Machine-learned interatomic potential¶

A trained surrogate that maps atomic configurations to interatomic energy, and often forces, within a validated materials domain.

Version: v1 · 2026-09-28 · History
Domain-specific #: 10516
Domain group: Natural Sciences
Origin domain: Chemistry & Materials Science
Subdomains: Computational Materials Science, Atomistic Simulation → Chemistry & Materials Science
Aliases: MLIP, Machine learning potential, MLP

Core Idea¶

A machine-learned interatomic potential is an energy model fitted from atomic configurations and reference results rather than specified solely by an analytic force-law recipe. For a given structure, elements and positions enter a learned mapping that returns a potential-energy estimate. Forces may be derived from the energy gradient or modeled with additional constraints. The reference may be density-functional calculations; the fitted model is a surrogate for that declared target, not a replacement for experiment or a proof of quantum accuracy everywhere.

Model families differ: Gaussian-process potentials use chosen structural descriptors, while neural and equivariant message-passing approaches learn different representations. Translational, rotational, and permutation behavior matters, but no one architecture or finite cutoff defines the whole class. An MLIP is useful in atomistic simulation because repeated evaluations can be cheaper than its reference method; the conclusion must be restricted to phases, elements, bond environments, and observables covered by validation. Deringer and Csányi's carbon potential is an attested case with explicit liquid/amorphous and surface uses.

Structural Signature¶

Sig role-phrases:

Atomic configuration — Supplies element identities and spatial positions or equivalent structural descriptors for evaluation. It is constitutive. Counterfactual: A molecular name without a geometry does not identify the modeled energy state.
Reference energy data — Provides the quantum or other chosen target energies, sometimes forces, used to fit and assess the surrogate. It is constitutive. Counterfactual: Without a reference target the learned output is not a potential approximation of a declared system.
Learned energy mapping — Maps structural input to a scalar potential-energy estimate, with force prediction where derived or separately trained. It is constitutive. Counterfactual: A classifier of element labels with no energy mapping is not an interatomic potential.
Symmetry and locality design — Handles translations, rotations, particle permutations, and interaction range as model choices tied to the physics. It is diagnostic. Counterfactual: A particular local cutoff or equivariant neural network is not mandatory for every MLIP.
Validation domain — Bounds error claims to represented elements, phases, configurations, and intended dynamics or relaxation task. It is boundary condition. Counterfactual: A good error on one carbon dataset does not license reliable predictions for an unseen metal or reaction path.

What It Is Not¶

Not a bare neural network. The model must map atomic geometry to interatomic energy for a declared material system.
Not a fixed empirical potential by default. An unlearned hand-parameterized pair law lacks the defining data-fitted mapping.
Not the training dataset. Reference structures and energies are inputs to fitting, not the fitted potential itself.
Not universally ab initio. Fidelity to one computational reference and test domain cannot establish accuracy for all matter.
Closest near-miss. A fixed analytic pair potential with manually assigned parameters and no data-driven training is the closest near miss: it maps configuration to energy but lacks the learned reference-surface fit.

Scope of Application¶

Atomistic simulation. Use validated energy/force estimates in repeated material-state calculations.
Surface studies. Evaluate eligible reconstructions within a trained chemistry and configuration range.
Model comparison. Contrast Gaussian, descriptor-based, and equivariant approaches under matched reference data.
Catalyst screening. Treat large DFT-labeled adsorbate–surface datasets as training and evaluation regimes, not universal guarantees.

Clarity¶

Identify the atoms, coordinates, reference calculation, learned model, and validation range. An accurate carbon model is not thereby a hydrogen, metal, or reactive interface model. A hand-coded force field may compute the same energy-shaped output without being machine-learned. Where forces are reported, check whether they come from the energy gradient and whether that consistency was tested.

Manages Complexity¶

The model compresses many reference evaluations into a reusable energy function. This makes long or numerous atomistic calculations feasible, but hides training-set coverage, descriptor choices, and extrapolation failures behind a single predicted energy. Reporting those limits is part of the abstraction, not an optional performance footnote.

Abstract Reasoning¶

Declare the target chemical system and reference energy/force method.
Describe atomic input, learned mapping, and symmetry or locality assumptions.
Fit the mapping conceptually on reference configurations and keep independent validation separate.
Check held-out energy, force, and relevant structural behavior in the intended regime.
Reject untested extrapolation from one element, phase, or bond environment to another.

Knowledge Transfer¶

The configuration–reference–surrogate–validation pattern transfers among different material systems, but a trained carbon potential's weights, cutoff, and error estimates do not. The representation relation is portable; the exact potential remains tied to its atomic species, reference method, and sampled energy landscape. A property predictor without an energy function may use ML but is not this abstraction.

Examples¶

Canonical¶

Consider an illustrative learned potential for one element: reference calculations label several atomic geometries with energies, and a model learns an invariant mapping from positions and species to an energy estimate. If differentiable, the negative energy gradient supplies force estimates. Evaluation on held-out configurations can reveal error near unfamiliar bond rearrangements. This worked construction defines the roles without asserting that any toy model has published accuracy.

Mapped back: Atomic configuration → illustrative element identities and three-dimensional atom positions; Reference energy data → specified reference energies for training geometries; Learned energy mapping → fitted structure-to-energy function and optional gradient forces; Symmetry and locality design → invariant input treatment, with cutoff choice left explicit; Validation domain → held-out geometries and stated failure near unfamiliar bonds.

Applied / In Practice¶

Deringer and Csányi's published Gaussian approximation potential learns a density-functional reference surface for liquid and amorphous carbon. They evaluate energetic and structural properties across several carbon densities and use the model in simulations of tetrahedral amorphous-carbon surfaces, including reconstruction. Their results support that carbon model's tested regime, not a promise of DFT-level fidelity for all elements, phases, or new bond environments.

Mapped back: Atomic configuration → liquid/amorphous carbon structures and surface geometries; Reference energy data → density-functional carbon energy data; Learned energy mapping → the published Gaussian approximation potential; Symmetry and locality design → hierarchical two-, three-, and many-body finite-range descriptors; Validation domain → tested carbon densities and surface reconstructions, not arbitrary chemistry.

Structural Tensions¶

T1 — Reference Fidelity versus Simulation Cost. A learned surrogate can reduce repeated high-cost reference evaluations while introducing approximation error and coverage dependence.

Diagnostic: Which reference calculations and independent tests support the speed–accuracy tradeoff?

T2 — Symmetry Constraints versus Model Flexibility. Physics-aware invariance or equivariance can improve consistency, yet a chosen architecture and cutoff can miss long-range or unfamiliar interactions.

Diagnostic: Which symmetries and interaction ranges are actually encoded?

Structural–Framed Character¶

The skeleton is representation: a surrogate maps configurations to estimated values while preserving selected properties of a reference relation within a validated domain. A machine-learned interatomic potential approximates an atomic energy surface and may derive forces from energy gradients. Its approved parent is Representation.

Evaluative weight: Accuracy is conditional on training coverage, symmetry handling, reference method, and held-out tests.

Human-practice-bound: Choice of atomic species, geometry descriptors, and simulation tasks defines intended use.

Institutional origin: Materials modeling and computational chemistry set energy and force conventions.

Vocabulary travels: A learned property predictor can be “machine-learned” without representing a potential energy function.

Import versus recognize: Surrogate-validation reasoning transfers, but weights and error bounds do not move unchanged to another material.

Its character: A trained atomic-energy representation with domain-bounded fidelity, not a prime for prediction.

Structural Core vs. Domain Accent¶

Skeletal core. A target relation is encoded in a manipulable surrogate with a declared mapping and fidelity boundary.

Domain-bound accent. The target is interatomic energy across atomic configurations; the learned model estimates that energy and may obtain forces via its gradient. Species, reference calculations, symmetries, and sampled configurations constrain the result.

Why not prime. Weather models and text embeddings can also be representations, but they lack the atomic energy surface. A predictor with no energy function does not become this potential through its learning method alone.

This entry is a kind of Representation.

Strict parent — representation. The reference energy surface is the target, trained parameters and descriptors form the medium, atomic configurations are mapped to estimated energies, and validation states which features are preserved sufficiently for simulation.
Related — empirical potential and Mie potential. Both yield configuration-dependent energies, but the learned reference mapping is the distinguishing MLIP property; Mie is a fixed analytic pair-potential family.

Relationships to Other Abstractions¶

Current abstraction Machine-learned interatomic potential Domain-specific

Parents (1) — more general patterns this builds on

Machine-learned interatomic potential is a kind of Representation Prime

A learned potential represents a reference interatomic energy relation with tested fidelity over atomic configurations.

Hierarchy path (1) — routes to 1 parentless root

Machine-learned interatomic potential → Representation → Abstraction

Neighborhood in Abstraction Space¶

Machine-learned interatomic potential sits in a moderately populated region (45^th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Molecular Structure & Interaction Models (20 abstractions)

Nearest neighbors

Energy minimization — 0.91
Principal interacting orbital — 0.87
Piola transformation — 0.87
Acidic — 0.86
Constraint (Computational Chemistry) — 0.86

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Empirical force field. Tell: Was the potential learned from reference configurations or fixed by an analytic recipe?
Energy dataset. Tell: Is there a fitted callable mapping for new atomic geometries?
Molecular property predictor. Tell: Does the output actually serve as an interatomic energy potential?
Density-functional calculation. Tell: Is this the expensive reference or a learned surrogate of it?

References¶

Deringer and Csányi, 'Machine-learning based interatomic potential for amorphous carbon,' author manuscript: https://arxiv.org/abs/1611.03277
Batatia et al., 'MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields,' author manuscript: https://arxiv.org/abs/2206.07697
Open Catalyst Project, overview and reference datasets: https://opencatalystproject.org/
Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Machine-learned_interatomic_potential (revision 1368845520).
Preserved source candidate: https://link.aps.org/doi/10.1103/PhysRevLett.104.136403
Preserved source candidate: https://gap-ml.org/
Preserved source candidate: https://link.aps.org/doi/10.1103/PhysRevB.87.184115
Preserved source candidate: https://link.aps.org/doi/10.1103/PhysRevB.95.094203
Preserved source candidate: https://link.aps.org/doi/10.1103/PhysRevX.8.041048
Preserved source candidate: https://link.aps.org/doi/10.1103/PhysRevB.90.104108
Preserved source candidate: https://pubs.acs.org/doi/10.1021/acs.jpcb.8b06476
Preserved source candidate: https://link.aps.org/doi/10.1103/PhysRevMaterials.8.033804

The frozen Wikipedia revision is discovery provenance. The retained source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; a thin authority surface is recorded as a nonblocking source-strengthening repair rather than concealed.