in PINN
PINN encodes each atomic environment with rotationally invariant local descriptors built from radial Gaussian functions and three-body angular terms. The descriptors feed a fixed-length neural-network feature vector, although their completeness and superiority are not claimed.
- Scope: The authors report that the descriptors performed well in tests but do not claim superiority over alternatives or completeness as a basis set.This limits the strength of conclusions about descriptor optimality.
- Descriptor construction: For single-component materials, the local environment is encoded using rotationally invariant three-body parameters involving neighbor distances and bond angles.Legendre polynomials provide the angular basis, while radial functions are Gaussian.
- Descriptor construction: The descriptor truncation radius is 1.5rc so atoms outside the BOP cutoff can be included when they affect atomic energy through screening.This extends descriptor coverage beyond the direct interaction cutoff.
- Descriptor construction: Rotationally invariant descriptors are formed by summing products of local-density projections, with j = k terms supplying additional two-body descriptors.The formulation uses Gaussian radial functions and the spherical-harmonics summation theorem.
- Feature representation: A fixed-length feature vector of size K = lmaxnmax is supplied to the K-node neural-network input layer.Multicomponent descriptors are formed by juxtaposing species-resolved parameters, whose size can grow quadratically or cubically with component count.