AENET-LAMMPS and AENET-TINKER: Interfaces for accurate and efficient molecular dynamics simulations with machine learning potentials.

Solving electronic structure problems represents a promising field of applications for quantum computers. Currently, much effort is spent in devising and optimizing quantum algorithms for near-term quantum processors, with the aim of outperforming classical counterparts on selected problem instances using limited quantum resources. These methods are still expected to feature a runtime preventing quantum simulations of large scale and bulk systems. In this work, we propose a strategy to extend the scope of quantum computational methods to large scale simulations using a machine learning potential trained on quantum simulation data. The challenge of applying machine learning potentials in today’s quantum setting arises from the several sources of noise affecting the quantum computations of electronic energies and forces. We investigate the trainability of a machine learning potential selecting various sources of noise: statistical, optimization, and hardware noise. Finally, we construct the first machine learning potential from data computed on actual IBM Quantum processors for a hydrogen molecule. This already would allow us to perform arbitrarily long and stable molecular dynamics simulations, outperforming all current quantum approaches to molecular dynamics and structure optimization.

Accurate interatomic force field for molecular dynamics simulation by hybridizing classical and machine learning potentials

Machine learning potentials have revolutionised the field of atomistic simulations in recent years and are becoming a mainstay in the toolbox of computational scientists. This paper aims to provide an overview and introduction into machine learning potentials and their practical application to scientific problems. We provide a systematic guide for developing machine learning potentials, reviewing chemical descriptors, regression models, data generation and validation approaches. We begin with an emphasis on the earlier generation of models, such as high-dimensional neural network potentials and Gaussian approximation potentials, to provide historical perspective and guide the reader towards the understanding of recent developments, which are discussed in detail thereafter. Furthermore, we refer to relevant expert reviews, open-source software, and practical examples-further lowering the barrier to exploring these methods. The paper ends with selected showcase examples, highlighting the capabilities of machine learning potentials and how they can be applied to push the boundaries in atomistic simulations.

In this study, we trained a deep potential (DP) for H2O, an accurate machine learning (ML) potential. We performed molecular dynamics (MD) simulations of liquid water using the DP model (or DeePMD simulations). Our results showed that the DP model exhibits DFT-level accuracy, and the DeePMD simulation is a promising approach for modeling the structural properties of liquid water. Based on the DeePMD simulation trajectories, we calculated the isotropic Raman spectra of the O-H stretching mode using the surface-specific velocity-velocity correlation function (ssVVCF), showing that the DeePMD/ssVVCF approach can correctly capture the bimodal characteristics of the experimental Raman spectra, with one peak located near 3400 cm-1 and the other near 3250 cm-1. The success of the DeePMD/ssVVCF approach should be credited to (1) the DFT-level accuracy of the DP model for H2O, (2) the ssVVCF formulation considering the coupling between vibrational modes, and (3) non-Condon effects. Furthermore, the DeePMD simulations revealed that the anharmonic interactions between the coupled water molecules in the first and second hydration shells should play an essential role in the strong mixing of the H-O-H bending mode and the O-H stretching mode, leading to the delocalization of the O-H stretching band. In particular, increasing the strength of hydrogen bonds would enhance the bend-stretch coupling, leading to the red-shifting of the O-H vibrational spectra and the increase in the intensity of the shoulder around 3250 cm-1.

Sampling reference data is crucial in machine learning potential (MLP) construction. Inadequate coverage of local configurations in reference data may lead to unphysical behaviors in MLP-based molecular dynamics (MLP-MD) simulations. To address this problem, this study proposes a new on-the-fly reference data sampling method called radial distribution function (RDF)-based data sampling for MLP construction. This method detects and extracts anomalous structures from the trajectories of MLP-MD simulations by focusing on the shapes of RDFs. The detected structures are added to the reference data to improve the accuracy of the MLP. This method allows us to realize a reasonable MLP construction for liquid water with minimal additional data. We prepare data from an H2O molecular cluster system and verify whether the constructed MLPs are practical for bulk water systems. MLP-MD simulations without RDF-based data sampling show unphysical behaviors, such as atomic collisions. In contrast, after applying this method, we obtain MLP-MD trajectories with features, such as RDF shapes and angle distributions, that are comparable to those of ab initio MD simulations. Our simulation results demonstrate that the RDF-based data sampling approach is useful for constructing MLPs that are robust to extrapolations from molecular cluster systems to bulk systems without any specialized know-how.

Machine learning potentials (MLPs) can enable simulations of ab initio accuracy at orders of magnitude lower computational cost. However, their effectiveness hinges on the availability of considerable data sets to ensure robust generalization across chemical space and thermodynamic conditions. The generation of such data sets can be labor-intensive, highlighting the need for innovative methods to train MLPs in data-scarce scenarios. Here, we introduce transfer learning of potential energy surfaces between chemically similar elements. Specifically, we leverage the trained MLP for silicon to initialize and expedite the training of an MLP for germanium. Utilizing classical force field and ab initio data sets, we demonstrate that transfer learning surpasses traditional training from scratch in force prediction, leading to more stable simulations and improved temperature transferability. These advantages become even more pronounced as the training data set size decreases. We also observe positive transfer learning effects for most out-of-target properties. Our findings demonstrate that transfer learning across chemical elements is a promising technique for developing accurate and numerically stable MLPs, particularly in a data-scarce regime.

A key component of research in the aerospace industry constitutes hypersonic flights (M>5) which includes the design of commercial high-speed aircrafts and development of rockets. Computational analysis becomes more important due to the difficulty in performing experiments and reliability of its results at these harsh operating conditions. There is an increasing demand from the industry for the accurate prediction of wall-shear and heat transfer with a low computational cost. Direct Numerical Simulations (DNS) create the standard for accuracy, but its practical usage is difficult and limited because of its high cost of computation. The usage of Reynold's Averaged Navier Stokes (RANS) simulations provide an affordable gateway for industry to capitalize its lower computational time for practical applications. However, the presence of existing RANS turbulence closure models and associated wall functions result in poor prediction of wall fluxes and inaccurate solutions in comparison with high fidelity DNS data. In recent years, machine learning emerged as a new approach for physical modeling. This thesis explores the potential of employing Machine Learning (ML) to improve the predictions of wall fluxes for hypersonic turbulent boundary layers. Fine-grid RANS simulations are used as training data to construct a suitable machine learning model to improve the solutions and predictions of wall quantities for coarser meshes. This strategy eliminates the usage of wall models and extends the range of applicability of grid sizes without a significant drop in accuracy of solutions. Random forest methodology coupled with a bagged aggregation algorithm helps in modeling a correction factor for the velocity gradient at the first grid points. The training data set for the ML model extracted from fine-grid RANS, includes neighbor cell information to address the memory effect of turbulence, and an optimal set of parameters to model the gradient correction factor. The successful demonstration of accurate predictions of wall-shear for coarse grids using this methodology, provides the confidence to build machine learning models to use DNS or high-fidelity modeling results as training data for reduced-order turbulence model development. This paves the way to integrate machine learning with RANS to produce accurate solutions with significantly lesser computational costs for hypersonic boundary layer problems.

Machine-learning potentials (MLPs) have become important tools for modern molecular simulations. However, developing models that simultaneously achieve high accuracy and high computational efficiency remains a significant challenge. In this work, we extend the recently proposed MB-PIPNet framework to general covalently bonded molecular systems by combining monomer-based energy decomposition, permutationally invariant polynomial (PIP) descriptors, and neural networks within a fragmentation-based strategy. Within this framework, the total potential energy is represented as a sum of effective monomeric contributions, where PIPs provide compact and chemically motivated descriptions of both monomer internal structures and their local chemical environments. As a proof-of-concept application, we apply the MB-PIPNet framework to linear alkanes, using C14H30 as a representative system, and benchmark its performance against established atomistic machine-learning models. The resulting MB-PIPNet potential accurately reproduces reference ab initio electronic energies and reliably captures key molecular properties, including torsional potential energy profiles, harmonic vibrational frequencies, and vibrational power spectra obtained from molecular dynamics simulations. Importantly, MB-PIPNet demonstrates a substantial advantage in computational efficiency over other MLP models for combined energy and force evaluations. These results establish MB-PIPNet as a scalable and efficient framework for constructing MLPs, providing an additional route for large-scale quantum and classical simulations of complex molecular systems.

Accurate and efficient simulation of liquids, such as water and salt solutions, using high-level wave function theories is still a formidable task for computational chemists owing to the high computational costs. In this study, we develop a deep machine learning potential based on fragment-based second-order Møller-Plesset perturbation theory (DP-MP2) for water through neural networks. We show that the DP-MP2 potential predicts the structural, dynamical, and thermodynamic properties of liquid water in better agreement with the experimental data than previous studies based on density functional theory (DFT). The nuclear quantum effects (NQEs) on the properties of liquid water are also examined, which are noticeable in affecting the structural and dynamical properties of liquid water under ambient conditions. This work provides a general framework for quantitative predictions of the properties of condensed-phase systems with the accuracy of high-level wave function theory while achieving significant computational savings compared to ab initio simulations.

Lithium borosilicate (LBS) glass is a prototypical lithium-ion conducting oxide glass available for an all-solid-state buttery. Nevertheless, the atomistic modeling of LBS glass using ab initio (AIMD) and classical molecular dynamics (CMD) simulations has critical limitations due to computational cost and inaccuracy in reproducing the glass microstructures, respectively. To overcome these difficulties, a machine-learning potential (MLP) was examined in this work for modeling LBS glasses using DeepMD. The glass structures obtained by this MLP possessed 4-fold coordinated boron (4B) confirmed well with the experimental data and abundance of three-membered rings. The models were energetically more stable compared with those constructed with a functional force field even though both the models included reasonable 4B. The results confirmed MLP to be superior to model the boron-containing glasses and address the inherent shortcomings of the AIMD and CMD. This study also discusses some limitations of MLP for modeling glasses.

Solvent environments play a central role in determining molecular structure, energetics, reactivity, and interfacial phenomena. However, modeling solvation from first principles remains difficult due to the complex interplay of interactions and unfavorable computational scaling of first-principles treatment with system size. Machine-learned potentials (MLPs) have recently emerged as efficient surrogates for quantum chemistry methods, offering first-principles accuracy at greatly reduced computational cost. MLPs approximate the underlying potential energy surface, enabling efficient computation of energies and forces in solvated systems. They are also capable of accounting for effects such as hydrogen bonding, long-range polarization, and conformational changes. This review surveys the development and application of MLPs in solvation modeling. We summarize the theoretical basis of MLP-based energy and force predictions and present a classification of MLPs based on training targets, model types, and design choices related to architectures, descriptors, and training protocols. Integration into established solvation workflows is discussed, with case studies spanning small molecules, interfaces, and reactive systems. We conclude by outlining open challenges and future directions toward transferable, robust, and physically grounded MLPs for solvation-aware atomistic modeling.

<sec><p indent="0mm">Catalysis, the science of accelerating chemical reactions and regulating the selectivity of reaction products, lies in the heart of modern chemical industry. Typically, catalysis can be categorized into heterogeneous catalysis and homogeneous catalysis. Over the past century, heterogeneous catalytic processes such as ammonia synthesis and Fischer-Tropsch synthesis have become an indispensable part of the chemical industry. In recent years, the rational design of catalysts has gained a lot of attention due to both economic and environmental concerns, which has raised new demands for a deep understanding of the mechanism of catalytic reactions. Machine learning methods as one of the most important areas of artificial intelligence (AI) are becoming more and more popular in practices of theoretical simulations. This review gives an overview on various machine learning methods and their applications in theoretical heterogeneous catalysis. </sec><sec> Starting with statistical learning models on small datasets based on specific physical descriptors, which play an important role in early theoretical heterogeneous research, this review analyzes how they connect linear scaling relationship and catalytic experiments. However, due to the complexity of heterogeneous catalytic systems, sometimes nonlinear models can be essential. In this way, the exact form of equations should be obtained before fitting, which is extremely challenging. This problem can be solved by methods based on symbolic regression like SISSO to obtain a set of expressions, and the researchers can select one by their domain knowledge. This review also shows the applications of some popular machine learning methods (such as SVM, PCA, RF, etc.) in research, most of which are based on specific descriptors, simple but have difficulty transferring to other systems. </sec><sec> First-principle calculations like DFT are crucial in catalysis rational design by computational simulations for surface structure determination and reaction channel exploration, but they are impractical for simulations on large and complex catalytic systems due to high computational costs. In recent years, machine learning potentials (MLPs), as tools to bridge the gap between first-principle accuracy and computational efficiency, have pushed forward the frontier of theoretical heterogeneous catalysis. Popular MLPs (LASP, EANN, DP, GemNet-OC, MACE, DPA, etc.) have numerous applications in this field, from global optimization and molecular dynamics of active surfaces under reaction conditions to automatic reaction network exploration. Recently, universal machine learning potentials have gained much attention and are very promising in heterogeneous catalysis simulations, for their promising capability to encapsulate huge chemical space of the whole periodic table in one pre-trained model, which is easy to fine-tune on a specific chemical system, efficiently deploying a specific MLP with high accuracy and transferability. </sec><sec> Recent advances in generative models have shown considerable promise for heterogeneous catalysis research, and the dramatic emergence of large language models like ChatGPT is especially important in this case. LLMs can incorporate information from massive natural language data, having a huge ability to push forward the understanding of topics in catalytic science. Furthermore, the concept of generative models themselves has much potential in generating novel catalytic information directly without natural language as the intermediate, but their application in heterogeneous catalysis is under limitation due to the lack of large datasets with high quality. </sec><sec> In a nutshell, machine learning methods play a key role in theoretical understanding of heterogeneous catalytic systems. Challenges remain due to the complexity of heterogeneous catalytic systems and lack of high-quality dataset, but the era of AI-driven rational design of surface catalysts is truly coming. </sec>

Accurate prediction of the mechanical and thermal properties of CO₂ hydrates is essential for their applications in carbon sequestration and refrigeration, yet remains challenging with empirical forcefields. In this work, a deep potential machine learning potential for CO₂ hydrate, trained on density functional theory datasets, is for the first time developed to serve as a unified and accurate computational framework. The as-developed deep potential machine learning potential achieves near-density functional theory accuracy in energy, force, and virial stress predictions while enabling large-scale molecular dynamics simulations at significantly reduced computational cost. Uniaxial stress-strain analyses demonstrate that the model captures the tensile strength and progressive ductile-like failure behavior. Thermal conductivity prediction agrees closely with experimental measurements within 2% deviation, outperforming empirical forcefields. Vibrational dynamics and phonon analyses reveal that the deep potential machine learning potential more accurately describes the anharmonicity and phonon scattering, especially in high-frequency modes, yielding physically realistic thermal transport behavior. This work establishes deep potential machine learning potential as a robust tool for advancing CO₂ hydrate-based technologies by providing a path for accurate and efficient multi-property prediction. Document Type: Original article Cited as: Xiong, K., Li, Y., Lin, Z., Luo, G., Wu, J. Machine learning potential insights into mechanical response and heat transfer in CO₂ hydrate. Advances in Geo-Energy Research, 2025, 18(1): 38-50. https://doi.org/10.46690/ager.2025.10.04

Purpose: We describe here a novel full-differential Monte Carlo (MC) event-by-event simulation, for modelling electron and positron histories in liquid and gaseous water, with impact energies ranging from the water excitation threshold (7.4 eV) to 10 keV. This new track-structure code is named EPOTRAN, an acronym for Electron and POsitron TRANsport in water. Material and methods: All the processes induced by both electrons and positrons are studied in detail via theoretical differential and total cross sections, calculated within the quantum mechanical framework by using the partial-wave method. Elastic and inelastic interactions are then successively reviewed, including in particular an original description of the positron-induced capture process leading to Positronium formation. Results: Total and differential cross sections are reported and compared with a large set of existing measurements. Rather good agreement is generally observed over the considered energy range. Conclusions: This work reports the theoretical cross sections used in a special purpose Monte Carlo simulation suitable for electron and positron transport in gaseous and liquid water. This MC code should represent an accurate tool for dose calculation at the nanometric scale, by providing a detailed spatial distribution of energy deposits. Furthermore, positron trajectory studies made possible by this approach should prove useful for evaluating the real contribution of the positron range on the overall spatial resolution of PET (Positron Emission Tomography) imaging.

Inorganic and hybrid lead halide perovskites are among the most promising candidates for next-generation optoelectronic devices. However, their further development is critically constrained by stability issues, including phase degradation, migration, and interfacial reactions. Overcoming these challenges requires a detailed understanding of the underlying atomic-scale mechanisms. Density functional theory (DFT) and ab initio molecular dynamics (AIMD) offer essential electronic and structural insights, but their high computational cost limits simulations to relatively small systems and short time scales, making it difficult to capture long-term evolution. Classical MD can access larger length and time scales but often lacks accurate and transferable force fields for lead halide perovskite systems. Machine learning potentials (MLPs) have emerged as an alternative solution to bridge this gap, enabling large-scale, long-time-scale MD simulations with near-DFT accuracy at a fraction of the computational cost. In this perspective, we review recent applications of MLPs to inorganic and hybrid lead halide perovskites, including phase behavior, ion migration, and perovskite interfaces. We further discuss current challenges related to model efficiency and transferability and outline future opportunities for deploying MLPs to tackle outstanding questions in perovskite stability, degradation, and device-relevant operation.