Machine Learning Interatomic Potentials: Simulation-in-the-Loop Discovery

A self-driving lab that synthesizes and characterizes one candidate material per hour is impressive until you count the search space. There are more plausible inorganic compositions than a robot could make in the lifetime of the sun. The bottleneck was never the robot; it was deciding what to make. Machine learning interatomic potentials are the component that turns that intractable outer loop into a tractable one — a surrogate for quantum mechanics fast enough to relax and screen millions of candidate structures before a single vial is touched. They sit between generative models that propose structures and the physical instruments that verify them, and they are the reason 2023-2026 saw discovery pipelines jump from thousands of candidates to millions.

This piece is a systems view, not a lab bench manual. We treat the potential as an engineering component with an accuracy budget, a cost model, failure modes, and integration contracts.

What this covers: what a potential is and why DFT is accurate but expensive; the architecture of modern equivariant and foundation potentials; how they plug into the closed discovery loop; active learning; and where they quietly break.

Context and Background

Every atomistic simulation reduces to one question asked billions of times: given the positions of a set of atoms, what is the total energy, and what force does each atom feel? Answer it accurately and you can relax a structure to its ground state, run molecular dynamics, compute phonons, and estimate whether a material is stable. The gold standard answer comes from density functional theory (DFT), which solves an approximation to the electronic Schrodinger equation. DFT is accurate to a few meV per atom for many properties, but it scales roughly as the cube of the number of electrons. A single relaxation of a modest unit cell can cost minutes to hours of CPU time, and a nanosecond of dynamics for a few hundred atoms is often simply out of reach.

The historical escape hatch was the classical force field: a hand-designed functional form for bonds, angles, and pairwise interactions with a few dozen fitted parameters. These are millions of times cheaper than DFT and enabled decades of biomolecular simulation. But they are brittle. A force field parameterized for one chemistry silently produces nonsense on another, and most cannot describe bonds breaking or forming at all. For high-throughput discovery across the periodic table, neither pole works: DFT is too slow to screen millions of candidates, and classical fields are too narrow to trust on novel chemistries.

Machine learning interatomic potentials target the gap. They learn the energy surface directly from DFT reference data, aiming for near-DFT accuracy at a cost closer to a classical field. This is the surrogate that makes the autonomous materials discovery pipeline economically sane. For the foundational chemistry, the Materials Project documentation is a good external primer on how large DFT datasets are curated at scale.

Reference Architecture: The Simulation-in-the-Loop Discovery Pipeline

The defining move of modern computational discovery is to put a fast, differentiable surrogate inside the loop that used to call DFT. The canonical pattern — popularized at scale by DeepMind’s GNoME effort, which reported roughly 2.2 million candidate crystals and about 380,000 predicted-stable structures — is a funnel. Cheap generation proposes many candidates, a potential relaxes and ranks them, DFT verifies only the survivors, and physical synthesis attempts only the tiny fraction that clears every filter. Each stage is orders of magnitude more expensive and orders of magnitude more selective than the last.

Figure 1: The simulation-in-the-loop discovery funnel, from candidate generation through MLIP screening and DFT verification to robotic synthesis and retraining.

Figure 1 shows the full loop. A generator (a structural pipeline that mutates known crystals, plus a compositional pipeline that samples formulas) emits candidate structures. The potential relaxes each candidate to a local minimum and estimates its energy, which is compared against the convex hull of known phases to get an “energy above hull” — the standard proxy for thermodynamic stability. Only candidates within a small tolerance of the hull survive. Those are handed to DFT for verification, and only DFT-confirmed structures proceed to a robotic synthesis and characterization stage. Crucially, both the DFT results and the experimental outcomes are fed back to retrain the potential, so the surrogate gets sharper exactly where the pipeline is exploring. This feedback edge is what separates a discovery engine from a one-shot screen.

The reason the funnel works is that selectivity and cost move in opposite directions at every stage. Generation is nearly free and hugely permissive — it will happily emit millions of formulas, most of them junk. The potential is cheap and moderately selective, discarding the obviously unstable in bulk. DFT is expensive and highly selective, and synthesis is the most expensive step of all and the final arbiter. If you plot survivors against cost-per-item, you get a staircase where each tread is roughly an order of magnitude taller and an order of magnitude narrower than the one below it. The engineering goal is to push as much of the rejection as possible up-funnel, into the cheap stages, without discarding true positives. That trade — recall versus cost — is the central tuning knob of the whole pipeline, and it is set by the potential’s accuracy and by how aggressively you threshold on energy above hull.

Direct answer: A machine learning interatomic potential is a model that predicts a system’s total energy from atomic positions and species, and derives forces and stresses as gradients of that energy. In a discovery pipeline it acts as a near-DFT-accuracy surrogate that relaxes and ranks millions of candidate structures cheaply, so that expensive DFT and physical experiments are spent only on the most promising few.

Where the potential sits in the stack

The potential is not a standalone product; it is a service consumed by three neighbors. Upstream, a generative or enumerative model supplies structures. Alongside it, a workflow engine drives relaxations and molecular dynamics. Downstream, a decision layer converts predicted stability into a synthesis queue. In a full closed-loop self-driving lab the potential’s outputs also gate scheduling: a robot arm is a scarce resource, so the ranking a potential produces is effectively a priority function on physical experiments.

Energy, forces, and why gradients matter

A well-designed potential predicts a single scalar — the total energy — and obtains forces as the negative gradient of that energy with respect to atomic positions, and stress as the gradient with respect to cell strain. This is not a stylistic choice. Deriving forces from a single energy by automatic differentiation guarantees the force field is conservative: energy is conserved along a trajectory, and there is no spurious work created or destroyed as atoms move. Models that predict forces as an independent output can be marginally more accurate per-force on a static benchmark but drift in long dynamics because their forces do not integrate back to any consistent energy. For anything that runs MD, the energy-gradient formulation is the safe default.

The accuracy budget

Practitioners speak in two numbers: energy mean absolute error in meV per atom, and force MAE in meV per Angstrom. On complex multi-component datasets, strong equivariant models report roughly 1-2 meV/atom and 15-50 meV/Angstrom, close enough to DFT that relaxations land in the same basin. Foundation-scale general potentials trade some per-system accuracy for breadth: published figures for early general models sit in the low tens of meV/atom for energy across dozens of elements — for instance, MACE-MP-0 has been reported in the neighborhood of 20 meV/atom for energy and tens of meV/Angstrom for forces on broad validation sets, with materially larger errors on chemistries far from its Materials Project training data. The units matter more than they look. An energy error of 25 meV/atom is roughly 2.4 kJ/mol per atom, which is comparable to the energy scale that separates competing polymorphs; a screen operating at that resolution will occasionally rank a metastable phase above the true ground state. That is tolerable for a coarse first pass whose survivors get DFT-verified, and unacceptable for a final ranking with no verification gate. Always read an accuracy figure against the decision it will drive.

A Deeper Walk-Through: Architecture, Foundations, and the Model Zoo

To understand both the power and the limits of these models, you have to look at how they represent atoms. The core problem is symmetry. The energy of a molecule does not change if you translate it, rotate it, or relabel identical atoms. Forces, by contrast, rotate with the system — they are equivariant, not invariant. A good architecture bakes these symmetries in rather than hoping to learn them from data.

Figure 2: An equivariant message-passing potential. Atoms become graph nodes within a radial cutoff; equivariant layers exchange messages; a readout produces per-atom energies that sum to a total, from which forces and stress follow by autograd.

Figure 2 traces the forward pass. Each atom is a node, and edges connect atoms within a fixed radial cutoff — typically 4 to 6 Angstroms. Early descriptor-based methods, such as the Smooth Overlap of Atomic Positions (SOAP) and the Atomic Cluster Expansion (ACE), hand-engineer a rotation-invariant fingerprint of each atom’s neighborhood and regress energy from it. Modern message-passing networks instead learn features that are equivariant to rotation: internal representations transform as proper geometric objects (scalars, vectors, higher tensors), so the network can represent directional information without breaking symmetry. NequIP demonstrated that E(3)-equivariant message passing is dramatically more data-efficient than invariant predecessors. MACE combined equivariant messages with higher-order ACE-style features to cut the number of message-passing layers needed, which matters because each layer widens the receptive field and slows inference. Allegro took a different route: a strictly local, equivariant model with no message passing between distant atoms, trading some receptive field for near-linear scaling and easy parallelism on large systems.

The receptive field is the quantity to reason about. With a cutoff of 5 Angstroms and three message-passing layers, an atom effectively “sees” information from up to 15 Angstroms away, because each layer passes messages one hop further. More layers mean a wider field and better handling of medium-range order, but each layer is a full pass over every edge in the graph, so inference cost grows with layer count and with the average number of neighbors. This is the tension MACE’s higher-order features exploit: by packing more of the neighborhood’s geometry into each message, it reaches competitive accuracy with fewer layers, which is why it is often faster at a given accuracy than a deeper invariant network. Allegro attacks the same cost from the opposite side by refusing to propagate messages between atoms at all — every atom’s energy depends only on its own neighborhood, so the model is embarrassingly parallel and scales to systems of hundreds of thousands of atoms on a single node, at the cost of a strictly local receptive field.

From bespoke to foundation potentials

Until recently every serious simulation trained a bespoke potential on system-specific data. The shift since 2023 is toward foundation interatomic potentials: single models trained on enormous DFT datasets that aim to generalize across most of the periodic table out of the box. MACE-MP-0 is the emblematic example — a MACE-architecture model trained on Materials Project relaxations, covering on the order of 89 elements, and usable zero-shot on chemistries it never saw explicitly. Meta’s Open Materials 2024 (OMat24) pushed the data axis hard, releasing over 110 million DFT calculations and EquiformerV2-based models in 31M, 86M, and 153M parameter sizes that topped the Matbench Discovery leaderboard. A parallel wave of industrial models (Orb-class potentials, Microsoft’s efforts, and others) chase the same target from different architectural bets. The common thesis: pretrain broadly, then fine-tune on a few hundred to a few thousand system-specific labels to recover bespoke accuracy for a fraction of the data cost.

Fine-tuning is worth understanding mechanically because it is where most of the practical value lands. A foundation potential arrives already knowing the general shape of chemical bonding across the periodic table; what it lacks is calibration to your specific functional, pseudopotentials, and chemistry. Fine-tuning nudges the readout and upper layers using a small, targeted dataset — often generated by the active-learning loop below — so the model corrects its systematic offset on your system without forgetting the broad prior. Empirically this recovers near-bespoke accuracy with one to two orders of magnitude less data than training from scratch, because the model is adjusting rather than learning bonding from nothing. The failure mode to watch is catastrophic forgetting: fine-tune too hard on a narrow set and the model becomes a bespoke potential again, losing the transferability you paid for. Freezing lower layers, using a small learning rate, and mixing a slice of the original pretraining data back in are the standard mitigations.

Choosing an approach

The table below compares the practical families. Numbers are indicative ranges from the public literature, not guarantees; benchmark on your own chemistry before committing.

Approach	Typical accuracy	Cost vs DFT	Data hunger	Best fit
Classical force field	Poor off-domain	~1e6-1e7x cheaper	None (hand-tuned)	Known, non-reactive chemistry
Descriptor + regression (SOAP, ACE)	~1-5 meV/atom in-domain	~1e4-1e6x cheaper	Moderate	Single-system, interpretable
Equivariant MPNN (NequIP, MACE)	~1-2 meV/atom in-domain	~1e3-1e5x cheaper	Low-moderate	High-accuracy bespoke
Local equivariant (Allegro)	~1-2 meV/atom in-domain	~1e3-1e5x cheaper	Moderate	Very large systems, HPC
Foundation potential (MACE-MP-0, OMat24)	~10-50 meV/atom zero-shot	~1e3-1e4x cheaper	Pretrained; fine-tune small	Broad screening, cold start

Figure 3: The active-learning loop. The potential drives simulation, uncertainty is estimated from committee spread, high-uncertainty configurations are labeled with DFT and folded back into training.

Figure 3 shows the mechanism that keeps a potential honest as a pipeline wanders into new chemistry. Because a surrogate is only trustworthy near its training distribution, you need a signal for “I have not seen anything like this.” The workhorse is committee (ensemble) disagreement: train several potentials on the same data with different seeds, and where they agree the prediction is likely sound, where they diverge you are extrapolating. During MD or relaxation, configurations whose committee spread exceeds a threshold are flagged, labeled with a single DFT calculation, and added to the training set. Retraining or fine-tuning on those hard cases sharpens the model exactly where it was weakest. This uncertainty-driven data acquisition is what let GNoME-style efforts grow their training data by iterating generation, prediction, and DFT verification rather than fixing the dataset up front — the same spirit as Bayesian optimization for autonomous experiments, but acting on the simulation rather than the wet lab. Newer variants add confidence heads or single-model uncertainty estimates to avoid the cost of an ensemble, but committee disagreement remains the robust default.

Two design choices govern how well the loop learns. The first is the acquisition rule — what you select for labeling. Pure uncertainty sampling grabs the single most uncertain configuration, but that greedily over-samples one pathological region and ignores the rest of the space. Batch acquisition with a diversity penalty — pick configurations that are both uncertain and unlike each other — spends the DFT budget far more efficiently, which matters because each label is an expensive quantum calculation. The second choice is the labeling budget per round. Label too few and the model barely moves between retrains, wasting compute on repeated near-identical structures; label too many and you pay for calculations the model would have gotten right anyway. A practical rhythm is to label tens to low hundreds of high-value configurations per cycle, retrain or fine-tune, and let the new model choose the next batch. Over several cycles the training set concentrates in exactly the regions the pipeline actually visits, which is why an actively grown dataset of a few thousand well-chosen structures routinely outperforms a passively collected one many times larger. The uncertainty signal, not raw data volume, is the scarce resource.

An illustrative cost model

Consider a concrete, illustrative screen (numbers rounded for intuition, not measured). Suppose you want to relax 1,000,000 candidate structures averaging 40 atoms. A DFT relaxation of such a cell might take ~1 CPU-hour; a foundation-potential relaxation might take ~1 GPU-second. Screening all million with DFT is on the order of 1,000,000 CPU-hours — infeasible for most groups. The same screen with a potential is roughly 1,000,000 GPU-seconds, about 280 GPU-hours, a job you can finish over a weekend. If the potential correctly narrows the field to the top 2,000 candidates, verifying those with DFT costs ~2,000 CPU-hours. The potential converted a 1,000,000-CPU-hour problem into ~280 GPU-hours plus ~2,000 CPU-hours — a three-to-four order of magnitude reduction in the expensive resource. The catch, of course, is the word “correctly,” which is where failure modes enter.

Operationally, extracting that speedup is its own engineering exercise. A relaxation is not one forward pass but tens to hundreds of them, one per optimizer step, and each step is bottlenecked by graph construction and neighbor lists as much as by the network itself. Throughput on a screen therefore depends on batching many small structures together to keep the GPU saturated, caching neighbor lists across steps where geometry changes slowly, and choosing a model whose per-call latency fits your candidate size. A larger foundation model may be more accurate but slow enough that you can afford fewer candidates; a smaller or strictly local model may let you screen ten times as many structures at slightly lower fidelity. Because the pipeline is a funnel, the right operating point is usually a cheaper, faster potential up-funnel where you are rejecting in bulk, and a larger fine-tuned model deeper in where precision decides which few candidates reach DFT. Treating “the potential” as a single fixed choice, rather than a tier of models matched to each funnel stage, leaves most of the available throughput on the table.

Reading a materials-discovery benchmark

Because a foundation potential is now a purchased or downloaded component, teams increasingly pick one off a leaderboard — and leaderboards are easy to misread. Matbench Discovery, the de facto standard for this task, does not score raw energy MAE alone; it frames the real job as a classification problem: of the structures a model predicts to be stable, how many actually are? The headline metrics are therefore precision, recall, and F1 on stability against DFT, alongside the energy-above-hull MAE. A model can post an excellent energy MAE and still have mediocre F1 if its errors cluster right at the decision boundary near the convex hull, which is exactly where a screen lives or dies. Rank on the leaderboard is a starting filter, not a verdict for your chemistry.

Three subtleties decide whether a leaderboard number transfers to your pipeline. First, the discovery acceleration factor — how much a model shrinks the DFT budget to find a fixed number of stable materials — depends on your stability threshold, not just the model; tightening the energy-above-hull cutoff raises precision but sacrifices recall, moving you along the same recall-versus-cost curve discussed earlier. Second, aggregate scores hide chemistry-specific weakness: a model that is superb on oxides can be poor on intermetallics or hydrides, and the single leaderboard number averages that away. If your target space is narrow, benchmark on that space or fine-tune into it. Third, watch for training-set overlap. A model trained on the same Materials Project relaxations used to define the hull will look better than it will on genuinely novel chemistries, because part of the “test” is effectively in-distribution. The honest read of any leaderboard is: it tells you which models are worth trialing, and nothing more. Run the two or three top candidates on a held-out slice of your own chemistry, measure F1 at your operating threshold, and let that decide. This is the same discipline that separates a real evaluation from a vanity metric anywhere in machine learning — the benchmark scopes the shortlist, your data picks the winner.

Trade-offs, Gotchas, and What Goes Wrong

Figure 4: A taxonomy of where potentials break — out-of-distribution chemistry, reactive events, long-range interactions, charged and spin states, rare configurations, and soft unphysical modes.

Figure 4 organizes the honest caveats. The first and most common is out-of-distribution failure: a potential trained on near-equilibrium bulk crystals will confidently produce garbage on a chemistry, coordination, or pressure regime it never saw. The failure is silent — the model returns a smooth, plausible-looking energy — which is precisely why the uncertainty loop of Figure 3 is not optional. Reactive events are a second class: many models trained on relaxed structures rarely see bonds mid-break, so transition states and reaction barriers can be badly wrong even when equilibrium energies are excellent.

Long-range interactions are a structural limitation. A model with a 5-Angstrom cutoff and a few message-passing layers physically cannot see electrostatics or dispersion beyond its receptive field, so charged defects, ionic surfaces, and polar interfaces are hard. Related is the charged- and spin-state problem: standard potentials take only positions and atomic numbers as input, with no notion of total charge or magnetic state, so they cannot distinguish two systems that differ only in oxidation or spin. A particularly dangerous mode is the “soft” unphysical configuration: during an aggressive relaxation the model may find a low-energy hole that does not exist in reality — atoms collapsing to unphysically short distances because the training data never penalized that region. These holes can crash MD or, worse, produce a “stable” structure that is an artifact.

A concrete, illustrative failure sequence makes the risk vivid. Suppose a screen built on a bulk-crystal foundation potential proposes a new layered oxide. The potential relaxes it to a tidy structure and reports it comfortably below the hull, so it advances. But the true material has a charged surface reconstruction and significant long-range electrostatics that the model’s short cutoff never captured; the “stable” structure is an artifact of an energy surface that is smooth but wrong in exactly the regime that matters. Without an uncertainty flag, the mistake propagates all the way to a synthesis attempt that quietly fails, burning reagents and instrument time. With committee uncertainty, the anomalous configuration lights up as high-disagreement, gets a DFT label, and either corrects the model or is dropped before it ever reaches the robot. The lesson is not that the potential is bad; it is that a surrogate without an out-of-distribution alarm is a liability precisely because its errors are silent and plausible.

Phonons and defects deserve special mention because they probe the curvature of the energy surface, not just its value. A potential can nail total energies while getting second derivatives wrong, producing imaginary phonon modes or incorrect defect formation energies. Always validate the specific property you care about; a good energy MAE does not certify a good phonon spectrum. Finally, remember that a foundation potential inherits every bias of its DFT training data — the same exchange-correlation functional, the same systematic errors. The surrogate cannot be more accurate than the theory that labeled it.

Practical Recommendations

Treat the potential as a filter with a known false-positive rate, not an oracle. Its job is to be cheap and to rank; DFT and experiment remain the arbiters of truth. Design the pipeline so that every synthesis decision is backed by a DFT verification, and so that the potential’s mistakes cost GPU-seconds, not reagents. Start from a foundation potential for cold-start breadth, then fine-tune on a few hundred system-specific labels once you know your target chemistry — this almost always beats training from scratch on a small dataset. Instrument uncertainty from day one; a screen without an uncertainty signal is a screen you cannot trust off-domain. And validate the exact property that drives your decision, not a generic energy metric.

Use this checklist before you trust a potential in a loop:

[ ] Confirm the target chemistry is inside the training distribution, or plan active learning to bring it in.
[ ] Derive forces and stress from the energy by autograd; verify energy conservation on a short NVE run.
[ ] Stand up a committee or other uncertainty estimate and set a labeling threshold.
[ ] Validate the decision-relevant property (stability, phonons, barriers) against DFT on a held-out set.
[ ] Guard against soft collapse with distance floors or short-range repulsion.
[ ] Keep a DFT verification gate before any physical synthesis.
[ ] Log every flagged configuration and fold it back into training.

Non-advisory note: this article is a systems and architecture overview for engineering audiences and is not scientific, safety, or investment advice; validate all methods and numbers against primary sources for your own application.

Frequently Asked Questions

What is a machine learning interatomic potential in one sentence?

It is a model that predicts the total energy of a group of atoms from their positions and chemical identities, and computes the forces on each atom as the gradient of that energy. Trained on quantum-mechanical reference data, it aims to reproduce DFT accuracy at a tiny fraction of the cost. That speed is what makes it usable inside high-throughput screening and molecular dynamics, where DFT itself would be far too slow to run at scale.

How is a foundation potential different from a traditional one?

A traditional potential is trained on data for one system or a narrow chemistry and is trustworthy only there. A foundation interatomic potential, such as MACE-MP-0 or the OMat24 models, is pretrained on very large DFT datasets spanning most of the periodic table, so it can be applied zero-shot to chemistries it never saw explicitly. In practice you use it for broad screening or cold starts, then fine-tune on a small system-specific dataset to recover bespoke accuracy cheaply.

Why not just use DFT for everything?

DFT is the accuracy reference, but its cost scales roughly as the cube of the number of electrons, so a single relaxation can take CPU-hours and long dynamics are infeasible. Screening millions of candidate structures with DFT would take millions of CPU-hours. A potential does the same relaxations in GPU-seconds, letting you spend scarce DFT budget only on the handful of candidates that survive the surrogate screen. DFT remains the verification gate, not the screening workhorse.

How do you know when a potential is wrong?

You cannot tell from the prediction alone — a potential returns a smooth, plausible energy even when extrapolating badly. The standard defense is an uncertainty estimate, most commonly the disagreement across a committee of independently trained models. Where they agree, trust the prediction; where they diverge, you are out of distribution and should label that configuration with DFT and retrain. This active-learning loop is what keeps a potential reliable as a pipeline explores new chemistry.

What does simulation-in-the-loop mean here?

It means placing the potential inside the discovery loop as a fast surrogate for physical or DFT evaluation. Candidates are generated, relaxed and ranked by the potential, verified by DFT, and only then synthesized by robots; results flow back to retrain the potential. The simulation is not a one-off pre-study but a live component that filters what the wet lab attempts, shrinking a combinatorial search into a short shortlist before any experiment runs.

Where do foundation potentials still fail?

Out-of-distribution chemistry, reactive bond-breaking, long-range electrostatics and dispersion beyond the model cutoff, charged and magnetic states the input cannot express, and rare configurations like specific defects or phonon modes. They can also find “soft” unphysical energy minima that do not exist in reality. And because they learn from DFT, they inherit that theory’s systematic errors — a surrogate is never more accurate than the labels it was trained on.

ML Interatomic Potentials: Simulation-in-the-Loop Discovery (2026)