A practical guide to machine learning interatomic potentials – Status and future

Machine learning (ML) interatomic potentials (ML-IAPs) are generated for alkane and polyene hydrocarbons using on-the-fly adaptive sampling and a sparse Gaussian process regression (SGPR) algorithm. The ML model is generated based on the PBE+D3 level of density functional theory (DFT) with molecular dynamics (MD) for small alkane and polyene molecules. Intermolecular interactions are also trained with clusters and condensed phases of small molecules. It shows excellent transferability to long alkanes and closely describes the ab inito potential energy surface for polyenes. Simulation of liquid ethane also shows reasonable agreement with experimental reports. This is a promising initiative toward a universal ab initio quality force-field for hydrocarbons and organic molecules.

Recent developments in machine learning interatomic potentials (MLIPs) have empowered even nonexperts in machine learning to train MLIPs for accelerating materials simulations. However, reproducibility and independent evaluation of presented MLIP results is hindered by a lack of clear standards in current literature. In this Perspective, we aim to provide guidance on best practices for documenting MLIP use while walking the reader through the development and deployment of MLIPs including hardware and software requirements, generating training data, training models, validating predictions, and MLIP inference. We also suggest useful plotting practices and analyses to validate and boost confidence in the deployed models. Finally, we provide a step-by-step checklist for practitioners to use directly before publication to standardize the information to be reported. Altogether, we hope that our work will encourage the reliable and reproducible use of these MLIPs, which will accelerate their ability to make a positive impact in various disciplines including materials science, chemistry, and biology, among others.

The integration ofmachine learning into reactive chemistry, materialsdiscovery, and drug design is transforming the development of novelmolecules and materials. Machine Learning Interatomic Potentials (MLIPs)predict potential energies and forces with quantum chemistry accuracy,surpassing traditional approaches. Incorporating force fitting inMLIP training enhances potential-energy surface predictions and improvesmodel transferability and reliability. This paper introduces and evaluatesthe integration of Hessian matrix training in MLIPs, which encodessecond-order information about the PES curvature. Our evaluation focuseson models trained only to equilibrium geometries and first-order saddlepoints (i.e., critical points on the potential surface), demonstratingtheir ability to extrapolate to nonequilibrium geometries. This integrationimproves extrapolation capabilities, allowing MLIPs to accuratelypredict energies, forces, and Hessian predictions for nonequilibriumgeometries. Hessian-trained MLIPs enhance reaction pathway modeling,transition state identification, and vibrational spectra predictions,benefiting molecular dynamics (MD) simulations and Nudged ElasticBand (NEB) calculations. By analyzing models trained with varyingcombinations of energy, force, and Hessian data on a small moleculereactive data set, we demonstrate that models including Hessian informationnot only extrapolate more accurately to unseen molecular systems,improving accuracy in reaction modeling and vibrational analysis,but also reduce the total amount of data required for effective training.However, the primary trade-off is increased computational expense,as Hessian training requires more resources than conventional energy-forcetraining. Our findings provide comprehensive insights into the advantagesand limitations of Hessian integration in MLIP training, allowingpractitioners in computational chemistry to make informed decisionsabout employing this method in accordance with their specific researchobjectives and computational constraints.

Dopants can tune the performance of MoS2 in various applications, but the use of molecular dynamics simulations for doped MoS2 materials discovery is limited by the lack of multidopant interatomic potentials. Universal machine learning interatomic potentials (MLIPs) could be a solution, but the accuracy of these potentials must first be evaluated. Here, we evaluate the accuracy of a recently developed MLIP, META's Universal Model for Atoms, for 25 different MoS2 dopants spanning metals, nonmetals, and transition metals in Mo-substitution, S-substitution, and intercalated positions by benchmarking the MLIP-predicted formation energy and the dopant-induced structural change against density functional theory (DFT) calculations. The computational framework for MLIP validation and simulations is described in detail, and the source code is made available online. The MLIP is then demonstrated by performing heating-cooling simulations of MoS2 supercells with all 25 dopants. These simulations capture complex phenomena including dopant clustering, MoS2 layer fracturing, interlayer diffusion, and chemical compound formation at orders-of-magnitude reduced computational cost compared to DFT. This work provides a computational workflow for the application-oriented design of doped-MoS2, enabling high-throughput screening of dopant candidates and optimization of compositions for targeted tribological, electronic, and optoelectronic performance.

<title>Abstract</title> Machine-Learning Interatomic Potentials (MLIPs) have surged in popularity due to their promise of expanding the spatiotemporal scales possible for simulating molecules with high fidelity. The accuracy of any MLIP is dependent on the data used for its training; thus, for large molecules, like polymers, where accurate training data is prohibitively difficult to obtain, it becomes necessary to pursue non-traditional methods to construct MLIPs, many of which are based on constructing MLIPs using smaller, analogous chemical systems. However, we have yet to understand the limits to which smaller molecules can be used as a proxy for extrapolating macromolecular energetics. Here, we provide a ''control study'' for such experiments, exploring the ability of MLIP approaches to extrapolate between n=1-8 n-polyalkanes at identical conditions. Through Principal Covariates Classification, we quantitatively demonstrate how convergence in chemical environments between training and testing datasets coincides with an MLIP's transferability. Additionally, we show how careful attention to the construction of an MLIP's neighbor list can promote greater transferability when considering various levels of the energetic hierarchy. Our results establish a roadmap for how one can create transferable MLIPs for macromolecular systems without the prohibitive cost of constructing system-specific training data.

Efficiently creating a concise but comprehensive data set for training machine-learned interatomic potentials (MLIPs) is an under-explored problem. Active learning, which uses biased or unbiased molecular dynamics (MD) to generate candidate pools, aims to address this objective. Existing biased and unbiased MD-simulation methods, however, are prone to miss either rare events or extrapolative regions—areas of the configurational space where unreliable predictions are made. This work demonstrates that MD, when biased by the MLIP’s energy uncertainty, simultaneously captures extrapolative regions and rare events, which is crucial for developing uniformly accurate MLIPs. Furthermore, exploiting automatic differentiation, we enhance bias-forces-driven MD with the concept of bias stress. We employ calibrated gradient-based uncertainties to yield MLIPs with similar or, sometimes, better accuracy than ensemble-based methods at a lower computational cost. Finally, we apply uncertainty-biased MD to alanine dipeptide and MIL-53(Al), generating MLIPs that represent both configurational spaces more accurately than models trained with conventional MD.

Machine learning interatomic potentials (MLIPs) enable accurate simulations of materials at scales beyond those accessible to conventional first-principles methods.1 Specifically, for solid electrolyte materials, MLIPs have bridged the gap between ionic conductivities simulated by ab initio molecular dynamics (AIMD) at room temperature and experimental measurements.2 MLIPs also extend the accuracy of AIMD to systems too computationally demanding for AIMD alone, including grain boundaries, amorphous structures, interfaces, and structures with varying short-range order.3 Recently, the emergence of universal MLIPs has opened opportunities for discovering new materials at unprecedented scales.4 In this talk, I will introduce MatPES, a fundamental dataset constructed to enhance the stability and accuracy of universal MLIPs in capturing both thermodynamic and kinetic properties.5 With benchmark results, I will also attempt to answer the question posed in the title of my talk. Unke, O. T. et al. Machine Learning Force Fields. Chem. Rev. 121, 10142–10186 (2021).Qi, J. et al. Bridging the gap between simulated and experimental ionic conductivities in lithium superionic conductors. Mater. Today Phys. 21, 100463 (2021).Lee, T. et al. Atomic-scale origin of the low grain-boundary resistance in perovskite solid electrolyte Li0.375Sr0.4375Ta0.75Zr0.25O3. Nat. Commun. 14, 1940 (2023).Jacobs, R. et al. A practical guide to machine learning interatomic potentials – Status and future. Curr. Opin. Solid State Mater. Sci. 35, 101214 (2025).Kaplan, A. D. et al. A Foundational Potential Energy Surface Dataset for Materials. Preprint at https://doi.org/10.48550/arXiv.2503.04070 (2025).

In nuclear and/or solar applications that involve molten salts, impurities frequently enter the salt as either fission products or via corrosion. Impurities can interact and make complexes, but the impact of such complexation on the properties of the salts and corrosion rates has not been understood. Common impurities in molten salts, such as FLiBe, include Cr, Ni, and Fe. Here, we investigate the complexation of Cr and Ni in FLiBe using molecular dynamics based on a machine learning interatomic potential (MLIP) fitted using the atomic cluster expansion (ACE) method. The MLIP allows us to overcome the challenges of simultaneously needing accurate energetics and long time scale to study complexation. We demonstrate that impurity behavior is more difficult to capture than that of concentrated elements with MLIPs due to less sampling in training data, but that this can be overcome by using active learning strategies to obtain a robust fit. Our findings suggest that there is a weak but potentially significant binding free energy between CrF2 and NiF2 in eutectic FLiBe of -0.112eV. Under certain conditions this binding creates a significant short-range order between the species and lowers the redox potential of NiF2 in the presence of CrF2 in FLiBe, making Ni dissolution more favorable in the presence of Cr as compared to its dissolution in pure FLiBe. However, we find little impact of this complexation on the diffusivity of Ni and Cr. Overall, the methodology presented here suggests an approach to modeling complexation with MLIPs and suggests that interactions between dissolved cations could be playing a significant role in some salt thermophysical properties.

Training machine learning interatomic potentials (MLIPs) on total energies of molecular clusters using differential or transfer learning is becoming a popular route to extend the accuracy of correlated wave-function theory to condensed phases. A key challenge, however, lies in validation, as reference observables in finite-temperature ensembles are not available at the reference level. Here, we construct synthetic reference data from pretrained MLIPs and evaluate the generalizability of cluster-trained models on ice-Ih, considering scenarios where both energies and forces and where only energies are available for training. We study the accuracy and data-efficiency of differential, single-fidelity transfer, and multi-fidelity transfer learning against ground-truth thermodynamic observables. We find that transferring accuracy from clusters to bulk requires regularization, which is best achieved through multi-fidelity transfer learning when training on both energies and forces. By contrast, training only on energies introduces artefacts: stable trajectories and low energy errors conceal large force errors, leading to inaccurate microscopic observables. More broadly, we show that accurate reproduction of microscopic structure correlates strongly with low force errors but only weakly with energy errors, whereas global properties such as energies and densities correlate with low energy errors. This&amp;#xD;highlights the need to incorporate forces during training or to apply careful validation before production. Our results highlight the promise and pitfalls of cluster-trained MLIPs for condensed phases and provide guidelines for developing - and critically, validating - robust and data-efficient MLIPs.

Low-dimensional hybrid organic-inorganic perovskites (HOIPs) are promising electronically active materials for light absorption and emission. The design space of HOIPs is extremely large, as a variety of organic cations can be combined with different inorganic frameworks. This not only allows for tunable electronic and mechanical properties but also necessitates the development of new tools for in silico high throughput analysis of candidate materials. In this work, we present an accurate, efficient, and widely applicable machine learning interatomic potential (MLIP) trained on 86 diverse experimentally reported HOIP materials. This MLIP was tested on 73 experimentally reported perovskite compositions and achieves a high accuracy, relative to density functional theory (DFT). We also introduce a novel random structure search algorithm designed for the crystal structure prediction of 2D HOIPs. The combination of MLIP and the structure search algorithm reliably recovers the crystal structure of 14 known 2D perovskites by specifying only the organic molecule and inorganic cation/halide. Performing this crystal structure search with ab initio methods would be computationally prohibitive but is relatively inexpensive with the MLIP. Finally, the developed procedure is used to predict the structure of a totally new HOIP with cation (cis-1,3-cyclohexanediamine). Subsequently, the new compound was synthesized and characterized, which matches the predicted structure, confirming the accuracy of our method. This capability will enable the efficient and accurate screening of thousands of combinations of organic cations and inorganic layers for further investigation.

Lithium-based disordered rocksalts (LDRs), which are an important class of positive electrode materials that can increase the energy density of current Li-ion batteries, represent a significantly complex chemical and configurational space for conventional density functional theory (DFT)-based high-throughput screening approaches. Notably, atom-centered machine-learned interatomic potentials (MLIPs) are a promising pathway to accurately model the potential energy surface of highly disordered chemical spaces, such as LDRs, where the performance of such MLIPs has not been rigorously explored yet. Here, we represent a comprehensive evaluation of the accuracy, transferability, and ease of training of five atom-centered MLIPs, including the artificial neural network potentials developed by the atomic energy network (AENET), the Gaussian approximation potential (GAP), the spectral neighbor analysis potential (SNAP) and its quadratic extension (qSNAP), and the moment tensor potential (MTP), in modeling a 11-component LDR chemical space. Specifically, we generate a DFT-calculated data set of 10,842 configurations of disordered LiTMO2 and TMO2 compositions, where TM = Sc, Ti, V, Cr, Mn, Fe, Co, Ni, and/or Cu. To provide a point-of-comparison on the performance of atom-centered MLIPs, we also trained the neural equivariant interatomic potential (NequIP) on a subset of our data. Importantly, we find AENET to be the best potential in terms of accuracy and transferability for energy predictions, while MTP is the best for atomic forces. While AENET is the fastest to train among the MLIPs considered at low number of epochs (300), the training time increases significantly as epochs increase (3300), with a corresponding reduction in training errors (∼60%). Note that AENET and GAP tend to overfit in small data sets, with the extent of overfitting reducing with larger data sets. Finally, we observe AENET to provide reasonable predictions of average Li-intercalation voltages in layered, single-TM LiTMO2 frameworks, compared to DFT (∼10% error on average). Our study should pave the way both for discovering novel disordered rocksalt electrodes and for modeling other configurationally complex systems, such as high-entropy ceramics and alloys.

The accurate description of phonon dispersion of two-dimensional (2D) materials demonstrates significance in many research fields of condensed matter physics. In this paper, we systematically calculate the phonon spectra and transport properties of six representative 2D materials (encompassing single-element and binary compounds with flat, buckled, and puckered backbone geometries) by means of density functional theory (DFT) and two machine learning interatomic potentials [MLIPs, on-the-fly machine learning potential (FMLP), and moment tensor potential (MTP)]. The results show that the acoustic out-of-plane flexural (ZA) dispersion of the 2D materials are always and easily exhibiting non-quadratic dispersion phenomena near the center of the Brillouin zone by using the pure DFT calculation method. This phenomenon contradicts physics and reflects intuitively from the non-zero group velocity at Γ point. However, no matter which MLIP (FMLP/MTP) the calculation is based on, it could solve such behavior perfectly, where the ZA mode conforms to the quadratic dispersion relationship in the long-wavelength limit. Our results further demonstrate that compared to the pure DFT calculation, the FMLP and MTP method could quickly and relatively accurately obtain the lattice thermal conductivities of graphene, silicene, phosphorene, SiC, MoS2, and GeS. The findings presented in this work provide a solution about the pseudophysical phenomenon of ZA dispersions in 2D materials with the pure DFT calculation, which will greatly facilitate research areas such as phonon thermal transport, flexural mechanics, and electron–acoustic coupling.

This work develops a hybrid machine learning/molecular mechanics (ML/MM) interface integrated into the AMBER molecular simulation package. The resulting platform is highly versatile, accommodating several advanced machine learning interatomic potential (MLIP) models while providing stable simulation capabilities and supporting high-performance computations. Building upon this robust foundation, we developed new computational protocols to enable pathway-based and end point-based free energy calculation methods utilizing ML/MM hybrid potential. In particular, we proposed an ML/MM-compatible thermodynamic integration (TI) framework that adequately addressed the challenge of applying MLIPs in TI calculations due to its indivisible nature of energy and force. Our results demonstrated that the hydration free energies calculated using this framework achieved an accuracy of 1.0 kcal/mol, outperforming the traditional approaches. Moreover, ML/MM enables more precise sampling of conformational ensembles for improved end point-based free energy calculations. Overall, our efficient, stable, and highly compatible interface not only broadens the application of MLIPs in multiscale simulations but also enhances the accuracy of free energy calculations from multiple aspects. By introducing a novel ML/MM-compatible thermodynamic integration framework, we offered a novel foundation for combining advanced multiscale simulation methodologies with highly accurate free energy calculation techniques, thereby opening new avenues and providing a robust theoretical framework for future developments in this field.

A combination of machine learned interatomic potentials (MLIPs) and enhanced sampling simulations is used to investigate the activation of methane on a Ni(111) surface. The work entails the development and iterative refinement of MLIPs, initially trained on a data set constructed via ab initio molecular dynamics simulations, supplemented by adaptive biasing forces, to enrich the sampling of catalytically relevant configurations. Our results reveal that upon incorporation of collective variables that capture the behavior of the reactant molecule, as well as additional frames that describe the dynamic response of the catalytic surface, it is possible to enhance considerably the accuracy of predicted energies and forces. By employing enhanced sampling schemes in the refinement of the MLIP, we systematically explore the potential energy surface, leading to a refined MLIP capable of predicting density functional theory-level energies and forces and replicating key geometric characteristics of the catalytic system. The resulting free energy landscapes at several temperatures provide a detailed view of the thermodynamics and dynamics of methane activation. Specifically, as methane approaches and dissociates on the catalytic surface, the process involves the dynamic interplay of CH4 and the Ni catalyst that includes both enthalpic and entropic contributions. The progression toward the transition state involves a CH4 moiety that is increasingly restrained in its ability to rotate or translate, while the stage following the transition state is characterized by a notable rise of the Ni atom that interacts with the cleaved C-H bond. This leads to an increase in the mobility of the adsorbed species, a feature that becomes more pronounced at higher temperatures.

While DFT is the computational method of choice for mechanistic insight in homogeneous catalysis, the recent rise of foundation-level machine learning interatomic potentials (MLIPs) invites reconsideration: are we approaching competition, or a deeper synergy? These pretrained, fast surrogates are able to map reaction space, sample conformers, and flag likely transition states, potentially displacing routine low-level DFT. Yet their reliability hinges on calibrated uncertainty, transferability across ligand and oxidation-state manifolds, and faithful treatment of long-range polarization, solvation, and open-shell or multireference character. We argue that the near future will likely be contested: MLIPs will handle everyday exploratory tasks, while DFT and higher-level methods will anchor electronic effects, validate high-stakes predictions, and resolve edge cases. If supported by FAIR catalysis datasets, standardized workflows, and robust error quantification, the two approaches will coevolve, enabling scalable, predictive discovery without sacrificing rigor or interpretability.