Machine-learned interatomic potentials (MLIPs) predict the potential energy surface of a material by learning from density functional theory calculations rather than from analytical force field expressions. The speed gain is real and large: a DFT calculation on a 100-atom unit cell might take eight to twelve hours on a modern cluster. An MLIP evaluation of the same cell takes milliseconds. That speed difference is what makes it computationally feasible to screen thousands of inorganic candidates in a single run rather than queuing DFT jobs for weeks.
What gets left out of the speed comparison is the fine print. An MLIP is only reliable within the region of chemical and structural space it was trained on, and the boundary of that region is not visible from the outside. A model will produce a number for any structure you give it, including structures it has never seen anything like. Whether that number means anything depends entirely on whether you are interpolating from the training set or extrapolating away from it. This post describes how MLIPs are built, what they predict reliably for inorganic crystals, and the specific failure modes we have encountered and built around in our screening workflow.
How MLIPs Are Trained
A modern MLIP for inorganic materials is typically built on a graph neural network. Each atom in the structure is a node, and edges encode interatomic distances and bond angles within a defined cutoff radius. The network takes this graph as input and predicts atomic contributions to total energy, from which forces are derived analytically as gradients. Training minimizes prediction error over a dataset of DFT-calculated energies and forces across a range of structural configurations: equilibrium structures, thermally displaced snapshots, configurations along transition pathways, and sometimes amorphous phases.
The quality and coverage of the training data determines the quality of the model. A model trained only on equilibrium 0 K structures will underperform on configurations sampled from a finite-temperature molecular dynamics trajectory. A model trained on a narrow composition range will extrapolate poorly to adjacent compositions, even if the crystal prototype is the same. A model with heavy representation of one phase family will have systematically lower accuracy on other families even when the constituent elements overlap.
The Descriptor and Architecture Question
The representation of local chemical environment fed to the network (the descriptor) is where most of the practical accuracy differences between MLIP architectures live. Older descriptors, Behler-Parrinello symmetry functions and SOAP kernels, encode local geometry but lose information about long-range interactions beyond the cutoff radius. Equivariant graph neural network architectures such as NequIP and MACE maintain rotational and reflection equivariance in the learned features, which matters for systems with directional bonding. For ionic inorganic crystals where electrostatic interactions are significant, descriptors that can represent charge state implicitly tend to outperform purely geometric descriptors near phase boundaries and for properties like intercalation voltage.
What MLIPs Predict Well in Inorganic Crystals
For inorganic crystal structures that are well-represented in the training data, modern MLIPs predict formation energy to within roughly 20 to 50 meV/atom against DFT reference values, and force predictions are accurate to within a few percent for near-equilibrium configurations. This is sufficient accuracy for relative stability comparisons: distinguishing a structure that is 15 meV/atom above the convex hull from one that is 120 meV/atom above is reliable enough to use as a screening filter.
Phonon calculations, which require force evaluations at many displaced configurations, are also computationally tractable with MLIPs and generally produce reasonable results for thermodynamically stable phases within the training distribution. For layered oxide cathode materials specifically, voltage estimates derived from MLIP formation energies align reasonably with DFT-calculated values for Li intercalation compounds, provided the model has sufficient training coverage of that composition space. The operative phrase is "sufficient coverage."
Where MLIPs Break
The failure modes fall into three categories: extrapolation to unseen chemistry, high-temperature and high-pressure regimes, and chemistries with strong electronic correlation effects that the underlying DFT functional handles poorly.
Extrapolation to Unseen Chemistry
An MLIP has no mechanism for recognizing that it is outside its training domain. If you submit a candidate structure with an element combination or coordination geometry not represented in the training data, the model will produce a number. That number may be wrong by hundreds of meV/atom, and nothing in the model output signals that this is happening. This is the most practically dangerous failure mode in screening workflows because it is silent. The model outputs a plausible-looking formation energy, the structure passes the stability filter, and synthesis resources are spent on a candidate whose predicted stability was a complete artifact of extrapolation.
We handle this by computing a training-set coverage score for each candidate before relying on the MLIP prediction. The coverage score is based on distance in descriptor space between the candidate and its nearest neighbors in the training set. Candidates with low coverage scores are either routed for DFT validation before synthesis dispatch, or excluded from the ranked output entirely with an explicit flag in the results. Neither approach eliminates the problem, but both make the failure mode visible rather than hidden.
High-Temperature Synthesis Regimes
Most MLIP training datasets are built from DFT calculations at or near ground-state conditions. A model trained primarily on 0 K structures will underestimate entropy contributions and will often mislabel thermodynamically unstable ground-state structures as preferable when the actual synthesis target is a high-temperature phase stabilized by configurational entropy or kinetic trapping during quench. For solid-state synthesis routes where target phases form at 700 to 1000 degrees Celsius and are quenched to room temperature, the 0 K convex hull prediction is a poor proxy for what the synthesis actually produces.
We partially address this by including finite-temperature DFT snapshots from ab initio molecular dynamics trajectories in the training set for phase families where this is known to matter. This is a targeted mitigation for specific chemistries, not a general solution. For any material class where the synthesis temperature is far from 0 K and the equilibrium phase diagram changes significantly with temperature, this limitation needs to be stated explicitly when reporting screening results.
Electronic Correlation Failures
Standard DFT with GGA exchange-correlation functionals, which provides the majority of public training data for inorganic MLIPs, systematically underpredicts band gaps and can misorder phase stability for transition metal oxides with strong d-electron correlation. An MLIP trained on GGA data inherits these errors directly. For Mn- and Ni-containing cathode oxides, this means the model may predict incorrect phase stability ordering for specific oxidation state configurations. GGA+U corrects part of this, but the U parameter value varies by element and by local chemical context, and is not consistently applied across public databases. The consequence is that MLIP predictions for transition metal oxide stability should be treated as indicative, not definitive, unless the model training data explicitly includes Hubbard-corrected calculations for the relevant chemistry.
How We Design Around These Failure Modes in Practice
Our production screening workflow layers three checks on top of raw MLIP output. First, training-set coverage scoring is applied to every candidate before the stability filter runs. Second, calibration offsets by phase family are applied where we have systematic DFT validation data: for each phase family we maintain a held-out test set of DFT-calculated formation energies against which we evaluate model error, and we apply a correction if a consistent bias is present. Third, uncertainty flags are attached to low-coverage and high-variance predictions in the synthesis dispatch payload, so the synthesis team has visibility into which candidates are high-confidence predictions versus genuine exploration into uncertain chemistry.
None of these is a substitute for DFT validation on candidates you are seriously considering for synthesis. What they do is allow you to run a large screening pass efficiently and concentrate your DFT compute budget on the candidates most likely to be both stable and promising. For a team running 100 DFT calculations per week as validation budget, routing those calculations toward the top-ranked high-coverage candidates is substantially more productive than validating randomly or validating every candidate that passes the MLIP stability filter.
The science page describes how we benchmark our MLIP models against held-out DFT data and what error bounds we report for each property target. The platform page covers how the coverage scoring and uncertainty flags are integrated into the screening output. If you are evaluating a screening workflow for your specific material class and want to understand how our model performs on that chemistry, reach out to the team and we can walk through the relevant benchmarks.