Implementing a neural network interatomic model with performance portability for emerging exascale architectures

Modelling ultrafast dynamics at a conical intersection with regularized diabatic states: An approach based on multiplicative neural networks

In silico property prediction based on density functional theory (DFT) is increasingly performed for crystalline materials. Whether quantitative agreement with experiment can be achieved with current methods is often an unresolved question, and may require detailed examination of physical effects such as electron correlation, reciprocal space sampling, phonon anharmonicity, and nuclear quantum effects (NQE), among others. In this work, we attempt first-principles equation of state prediction for the crystalline materials ScF3 and CaZrF6, which are known to exhibit negative thermal expansion (NTE) over a broad temperature range. We develop neural network (NN) potentials for both ScF3 and CaZrF6 trained to extensive DFT data, and conduct direct molecular dynamics prediction of the equation(s) of state over a broad temperature/pressure range. The NN potentials serve as surrogates of the DFT Hamiltonian with enhanced computational efficiency allowing for simulations with larger supercells and inclusion of NQE utilizing path integral approaches. The conclusion of the study is mixed: while some equation of state behavior is predicted in semiquantitative agreement with experiment, the pressure-induced softening phenomenon observed for ScF3 is not captured in our simulations. We show that NQE have a moderate effect on NTE at low temperature but does not significantly contribute to equation of state predictions at increasing temperature. Overall, while the NN potentials are valuable for property prediction of these NTE (and related) materials, we infer that a higher level of electron correlation, beyond the generalized gradient approximation density functional employed here, is necessary for achieving quantitative agreement with experiment.

Modeling palladium surfaces with density functional theory, neural networks and molecular dynamics

Phonon thermal transport is a key feature for the operation of thermoelectric materials, but it is challenging to accurately calculate the thermal conductivity of materials with strong anharmonicity or large cells. In this work, a deep neural network potential (NNP) is developed using a dataset based on density functional theory (DFT) and applied to describe the lattice dynamics of Sb2Te3 and Bi2Te3/Sb2Te3 superlattices. The lattice thermal conductivities of Sb2Te3 are first predicted using equilibrium molecular dynamics (EMD) simulations combined with an NNP and the results match well with experimental values. Then, through further exploration of weighted phase spaces and the Grüneisen parameter, we find that there is a stronger anharmonicity in the out-of-plane direction in Sb2Te3, which is the reason why the thermal conductivities are overestimated more in the out-of-plane direction than in the in-plane direction by solving the phonon Boltzmann transport equation (BTE) with only three-phonon scattering processes being considered. More importantly, the lattice thermal conductivities of Bi2Te3/Sb2Te3 superlattices with different periods are accurately predicted using non-equilibrium molecular dynamics (NEMD) simulations together with an NNP, which serves as a good example to explore the thermal transport physics of superlattices using a deep neural network potential.

Machine learned potentials based on artificial neural networks are becoming a popular tool to define an effective energy model for complex systems, either incorporating electronic structure effects at the atomistic resolution, or effectively renormalizing part of the atomistic degrees of freedom at a coarse-grained resolution. One main criticism regarding neural network potentials is that their inferred energy is less interpretable than in traditional approaches, which use simpler and more transparent functional forms. Here we address this problem by extending tools recently proposed in the nascent field of explainable artificial intelligence to coarse-grained potentials based on graph neural networks. With these tools, neural network potentials can be practically decomposed into n-body interactions, providing a human understandable interpretation without compromising predictive power. We demonstrate the approach on three different coarse-grained systems including two fluids (methane and water) and the protein NTL9. The obtained interpretations suggest that well-trained neural network potentials learn physical interactions, which are consistent with fundamental principles.

Neural network potentials trained on quantum-mechanical data can calculate molecular interactions with relatively high speed and accuracy. However, not all neural network potentials are suitable for molecular simulations, as they might exhibit instabilities, nonphysical behavior, or lack accuracy. To assess the reliability of neural network potentials, a series of tests is conducted during model training, in the gas phase, and in the condensed phase. The testing procedure is performed for eight in-house neural network potentials based on the ANI-2x data set, using both the ANI-2x and MACE architectures. This consistent framework allows an evaluation of the effect of the model architecture on its performance. For comparison, we also perform stability tests of the publicly available neural network potentials: ANI-2x, ANI-1ccx, MACE-OFF23, and AIMNet2. The results show that the different models have different weaknesses. A normal-mode analysis of 14 simple benchmark molecules with large displacements from the energy minima revealed that the published MACE-OFF23-S model shows large deviations from the reference quantum-mechanical energy surface. Also, some MACE models with a reduced number of parameters failed to produce stable molecular dynamics simulations in the gas phase, and all MACE models exhibit unfavorable behavior during steric clashes. In addition, the published ANI-2x and one of the in-house MACE models are not able to reproduce the structure of liquid water at ambient conditions, forming an amorphous solid phase instead. For the ANI-1ccx model, the multibody interactions in the condensed water phase lead to nonphysical additional energy minima in bond length and bond angle space, which caused a phase transition to an amorphous solid. Out of all 13 considered public and in-house models, only one in-house model based on the ANI-2x B97-3c data set shows better agreement with the experimental radial distribution function of water than the simple molecular mechanics TIP3P and OPC models. Protein-ligand interaction energies for the four benchmark systems TYK2, CDK2, JNK1, and P38 show that almost all models exhibit a higher correlation with experimental binding affinities than the Chemgauss4 docking score (average R2 > 0.16). With an average R2 of 0.43, the ANI-2x model outperforms molecular mechanics calculations with the GAFF2 force field and DFTB3 semiempirical calculations (average R2 of 0.39 and 0.38), approaching the accuracy of absolute binding free energy calculations (average R2 of 0.52). However, the rather mixed results for the different machine learning potentials show that great care must be taken during model training and when selecting a neural network potential for real-world applications.

Modeling the effects of helium-vacancy clusters on the stress-strain response of a grain boundary in iron by a mechanistic finite element approach informed by molecular dynamics data

Transition metal dichalcogenides are promising materials for high-performance electronics, whereas the impact of defects on their electronic properties remains elusive. Here, we employ neural network potentials (NNPs) constructed from density functional theory (DFT) data to investigate defect-laden WS2 thin films. Molecular dynamics simulations reveal that at low defect concentrations (S/W ratio of 1.9), single sulfur vacancies are predominant. Conversely, at high defect concentrations (S/W ratio of 1.7), complex defects with short lifetimes appear. Additionally, DFT results indicate that the band gap persists at S/W = 1.9 but disappears at 1.7, aligning with observed device degradation at high defect concentrations.

We developed a high-dimensional neural network potential (NNP) to describe the structural and energetic properties of borophene deposited on silver. This NNP has the accuracy of density functional theory (DFT) calculations while achieving computational speedups of several orders of magnitude, allowing the study of extensive structures that may reveal intriguing moiré patterns or surface corrugations. We describe an efficient approach to constructing the training data set using an iterative technique known as the "adaptive learning approach". The developed NNP is able to produce, with excellent agreement, the structure, energy, and forces obtained at the DFT level. Finally, the calculated stability of various borophene polymorphs, including those not initially included in the training data set, shows better stabilization for ν ∼ 0.1 hole density, and in particular for the allotrope α . The stability of borophene on the metal surface is shown to depend on its orientation, implying structural corrugation patterns that can be observed only from long-time simulations on extended systems. The NNP also demonstrates its ability to simulate vibrational densities of states and produce realistic structures with simulated STM images closely matching the experimental ones.

Neural network interatomic potentials (NNPs) have recently proven to be powerful tools to accurately model complex molecular systems while bypassing the high numerical cost of abinitio molecular dynamics simulations. In recent years, numerous advances in model architectures as well as the development of hybrid models combining machine-learning (ML) with more traditional, physically motivated, force-field interactions have considerably increased the design space of ML potentials. In this paper, we present FeNNol, a new library for building, training, and running force-field-enhanced neural network potentials. It provides a flexible and modular system for building hybrid models, allowing us to easily combine state-of-the-art embeddings with ML-parameterized physical interaction terms without the need for explicit programming. Furthermore, FeNNol leverages the automatic differentiation and just-in-time compilation features of the Jax Python library to enable fast evaluation of NNPs, shrinking the performance gap between ML potentials and standard force-fields. This is demonstrated with the popular ANI-2x model reaching simulation speeds nearly on par with the AMOEBA polarizable force-field on commodity GPUs (graphics processing units). We hope that FeNNol will facilitate the development and application of new hybrid NNP architectures for a wide range of molecular simulation problems.

Molecular simulations of high energetic materials (HEMs) are limited by efficiency and accuracy. Recently, neural network potential (NNP) models have achieved molecular simulations of millions of atoms while maintaining the accuracy of density functional theory (DFT) levels. Herein, an NNP model covering typical HEMs containing C, H, N, and O elements is developed. The mechanical and decomposition properties of 1,3,5-trinitroperhydro-1,3,5-triazine (RDX), hexahydro-1,3,5-trinitro-1,3,5-triazine (HMX), and 2,4,6,8,10,12-hexanitrohexaazaisowurtzitane (CL-20) are determined by employing the molecular dynamics (MD) simulations based on the NNP model. The calculated results show that the mechanical properties of α-RDX, β-HMX, and ε-CL-20 agree with previous experiments and theoretical results, including cell parameters, equations of state, and elastic constants. In the thermal decomposition simulations, it is also found that the initial decomposition reactions of the three crystals are N-NO2 homolysis, corresponding radical intermediates formation, and NO2-induced reactions. This decomposition trajectory is mainly divided into two stages separating from the peak of NO2: pyrolysis and oxidation. Overall, the NNP model for C/H/N/O elements in this work is an alternative reactive force field for RDX, HMX, and CL-20 HEMs, and it opens up new potential for future kinetic study of nitramine explosives.

Environmental monitoring is increasingly reliant on image recognition through neural networks, automating the identification and classification of environmental threats. This automation optimizes resource allocation, enhances monitoring accuracy, and empowers researchers to address climate-related challenges. In response to this, we present a web application built using React.js and TensorFlow.js, which enables real-time image recognition. This application facilitates tasks ranging from climate analysis to disaster response and wildlife conservation, offering valuable insights into environmental trends. Moreover, it engages the public, promotes awareness, and contributes to conservation efforts, making it a versatile and cost-effective tool for environmental monitoring. The application leverages advancements in image recognition technology, availability of datasets, open-source frameworks, and cloud computing infrastructure. It supports a wide range of applications, promotes global collaboration, and ensures data security. While challenges exist, the potential of neural networks in environmental monitoring is promising. The web application combines the power of React.js and artificial neural networks, enhancing environmental monitoring efficacy and fostering sustainable practices on a global scale. Ongoing research and standardization efforts are essential for refining and expanding this approach.

The atomic-scale response of inhomogeneous fluids at interfaces and surrounding solute particles plays a critical role in governing chemical, electrochemical, and biological processes. Classical molecular dynamics simulations have been applied extensively to simulate the response of fluids to inhomogeneities directly, but are limited by the accuracy of the underlying interatomic potentials. Here, we use neural network potentials (NNPs) trained to ab initio simulations to accurately predict the inhomogeneous responses of two distinct fluids: liquid water and molten NaCl. Although NNPs can be readily trained to model complex bulk systems across a range of state points, we show that to appropriately model a fluid's response at an interface, relevant inhomogeneous configurations must be included in the training data. In order to sufficiently sample appropriate configurations of such inhomogeneous fluids, we develop protocols based on molecular dynamics simulations in the presence of external potentials. We demonstrate that NNPs trained on inhomogeneous fluid configurations can more accurately predict several key properties of fluids-including the density response, surface tension and size-dependent cavitation free energies-for liquid water and molten NaCl, compared to both empirical interatomic potentials and NNPs that are not trained on such inhomogeneous configurations. This work therefore provides a first demonstration and framework to extract the response of inhomogeneous fluids from first principles for classical density-functional treatment of fluids free from empirical potentials.

Neural networks offer an unbiased and numerically very accurate approach to represent high-dimensional ab initio potential-energy surfaces. Once constructed, neural network potentials can provide the energies and forces many orders of magnitude faster than electronic structure calculations, and thus enable molecular dynamics simulations of large systems. However, Cartesian coordinates are not a good choice to represent the atomic positions, and a transformation to symmetry functions is required. Using simple benchmark systems, the properties of several types of symmetry functions suitable for the construction of high-dimensional neural network potential-energy surfaces are discussed in detail. The symmetry functions are general and can be applied to all types of systems such as molecules, crystalline and amorphous solids, and liquids.

Secondary structures of amino acid sequences can be predicted with over 70% accuracy in Heidelberg with the aid of artificial neural networks. This improvement over the accuracy of statistical methods is extremely important in view of the rational design of peptides and proteins and the processing of data in sequence data banks. The potential of neural networks is thus demonstrated once again (see also the review “Neural Networks in Chemistry” by J. Gasteiger and J. Zupan in the April issue of Angewandte Chemie).