PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ryan Jacobs, Dane Morgan, Siamak Attarian, Jun Meng, Chen Shen, Zhenghao Wu, Clare Yijia Xie, Julia H. Yang, Nongnuch Artrith, Ben Blaiszik, Gerbrand Ceder, Kamal Choudhary, Gabor Csanyi, Ekin Dogus Cubuk, Bowen Deng, Ralf Drautz, Xiang Fu, Jonathan Godwin, Vasant Honavar, Olexandr Isayev, Anders Johansson, Boris Kozinsky, Stefano Martiniani, Shyue Ping Ong, Igor Poltavsky, KJ Schmidt, So Takamoto, Aidan Thompson, Julia Westermayr, Brandon M. Wood

arXiv:2503.09814v1cond-mat.mtrl-scics.LG

TL;DR

Researchers who are not MLIP experts face difficulty navigating the rapidly expanding literature and deciding how to use these tools. This review responds with an accessible practical guide covering MLIP foundations, applications, workflows, selection, infrastructure, limitations, and future directions. It also surveys universal MLIPs and compares practical considerations for using or fitting potentials.

  • Problem

    The rapidly growing MLIP literature makes it difficult for non-experts to determine how to proceed when using these tools.

  • Method

    The paper provides an accessible practical review of MLIP foundations, workflows, universal potentials, execution speed, selection, infrastructure, limitations, and future directions.

  • Results

    The review finds that MTP was highly accurate and very fast, generally performing somewhat better than GAP, SNAP, and Behler-Parrinello NN potentials, while ACE offers flexibility.

  • Takeaways & Limitations

    The paper provides practical guidance for choosing pre-trained or newly fitted MLIPs according to hardware, speed, accuracy, and application requirements.

  • Takeaways & Limitations

    Present MLIPs have limitations including instability during long molecular-dynamics simulations and challenges involving long-range interactions, magnetism, and excited states.

Abstract

from arXiv · show

The rapid development and large body of literature on machine learning interatomic potentials (MLIPs) can make it difficult to know how to proceed for researchers who are not experts but wish to use these tools. The spirit of this review is to help such researchers by serving as a practical, accessible guide to the state-of-the-art in MLIPs. This review paper covers a broad range of topics related to MLIPs, including (i) central aspects of how and why MLIPs are enablers of many exciting advancements in molecular modeling, (ii) the main underpinnings of different types of MLIPs, including their basic structure and formalism, (iii) the potentially transformative impact of universal MLIPs for both organic and inorganic systems, including an overview of the most recent advances, capabilities, downsides, and potential applications of this nascent class of MLIPs, (iv) a practical guide for estimating and understanding the execution speed of MLIPs, including guidance for users based on hardware availability, type of MLIP used, and prospective simulation size and time, (v) a manual for what MLIP a user should choose for a given application by considering hardware resources, speed requirements, energy and force accuracy requirements, as well as guidance for choosing pre-trained potentials or fitting a new potential from scratch, (vi) discussion around MLIP infrastructure, including sources of training data, pre-trained potentials, and hardware resources for training, (vii) summary of some key limitations of present MLIPs and current approaches to mitigate such limitations, including methods of including long-range interactions, handling magnetic systems, and treatment of excited states, and finally (viii) we finish with some more speculative thoughts on what the future holds for the development and application of MLIPs over the next 3-10+ years.

1 Introduction

This review is a practical, accessible guide to MLIPs for technically literate materials scientists and chemists who are not MLIP experts. It surveys their foundations, applications, performance, selection, infrastructure, limitations, and future directions.

  • Motivation: The review addresses the challenge of learning how to use MLIPs from a rapidly developing literature.It targets researchers who want to use MLIPs but are not experts.
  • Purpose and audience: It provides an in-depth practical examination of the current MLIP landscape, complementing shorter, higher-level discussions.The stated audience is technically literate materials scientists and chemists with molecular-modeling backgrounds.
  • Foundations: It compares explicit featurization and implicit graph-neural-network approaches while minimizing deep mathematics and formalism.Examples include atomic cluster expansion and graph neural networks.
  • Scope: The guide explains fundamental MLIP aspects, broad application potential, general frameworks, workflows, speed, accuracy, and supporting infrastructure.Workflows include fitting new potentials and using pre-fitted ones.
  • Universal MLIPs: The review covers universal MLIPs, including their capabilities, downsides, applications, and relevance to organic and inorganic systems.Universal MLIPs are presented as a nascent class with potentially transformative impact.
  • Practical use: It offers practical guidance for estimating execution speed and choosing potentials based on hardware, simulation demands, accuracy, and fitting strategy.The selection guidance covers pre-trained potentials and fitting from scratch.
  • Infrastructure, limitations, and future: The review discusses training data, pre-trained potentials, training hardware, current limitations, mitigation strategies, and MLIP prospects over the next 3-10+ years.Limitations include long-range interactions, magnetic systems, and excited states.

2 A List of MLIPs

The review lists MLIPs by acronym and standard citation to clarify terminology, while explicitly noting that the list is not comprehensive.

  • The paper standardizes acronyms and citations for the MLIPs discussed, including BP-NN, CHGNet, DeepMD, GAP, MACE, NequIP, and SNAP.
  • The authors caution that the catalog is not meant to be a comprehensive list of existing MLIPs.
  • The catalog spans neural-network, graph-based, cluster-expansion, tensor-potential, and other MLIP approaches.
  • Named examples include universal and domain-focused potentials such as MACE-MP-0, MACE-OFF23, PFP, and ALIGNN-FF.
  • Additional listed methods include MTP, SevenNet, M3GNet, UF3, AIMNet2, EquiformerV2-OMAT24, FLARE, GPTFF, GNoME, and grACE.

3 What Makes MLIPs So Exciting?

MLIPs provide flexible energy-surface models that bridge the scale–accuracy gap between ab initio and physics-based simulations. Their practical advantages include broad chemical applicability, improved accuracy, and much higher efficiency, but they still require substantial data and have present capability constraints.

  • An MLIP maps atomic positions, element types, and optionally lattice vectors to a total energy and generally provides forces and stresses through derivatives.
  • MLIPs bridge AIMD’s accuracy and small scale with physics-based MD’s larger scale but limited quantitative accuracy for specific materials.
  • Flexible MLIP functional forms can represent complex potential-energy landscapes and chemical changes when adequate training data are available.
  • Current limitations include large training-data requirements, higher computational cost depending on MLIP type, and constraints on present capabilities.
  • A few to tens of meV/atom is typical MLIP accuracy, often an order of magnitude better than typical physics-based potentials.
  • MLIPs can access AIMD-like time and length scales with orders-of-magnitude less compute time, while some universal models target millions to billions of atoms and microsecond-or-longer timescales.

4 Understanding the Types of MLIPs - Basic Formalisms

MLIPs share a common structure in which atomic environments are featurized and mapped to energies, forces, or stresses, while explicit and implicit approaches differ mainly in whether features are predefined or learned. These distinctions are increasingly converging into a unified formalism.

  • 4 Understanding the Types of MLIPs - Basic Formalisms: The review organizes MLIP formalisms around training, execution speed, accuracy, ease of use, and suitability for specific problems.It provides a high-level guide rather than a fully detailed mathematical treatment.
  • 4.1 The Basic Structure of an MLIP: MLIPs generally represent each atom’s local environment numerically, then pass those features to a regression model whose accuracy depends on the representation’s expressiveness.Larger feature sets can capture environments more effectively.
  • 4.2 Explicit vs. Implicit MLIPs: Explicit AEF methods define features for each element, whereas implicit AEF methods learn atomic and bond embeddings, typically through deep-learning architectures such as GNNs.Implicit features can scale more efficiently with many chemical species, while their architectures may be harder or slower to train and execute.
  • 4.4 Implicit AEF and GNN MLIPs: Implicit AEF and GNN approaches learn effective features, with recent architectures using graph connectivity and symmetry-aware designs to reduce representation complexity and improve scalability.Examples include SchNet, M3GNet, NequIP, and newer DeepMD implementations.
  • 4.3 Explicit AEF Type MLIPs: Explicit AEF construction represents local environments with density expansions using radial and angular basis functions while respecting permutation, translation, and rotation symmetries.The resulting invariant or covariant features are supplied to an ML model, often to predict scalar energy.
  • 4.3 Explicit AEF Type MLIPs: Reference targets for explicit AEF models typically include energies, forces, and stresses obtained from ab initio calculations such as DFT.Stable structures, relaxation structures, and AIMD structures can all provide training data.
  • 4.5 Unifying explicit and implicit AEFs: Explicit AEFs on multilayer tree-structured graphs can be transformed into implicit AEFs for efficient evaluation, yielding message-passing neural networks.This connection supports viewing apparently different MLIP architectures as choices within a broader framework.

5 Universal MLIPs

Universal MLIPs extend learned potentials across broad chemical and structural domains, with applications spanning materials discovery, molecular dynamics, and accelerated structure optimization. They show strong average accuracy and emerging physics-aware capabilities, but targeted potentials can remain more accurate for specific systems and some properties remain difficult.

  • Scope and applications: U-MLIPs aim to model many chemical species and material arrangements, with semi-universal and true universal categories distinguished by application domain.Their broad applicability can increase modeled chemical and structural complexity with typically small or minimal accuracy loss.
  • Training data and performance: U-MLIP training databases expanded from roughly 2⨯105 structures in 2022 to a maximum of 1.18⨯108 in 2024, exceeding an order-of-magnitude average annual increase.The review identifies continued database growth as a demanding route toward further performance improvements.
  • Training data and performance: Across large test sets, U-MLIPs achieve energy, force, and stress accuracy that generally surpasses many-body PBPs and approaches explicit AEF accuracy, potentially nearing DFT accuracy across varied atomic arrangements.The review also reports strong benchmark performance for individual models, including MatBench results and faster execution for Orb.
  • Physics-aware capabilities: CHGNet incorporates charge-related and magnetic information, enabling modeling of valence states, charge transfer, and electronic-entropy effects in battery and transition-metal systems.Including magnetic moments reduced energy, force, and stress errors by 1-10% depending on the property, while enabling phenomena unavailable to earlier MLIPs.
  • Limitations and trade-offs: Surface energies and alloy mixing energies remain challenging because relevant environments may be absent from training data, with U-MLIPs showing larger surface errors and failing to reproduce metallic-alloy mixing energies adequately.Targeted MLIPs can lower errors for specific systems, but this comes at the cost of generality.
  • Scope and applications: U-MLIPs support materials exploration, DFT acceleration, and molecular dynamics at longer length and time scales than conventional calculations.Reported examples include structure optimization speedups by a factor of three and stable MD simulations lasting many nanoseconds.

6 Execution (Inference) Speed of MLIPs

MLIP execution speed depends on hardware, model complexity, system size, and parallelism, with simple models generally fastest and deep-learning models often slower. Reported timings span many orders of magnitude, while some MLIPs scale to hundreds of billions of atoms.

  • Benchmarking caveats: Benchmark values are approximate and should be treated cautiously because optimization and software versions change continually.The review recommends careful benchmarking of current package versions under the hardware and conditions of the specific project.
  • Performance metrics: Processor-seconds per atom per step enables performance comparisons across simulations with different atom counts, processor counts, and step numbers.For one processor, performance is reported as s/atom/step; with sufficient atoms per processor, proc-s/atom/step is largely insensitive to system and processor count.
  • Deep-learning MLIPs: Deep-learning MLIPs are generally similar to or somewhat slower than non-deep-learning explicit AEF approaches, although hardware differences complicate comparisons.DeepMD was reported at about 10^-3 proc-s/atom/step on a CPU and approximately 10^-4 proc-s/atom/step on a GPU after ~40× acceleration.
  • Universal MLIPs: 10^-3 proc-s/atom/step was reported for M3GNet on one CPU and 10^-3 s/atom/step for PFP on a GPU, with PFP about 5× slower than identical single-CPU MTP and ACE runs.PFP was about 50–100× slower than the same runs on a large CPU set.
  • Scaling: Some MLIPs scale to hundreds of billions of atoms, but these calculations generally require multiple GPUs.Parallel performance for SNAP, FLARE, and Allegro begins to deviate significantly from linear inverse-time scaling at around 10^5 atoms per GPU on the cited Summit hardware.
  • General trends: Simple physics-based potentials are fastest, followed by explicit AEF MLIPs and then implicit AEF deep-learning MLIPs.Single-processor performance is usually better than larger parallel calculations because MPI communication adds overhead.

7 MLIP Choices – What Should I Use When?

Choosing an MLIP requires balancing hardware resources, accuracy, speed, and the practical details of fitting or using a potential. The review presents these factors as a progression from basic constraints to a specific MLIP choice.

  • Selection criteria: MLIP selection should consider hardware, accuracy, speed, and whether to fit a new potential or use a pre-trained one.The section begins with hardware, accuracy, and speed before addressing the details of pursuing a specific potential.

7.1 Hardware Resources

Available hardware can strongly constrain MLIP selection. Explicit AEF models generally suit CPU-only workflows, whereas neural-network and graph-based models primarily rely on GPUs, with hosted inference providing another option.

  • Hardware compatibility: Explicit AEF MLIPs such as MTP, ACE, SNAP, and GAP generally run well on CPUs, while NN- and GNN-based MLIPs mostly rely on GPUs.NN-based MLIPs can run on CPUs but typically incur a 10–100× slowdown relative to GPU calculations.
  • Hardware compatibility: CPU-only users are likely best served by explicit AEF MLIPs, while GPU access supports both explicit AEF and NN-based choices.MTP is described as presently CPU-only; ACE can be fitted using GPUs and used for molecular dynamics on either CPUs or GPUs.
  • Infrastructure options: Access to at least one high-performing GPU is likely advisable for extensive MLIP use because U-MLIPs increasingly use GPUs for training and execution.Matlantis provides PFP through an API, with inference executed on backend GPUs or MN-Core deep-learning accelerators.

7.2 Speed Requirements

Simulation time is determined by system size, execution steps, hardware resources, and MLIP computational cost. When other factors are fixed, the MLIP framework becomes the determining cost factor, creating a speed–accuracy trade-off.

  • Determining simulation time: Overall simulation time depends on system size, number of execution steps, available hardware, and MLIP computational cost.If the first three factors are fixed by project and infrastructure requirements, the MLIP framework determines overall simulation time.
  • Speed–accuracy trade-off: Explicit AEF MLIPs are about 10–100× faster than implicit AEF deep-learning MLIPs.When computational cost is not limiting, deep-learning MLIPs typically provide the highest accuracy; otherwise, explicit AEF MLIPs are an alternative.

7.3 Accuracy Requirements

MLIP accuracy depends on the completeness of atomic-environment descriptions and the trade-off between model accuracy, computational cost, and available hardware. Recent equivariant graph neural networks offer high accuracy at practical computational cost, while CPUs currently favor MTP or ACE.

  • Model accuracy: Earlier MLIPs were limited by incomplete three-body and four-body descriptions of local atomic environments.MTP and ACE introduced systematically improvable high-body-order basis functions while retaining tractable linear regression.
  • Model accuracy: MTP and ACE increase MLIP accuracy through systematically improvable high-body-order descriptions while keeping computational cost tractable.These formalisms use linear regression to learn the potential-energy surface.
  • Hardware and model choice: Equivariant GNNs such as NequIP, Allegro, TeaNet, SO3krates, and MACE can achieve very high accuracies with practical computational cost.The review identifies these approaches as current examples of accurate equivariant models.
  • Hardware and model choice: When GPUs are available, the authors recommend equivariant models, whereas CPU-only users are directed toward MTP or ACE for the present.The recommendation reflects current implementation efficiency and is expected to change as packages improve.
  • Training strategy: Fine-tuning a pre-trained potential may provide higher accuracy than training a new MLIP from scratch, including for some out-of-distribution tasks.This finding is based partly on the authors’ personal experiences and should be interpreted in that context.

7.4 Using A Pre-trained Potential

Pre-trained potentials can save development time, but their suitability must be checked against the target system, chemical states, relevant physics, properties, and speed requirements. Validation should compare energies and forces and, where possible, the properties of interest before larger-scale application.

  • Selection: Starting from a pre-trained potential can save substantial time, and such potentials are available through online repositories.The review cites repositories including the NIST Interatomic Potentials Repository and OpenKIM.
  • Limitations: Low energy and force errors do not ensure accurate materials-property predictions for universal MLIPs.Many properties, including vacancy formation energies and melting temperatures, may remain incorrectly predicted.
  • Validation: Users should validate energies, forces, and target properties against consistent ab initio calculations before applying a potential broadly.A handful of static calculations on small unit cells can provide an initial benchmark, while property agreement gives greater confidence than energies and forces alone.
  • Validation: Property-specific validation can compare quantities such as phonon dispersions, diffusion coefficients, and defect formation energies before larger or compositionally different applications.The review recommends extending small-system benchmarks to the properties of interest whenever feasible.
  • Practical applications: Using universal MLIPs for pre-relaxation and initial screening can tolerate some failures because later ab initio calculations can identify or correct them.M3GNet pre-relaxation of hypothetical structures produced approximately 3⨯ time savings compared with ab initio calculations on un-relaxed structures.
  • Practical applications: Universal MLIP failures during screening can create false positives or false negatives, so promising candidates require later full ab initio screening.The review notes that screening is often focused on obtaining some successes rather than ensuring no false negatives.
  • Limitations: Because applicability remains uncertain, many practitioners still fit their own potential or use one specifically fit to the material under investigation.The review also notes an emerging practice of using some universal potentials without fine-tuning.
  • Selection: A pre-trained potential must match the study’s atomic structures, chemical states, and relevant physics.Examples include amorphous phases, high-valence transition-metal states, and systems with significant van der Waals contributions.

7.5 Fitting A New Potential: General workflow

Fitting a new MLIP is presented as an iterative, science-informed workflow: generate problem-relevant ab initio data, fit and evaluate the model, then add configurations until accuracy is adequate. Testing should include expected use conditions and target properties, not only randomly held-out structures.

  • Fitting strategy: Science-informed fitting uses physics, chemistry, and materials-science knowledge because no fit will be perfect for all atomic configurations.The goal is a model adequate for the user’s needs rather than universally perfect behavior.
  • Fitting strategy: There is currently no agreed-upon standard or widely accepted optimal workflow for fitting an MLIP.The review therefore provides general strategies and practical guidance rather than a single prescribed procedure.
  • General workflow: The workflow begins by generating ab initio configurations relevant to the intended application, such as liquids, phonons, or molecular distortions.Configuration selection can combine domain intuition, MD samples, and active learning.
  • General workflow: An initial model is fitted on about 80% of the data and tested on the remaining 20% for energy and force accuracy.For GNNs, a common split is 80% training, 10% validation, and 10% testing.
  • General workflow: If the fit is inadequate, additional data are added iteratively and the model is reassessed until the desired accuracy is reached.Complex systems or small additions may require many iterations, whereas sufficiently large initial data may require none.
  • Testing: Testing should target expected or near-expected use conditions, such as low-temperature Li hopping when diffusion is the objective.Configurations from those conditions can be evaluated with both ab initio methods and the MLIP.
  • Testing: Random cross-validation can be overly optimistic for highly correlated trajectory or perturbation data because similar configurations appear in both training and test sets.New conditions, such as MD at a different temperature, provide a more robust assessment.
  • Testing: Low energy and force RMSE can still accompany inaccurate properties when important activated or outlier configurations are poorly represented.The review gives Li diffusion and liquid-water properties as examples of this mismatch.

7.6 Fitting A New Potential: More specific considerations

Choosing and fitting an MLIP requires balancing chemical complexity, training-data demands, hardware, data efficiency, and hyperparameter effort. Explicit atomic-environment models can become costly as species and environments increase, while newer equivariant models may reduce data requirements but often require more complex training.

  • Chemical complexity: Explicit atomic-environment MLIPs can scale poorly with chemical complexity and training-set size, limiting both training and execution for multicomponent systems.The review identifies quadratic scaling with the number of species as a particularly important constraint.
  • Chemical complexity: FLARE is generally extremely fast in execution but can become prohibitively expensive to update near N_env = 1,000,000 environments.N_env scales as O[(training ab initio frames) × (atoms/frame) × (species)^2].
  • Chemical complexity: For FLARE, offline training is estimated at about 4000 frames for one species but only about 150 frames for five species, with 250 atoms per frame.The five-species limit may be too small to fit an accurate potential.
  • Chemical complexity: An ACE fit for a 12-species chloride-salt system with 3500 training data failed to start on an Nvidia Tesla V100-32 GB GPU.The review presents this as a concrete example of scaling difficulty in chemically complex systems.
  • Training requirements: For molten-salt FLiBe, MTP fitting used less than 1% of the DeepMD data, although the comparison was not head-to-head under identical conditions.Both approaches nevertheless produced excellent potentials in the reported studies.
  • Training requirements: Newer equivariant deep-learning MLIPs can be substantially more data-efficient than earlier deep-learning approaches.Studies with NequIP report a 1000⨯ improvement versus DeepMD with respect to data requirements.
  • Ease of fitting: MTP currently lacks GPU training support, so large training sets require many CPU cores.The review notes that this limitation may be removed by future code updates.
  • Ease of fitting: Hyperparameter complexity can multiply fitting time, with ACE offering many interaction-related parameters and Allegro requiring substantial optimization experience.MTP is described as comparatively easy to fit because it has one hyperparameter.

7.7 Summary of Considerations for Choosing a Potential

Choosing an MLIP requires balancing speed, hardware, accuracy, system complexity, and whether to reuse or fit a model. The review recommends practical starting points while emphasizing validation and trade-offs.

  • For fast simulations or CPU-only workflows, explicit AEF models such as MTP or ACE are recommended starting points.This guidance targets simulations lasting tens of nanoseconds or longer and users without GPU access.
  • Deep learning potentials are practical when speed is less critical or GPUs are available, but they are not always necessary or optimal.The review gives exploring a few thousand structural energies as an example of a less speed-constrained use case.
  • MTP is a practical starting point because comparisons found it highly accurate and fast, generally outperforming GAP, SNAP, and Behler-Parrinello NN potentials.The review presents MTP as a default when application-specific information is unavailable.
  • ACE unifies many explicit AEF basis functions and offers flexibility, but its larger hyperparameter space can make model selection more complex.ACSF, SOAP, HBFs, and MTFs are described as special cases within the ACE formalism.
  • Pre-trained or targeted MLIPs can avoid training that typically takes days to months, but users should validate them for the intended application.U-MLIPs require careful vetting, while reused targeted potentials may differ from their original assessed use cases.
  • Deep learning models generally support systems with more than ~5 elements, but they can require more data, training time, and hyperparameter optimization.Recent methodological developments may reduce the element-count constraint.

8 MLIP Infrastructure

MLIP infrastructure now spans data repositories, fitting and verification tools, simulation interfaces, and heterogeneous computing resources. These systems aim to make models easier to train, evaluate, deploy, and reproduce.

  • ColabFit Exchange provides curated training data, with nearly 400 datasets comprising more than 230 million unique atomic arrangements as of January 2025.
  • KLIFF supports general-purpose MLIP fitting, model deployment through LAMMPS via OpenKIM, automated testing, archiving, and uncertainty quantification.Its uncertainty quantification assesses prediction reliability and confidence.
  • Garden packages containerized models into domain-specific collections and simplifies deployment across local machines, cloud resources, and HPC clusters.Globus Compute integration supports access across these computing environments.
  • LAMMPS remains widely used for MLIP-based molecular dynamics, while ASE supports simulations and static energy or force calculations for testing and comparison.Most established MLIPs provide LAMMPS or ASE libraries, though installation requirements vary.
  • Earlier MLIPs have well-tested LAMMPS libraries, whereas MTP and newer models often require separately downloaded integrations; streamlined integration remains ongoing.NequIP and MACE also provide native integration with JAX-MD for GPU-based differentiable molecular dynamics.
  • Many MLIPs require GPUs for efficient operation, with large studies drawing on national facilities or pay-as-you-go cloud access to large GPU resources.Modest GPU access is becoming widespread, but obtaining many GPUs can remain difficult.

9 Limits of Standard MLIPs and Advanced MLIPs to Overcome Those Limits

Standard MLIPs face difficulties with long-range interactions, magnetism, and excited states because local, short-range models may omit essential physics. Advanced approaches add explicit interactions or specialized degrees of freedom, but often introduce scope or complexity trade-offs.

  • Long-range interactions: Standard MLIPs typically omit long-range interactions, so long-range forces, magnetism, and electronic excitation states generally require modifications beyond adding training data.
  • Long-range interactions: Short-range MLIPs can succeed for screened ionic systems, but higher-energy states without strong local charge neutrality may require explicit long-range electrostatics.The review connects successful molten-salt models to screening and cautions against relying only on short-range interactions in less-neutral states.
  • Long-range interactions: Explicit long-range treatments include physics-based electrostatics, Wannier-center models, self-consistent refinement, and dispersion or electrostatic corrections.Examples include DeepPot, SCFNN, AIMNet2, and DFT-D3 corrections added to short-range models.
  • Long-range interactions: Global sGDML captures chemical interactions and long-range forces by treating the molecular system as a unified entity, but it is trained for a specific molecule and is not transferable.
  • Magnetic systems: Non-collinear magnetic ACE reproduces iron ground states across magnetic phases and captures magnetic and lattice excitations relevant to finite-temperature and defect behavior.
  • Magnetic systems: TENN incorporates time-reversal symmetry and spin-lattice coupling, providing a route to spin-lattice dynamics while potentially reducing required DFT reference data.
  • Excited states: Advanced excited-state frameworks address phase-free training, rotationally covariant non-adiabatic couplings, spin-orbit couplings, and simultaneous same-multiplicity states.These capabilities improve prediction of excited-state dynamics and excited-state potential-energy surfaces.

10 The Future of MLIPs

The review anticipates MLIPs becoming easier to fit, evaluate, deploy, and select as infrastructure and datasets mature. Universal models may expand toward quantitative modeling, while specialized models remain practical for constrained problems.

  • Specialized universal or targeted MLIPs may be more practical than fully universal models because limiting species and phases simplifies fitting despite reducing compositional and structural coverage.
  • Within current explicit AEF and equivariant deep-learning frameworks, future work is expected to emphasize turn-key fitting, evaluation, and scaling to larger datasets.
  • The field needs standard methods for long-range forces and excited states, including magnetic states, while advanced magnetic models are expected to enter fitting packages within years.
  • In the next 3-5 years, general excited-state methods are expected to become widely accessible alongside broader infrastructure improvements and adoption.
  • Future software may fit multiple potentials, recommend the best model for a system, and provide searchable repositories with automated quality assessment.
  • Hardware portability and accelerators will increasingly influence MLIP choice, with new hardware demonstrating potential for more than two orders of magnitude faster MD than CPUs.
  • Universal MLIPs are currently mostly useful for qualitative or semi-quantitative screening but are rapidly becoming quantitative tools for specific properties such as Li diffusion.
  • Future universal MLIPs may act as foundational models, enabling simulations beyond 1 ms and access to materials phenomena unavailable to current MLIPs.
Loading 2503.09814v1…