The design space of E(3)-equivariant atom-centred interatomic potentials

Nature Machine Intelligence, Journal Year: 2025, Volume and Issue: 7(1), P. 56 - 67

Published: Jan. 15, 2025

Abstract Molecular dynamics simulation is an important tool in computational materials science and chemistry, the past decade it has been revolutionized by machine learning. This rapid progress learning interatomic potentials produced a number of new architectures just few years. Particularly notable among these are atomic cluster expansion, which unified many earlier ideas around atom-density-based descriptors, Neural Equivariant Interatomic Potentials (NequIP), message-passing neural network with equivariant features that exhibited state-of-the-art accuracy at time. Here we construct mathematical framework unifies models: expansion extended recast as one layer multi-layer architecture, while linearized version NequIP understood particular sparsification much larger polynomial model. Our also provides practical for systematically probing different choices this design space. An ablation study NequIP, via set experiments looking in- out-of-domain smooth extrapolation very far from training data, sheds some light on critical to achieving high accuracy. A much-simplified call BOTnet (for body-ordered tensor network), interpretable architecture maintains its benchmark datasets.

Language: Английский

---

Cartesian atomic cluster expansion for machine learning interatomic potentials

npj Computational Materials, Journal Year: 2024, Volume and Issue: 10(1)

Published: July 18, 2024

Abstract Machine learning interatomic potentials are revolutionizing large-scale, accurate atomistic modeling in material science and chemistry. Many use atomic cluster expansion or equivariant message-passing frameworks. Such frameworks typically spherical harmonics as angular basis functions, followed by Clebsch-Gordan contraction to maintain rotational symmetry. We propose a mathematically equivalent simple alternative that performs all operations the Cartesian coordinates. This approach provides complete set of polynormially independent features environments while maintaining interaction body orders. Additionally, we integrate low-dimensional embeddings various chemical elements, trainable radial channel coupling, inter-atomic message passing. The resulting potential, named Atomic Cluster Expansion (CACE), exhibits good accuracy, stability, generalizability. validate its performance diverse systems, including bulk water, small molecules, 25-element high-entropy alloys.

Language: Английский

---

Vibrational and thermal properties of amorphous alumina from first principles

Physical Review Materials, Journal Year: 2024, Volume and Issue: 8(4)

Published: April 2, 2024

Amorphous alumina is employed ubiquitously as a high-dielectric-constant material in electronics, and its thermal-transport properties are of key relevance for heat management electronic chips devices. Experiments show that the thermal conductivity depends significantly on synthesis process, indicating need theoretical study to elucidate atomistic origin these variations. Here we employ first-principles simulations characterize structure, vibrational properties, at densities ranging from 2.28 $3.49 \mathrm{g}/{\mathrm{cm}}^{3}$. Moreover, using machine-learned interatomic potential trained data, investigate how system size affects predictions conductivity, showing containing 120 atoms can already reproduce bulk limit conductivity. Finally, relying recently developed Wigner formulation transport, shed light interplay between topological disorder anharmonicity context conduction, former dominates over latter determining alumina.

Language: Английский

---

Benchmarking of machine learning interatomic potentials for reactive hydrogen dynamics at metal surfaces

Machine Learning Science and Technology, Journal Year: 2024, Volume and Issue: 5(3), P. 030501 - 030501

Published: July 3, 2024

Abstract Simulations of chemical reaction probabilities in gas surface dynamics require the calculation ensemble averages over many tens thousands events to predict dynamical observables that can be compared experiments. At same time, energy landscapes need accurately mapped, as small errors barriers lead large deviations probabilities. This brings a particularly interesting challenge for machine learning interatomic potentials, which are becoming well-established tools accelerate molecular simulations. We compare state-of-the-art potentials with particular focus on their inference performance CPUs and suitability high throughput simulation reactive chemistry at surfaces. The considered models include polarizable atom interaction neural networks (PaiNN), recursively embedded (REANN), MACE equivariant graph network, atomic cluster expansion (ACE). applied dataset hydrogen scattering low-index facets copper. All assessed accuracy, time-to-solution, ability simulate sticking function rovibrational initial state kinetic incidence molecule. REANN provide best balance between accuracy time-to-solution current gas-surface dynamics. PaiNN features causes significant losses computational efficiency. ACE fastest however, trained existing were not able achieve sufficiently accurate predictions all cases.

Language: Английский

---

Machine-learned atomic cluster expansion potentials for fast and quantum-accurate thermal simulations of wurtzite AlN

Journal of Applied Physics, Journal Year: 2024, Volume and Issue: 135(8)

Published: Feb. 22, 2024

Thermal transport in wurtzite aluminum nitride (w-AlN) significantly affects the performance and reliability of corresponding electronic devices, particularly when lattice strains inevitably impact thermal properties w-AlN practical applications. To accurately model with high efficiency, we develop a machine learning interatomic potential based on atomic cluster expansion (ACE) framework. The predictive power ACE against density functional theory (DFT) is demonstrated across broad range w-AlN, including ground-state parameters, specific heat capacity, coefficients expansion, bulk modulus, harmonic phonon dispersions. Validation conductivity further carried out by comparing ACE-predicted values to DFT calculations experiments, exhibiting overall capability our sufficiently describing anharmonic interactions. As application, perform dynamics analysis using unravel effects biaxial which identified as significant tuning factor for near-junction design w-AlN-based electronics.

Language: Английский

---

A theoretical case study of the generalization of machine-learned potentials

Computer Methods in Applied Mechanics and Engineering, Journal Year: 2024, Volume and Issue: 422, P. 116831 - 116831

Published: Feb. 14, 2024

Machine-learned interatomic potentials (MLIPs) are typically trained on datasets that encompass a restricted subset of possible input structures, which presents potential challenge for their generalization to broader range systems outside the training set. Nevertheless, MLIPs have demonstrated impressive accuracy in predicting forces and energies simulations involving intricate complex structures. In this paper we aim take steps towards rigorously explaining excellent observed properties MLIPs. Specifically, offer comprehensive theoretical numerical investigation context dislocation simulations. We quantify precisely how such is directly determined by few key factors: size choice observations (e.g., energies, forces, virials), level achieved fitting process. Notably, our study reveals crucial role virials ensuring consistency Our series careful experiments encompassing screw, edge, mixed dislocations, supports existing best practices literature but also provides new insights into design data sets loss functions.

Language: Английский

---

Surrogate Models for Vibrational Entropy Based on a Spatial Decomposition

Multiscale Modeling and Simulation, Journal Year: 2025, Volume and Issue: 23(1), P. 514 - 544

Published: Feb. 17, 2025

Language: Английский

---

The atomic formation mechanism of GP zones in Al-Cu alloys: Insights from cluster expansion coupled with Monte Carlo simulation

Computational Materials Science, Journal Year: 2025, Volume and Issue: 252, P. 113798 - 113798

Published: Feb. 26, 2025

Language: Английский

---

The evolution of machine learning potentials for molecules, reactions and materials

Chemical Society Reviews, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 1, 2025

This review offers a comprehensive overview of the development machine learning potentials for molecules, reactions, and materials over past two decades, evolving from traditional models to state-of-the-art.

Language: Английский

---

Atomic cluster expansion without self-interaction

Journal of Computational Physics, Journal Year: 2024, Volume and Issue: 515, P. 113271 - 113271

Published: July 10, 2024

The Atomic Cluster Expansion (ACE) (Drautz (2019) [14]) has been widely applied in high energy physics, quantum mechanics and atomistic modeling to construct many-body interaction models respecting physical symmetries. Computational efficiency is achieved by allowing non-physical self-interaction terms the model. We propose analyze an efficient method evaluate parameterize orthogonal, or, non-self-interacting cluster expansion present numerical experiments demonstrating improved conditioning more robust approximation properties than original regression tasks both simplified toy problems applications machine learning of interatomic potentials.

Language: Английский

---