Atomistic Materials
Pretrain crystal potentials from scratch across eight physics suites.
Task & evaluation
- Objective
- Learn one model of atomic interactions that works across many physical settings.
- Input
- Elements, atomic positions and the repeating cell.
- Output
- Energy, forces and response to deformation.
- Data
- About 42 million structures labeled by quantum calculations.
- Evaluation
- Eight physical tests, from crystal stability to vibrations and ion migration.
- Research challenge
- Remain accurate when atoms move away from their resting arrangements.
Understanding the task
Follow one learned energy landscape from a stretching bond to the many physical properties it must explain.
Can one model describe many kinds of atomic motion?
Atoms in a material pull and push on one another. An interatomic potential predicts the energy of an arrangement and the forces that make it change. A fast, accurate potential can help simulate vibrations, deformation, heat transport and ion motion without repeating an expensive quantum calculation at every step.
The hard part is reliability away from familiar structures. Two models can agree near a crystal’s resting arrangement yet disagree when a bond stretches or an ion squeezes through a narrow passage. The interactive example shows how matching the bottom of an energy well can still miss the physics farther away.
INTERACTIVE · ONE LANDSCAPE, MANY PROBES
Same minimum. Different physics.
A harmonic fit and an anharmonic reference agree in energy, force and curvature at equilibrium. Move the atoms, then strain the cell to see where that agreement ends.
— Anharmonic reference┄ Candidate
1 · Distance → energy → force
Stretch a bond. The force on the right atom is F = −dE/dr; arrows point in the direction it would move.
2 · Cell strain → energy → stress
Change the horizontal cell length at fixed fractional coordinates. Every horizontal bond stretches together.
What this connects to in the task
Stability probes relaxed energies; phonons probe local curvature; conductivity needs higher-order force constants. Disorder, cleavage and migration visit other parts of the landscape. All eight suites reuse the same ASE calculator: total energy (eV), forces (eV/Å) and stress (eV/ų, positive tensile). The evaluator chooses the physical perturbations.
Synthetic illustration, not benchmark data or a materials simulator. Here r₀ = D = 1, Eref = (1 − exp[−2(r − 1)])² − 1 and Efit = 4(r − 1)² − 1. The cell is a 2D nearest-neighbor sketch; its dimensionless stress proxy is dEbond/dε, not a calculated 3D material stress. Controls explore a toy model, not alternative benchmark settings.
HOW THE REFERENCE MODEL WORKS
One learned energy surface, reused by every protocol
- 1 · StructureElements, atomic positions and repeating cell
- 2 · Learned representationLearn relationships between neighboring atoms
- 3 · Energy predictionThe energy surface determines forces and response to deformation
- 4 · Shared predictionsEnergy, forces and stress from one model
- 5 · Evaluator-controlled physicsRelaxations, displacements, strain, volume scans, surfaces and migration paths
Forces follow the slope of the energy surface. Matching a few energy values does not guarantee that its slopes and curvature are correct.
SCORING LAB · INVENTED FOUR-PATH EXAMPLE
A failed migration path still contributes error
Three ion-migration paths have absolute barrier errors of 0.10, 0.20 and 0.30 eV. Set the fourth path’s error, or mark it failed. The documented protocol caps each path’s error at 2 eV and assigns failures 2 eV.
This is arithmetic teaching data, not a measured material or actual NEB run. An unconverged but finite path is separate from an exception: the evaluator additionally reports converged-only error and convergence fraction. The eight-suite task has no aggregate score.
What the model learns from
The training pool contains about 42 million atomic configurations from several quantum-calculation datasets. Some follow crystals relaxing toward equilibrium; others deliberately displace atoms or simulate their motion. Each example supplies an arrangement, its energy, forces and stress—the internal response to deforming the cell.
More data does not automatically mean better supervision. Different calculations can use different energy references and approximations. The task separates evaluation materials from training, with structure-level separation for several amorphous systems whose compositions remain represented.
The results below concern pretrained UMA models and subsequent adaptation. They illustrate physical tradeoffs but do not establish performance under the newer requirement to learn from scratch.
Technical details
| Source | Frames | Physical coverage and label caveat |
|---|---|---|
| MPtrj | 1,065,435 | Crystal relaxation trajectories; MP PBE/PBE+U, 520 eV |
| sAlex | 9,434,535 | Alexandria subsample; MP-compatible PBE/PBE+U |
| MatPES-PBE 2025.1 | 193,171 | DIRECT-sampled MD, 85 elements; PBE without +U, 680 eV |
| OMat24 | 31,716,685 | Rattled, rattled-relax and AIMD; different pseudopotentials and energy offsets |
| Total | 42,409,826 | 91 GB; a shared pool, not a requirement to train on every frame |
Eight views of the same landscape
Stability screening asks which structures resist decomposition. Vibrations probe local curvature; elasticity probes response to strain. Heat conductivity depends on more complex interactions between vibrations. Flexible frameworks, disordered solids, cleaved surfaces and ion migration explore still other regions of the energy landscape.
The task reports fifteen physical metrics rather than averaging incompatible units into one score. A model can improve discovery accuracy while making dynamics worse. Failed calculations also matter: an accurate answer on the easy cases is insufficient if difficult simulations break.
Technical details
| Suite | Validation / test | What is measured |
|---|---|---|
| Stability screening | 1,000 / 5,000 | Stable/unstable F1; hull-distance MAE (meV/atom) |
| Phonons | 300 / 500 | Maximum and mean frequency (K); entropy/heat capacity (J/K/mol); free energy (kJ/mol) |
| Elasticity | 300 / 500 | Bulk and shear modulus MAE (GPa) |
| Thermal conductivity | 40 / 63 | Conductivity SRME at 300 K |
| MOF mechanics | 40 / 60 | Bulk-modulus MAPE; evaluator adds D3(BJ) |
| Amorphous solids | 376 / 558 structures | Energy and force relative RMSE, averaged over five systems |
| Surfaces | 2,000 / 2,000 slabs | Cleavage relative MAE; disjoint bulk materials |
| Ion migration | 256 / 390 paths | Barrier MAE (eV); all paths, converged paths and convergence fraction |
Learning an energy landscape
The reference model learns an energy surface over atomic arrangements. Its slope gives forces: atoms tend to move downhill. How that slope changes determines vibrations and other responses. Accurate energies at a few points therefore do not guarantee accurate motion between them.
One experiment adapts the model to Materials Project data near familiar crystals. Another replaces half the training examples with displaced structures from OMat24. For those additional examples it learns from forces and stress, avoiding incompatible energy references. The idea is to retain knowledge of configurations that atoms visit away from equilibrium.
Every evaluation uses the same potential. Researchers cannot choose a different model for each physical property, making this a test of broad reliability.
A clear tradeoff, not a universal improvement
Adapting UMA to Materials Project data raises stability-screening F1—a combined measure of finding stable materials and avoiding false alarms—from 0.54 to 0.86, but increases several physical errors. Adding displaced structures partly repairs heat-conductivity and framework-mechanics performance. It does not solve the whole problem.
For example, the mixed-data model lowers conductivity error by about 11% relative to the Materials Project adaptation, while increasing the maximum-vibration-frequency error by about 18%. Better performance in one part of the landscape can coexist with worse local curvature elsewhere.
The original Materials Project adaptation included a small overlap with later evaluation materials: about 0.2% of training frames. Treat it as a historical diagnostic, not a clean current-protocol baseline. Multiple training seeds and a from-scratch comparison are still missing.
Detailed physical results
| Metric | Original UMA | Crystal-focused training | Mixed-data training |
|---|---|---|---|
| Discovery F1 | 0.54 | 0.86 | 0.86 |
| Phonon max-frequency MAE, K | 7.9 | 47.1 | 55.5 |
| Conductivity SRME | 0.28 | 0.44 | 0.39 |
| MOF modulus MAPE | 25.2% | 25.0% | 21.4% |
| AM26 force relative RMSE | 11.8% | 14.7% | 14.1% |
| NEB barrier MAE, eV | 0.126 | 0.126 | 0.134 |
Compute and resources
Materials adaptation
Reported measurement
- Time
- 2 h 16 min
- Hardware
- 8 × H100
- GPU allocation
- ≈18.1 GPU-h
Training on material configurations.
Elapsed time by stage
- Materials adaptation
- 136 min
No finer stage timing is reported for this workload. Separate cards describe separate measurements, not parts of one total.
Original pretraining is excluded.
Mixed-data adaptation
Reported measurement
- Time
- 2 h 17 min
- Hardware
- 8 × H100
- GPU allocation
- ≈18.3 GPU-h
Training with a broader mixture of atomic environments.
Elapsed time by stage
- Mixed-data adaptation
- 137 min
No finer stage timing is reported for this workload. Separate cards describe separate measurements, not parts of one total.
A separate alternative run, not an additional stage.
Mixed-data preparation
Reported measurement
- Time
- 16 min
- Hardware
- 96 workers
- GPU allocation
- Not applicable (CPU stage)
Preparing the broader training mixture.
Elapsed time by stage
- Data preparation
- 16 min
No finer stage timing is reported for this workload. Separate cards describe separate measurements, not parts of one total.
Full physical evaluation
Reported measurement
- Time
- 6 h 18 min
- Hardware
- 8 × H100, shared with three other evaluation jobs
- GPU allocation
- Not established (shared GPUs)
All eight test suites for the mixed-data model.
Elapsed time by stage
- MOF
- 159 min
- Physical tests
- 104 min
- NEB
- 85 min
- Conductivity
- 28 min
The bar covers the four timed major suites (6 h 16 min). The reported full evaluation is 6 h 18 min; minor suites and rounding are not assigned invented durations. GPUs were shared.
Shared hardware: this timing is not an exclusive GPU-hours measurement. Largest components: MOF 2 h 39 min, physical tests 1 h 44 min, NEB 1 h 25 min and conductivity 28 min.
Learning from the failures
The next question is which atomic arrangements are missing from training. Sampling more distorted structures may help bond breaking and migration, while examples chosen to reveal curvature may help vibrations. Controlled comparisons should check all physical suites: improving force error alone is not enough to claim a better universal potential.
References and provenance
UMA: A Family of Universal Models for Atoms
Introduces the model family used to predict atomic energies and forces across different kinds of matter.
Task definition and experimental evidence
Documents the mixed training data and physical tests used here. These measurements use a different protocol from the public paper results.
Code links require access to the project’s GitHub repository.
Leaderboard
SimulatedDesign preview · leaderboard values and research histories are simulated.
| # | Method | Type | Discovery F1 ↑ | Force MAE ↓ | Energy MAE ↓ | Research time | Run |
|---|---|---|---|---|---|---|---|
| 1 | Agent search | 0.6605 | 0.5405 | 0.6991 | 76 h | ||
| 2 | Agent search | 0.6573 | 0.5431 | 0.7025 | 60 h | ||
| 3 | UMA-S-1.1 (MP + OMat24 mixed replay) | Reference | 0.6550 | 0.5450 | 0.7050 | — | — |
| 4 | UMA-S-1.1 (zero-shot) | Reference | 0.5371 | 0.6646 | 0.8598 | — | — |
How to read these leaderboard metrics
These leaderboard values and research histories are simulated design examples. The leaderboard is ranked by Discovery F1. The research curves show relative improvement in this metric over UMA-S-1.1 (MP + OMat24 mixed replay); only accepted new-best results update the retained trajectory. A positive gain means a better result, with the sign adjusted for metrics where lower is better. The task protocol is described in Task & evaluation above. These preview columns and trajectories are not measured benchmark results and must not be compared with the recorded experiments. The leaderboard and agent-search curves are simulated UI examples, not measured benchmark runs. Its three columns do not represent the documented 15-metric, eight-suite protocol, which has no aggregate score. Historical measured results, units, training eligibility caveats and failure handling appear in Understanding the task; leaderboard baseline labels must not be read as those measurements.
- Discovery F1Primary metric · Higher is better
- Balances precision and recall when identifying stable candidate materials. A high score requires finding useful candidates while limiting false positives.
- Force MAELower is better
- Mean absolute difference between predicted and reference atomic forces. It measures local force accuracy rather than material-discovery performance.
- Energy MAELower is better
- Mean absolute difference between predicted and reference energies. Interpret values using the units and normalization in the task protocol.
AI4AI research progress
Research trajectory
Runtime–performance frontier
How improvement is measured
Relative improvement over UMA-S-1.1 (MP + OMat24 mixed replay) in the primary metric; only retained results update the best-so-far curve.
Accepted new-best experiments raise the trajectory. The frontier includes evaluated candidates; trial, partial and stopped experiments are excluded.