Universal ML potentials accurate on bond breaking, transition states and high-energy configurations
Difficulty 48: the failure mode is diagnosed and fine-tuning or more diverse data largely fixes it in tests, but 1 kcal/mol barrier accuracy across chemistries is not shown, and reference-functional error caps the target. Application value 64: reactive simulation, catalysis and battery interphases. Leverage 50: it changes how foundation potentials are built and benchmarked.
The problem
Universal machine-learned potentials tend to be too soft away from equilibrium, so reaction barriers come out wrong.
- 1Training sets show relaxed structures
- 2Strained bonds are rare in data
- 3Model too soft on hills
- 4Rates are exponentially sensitive
- 5More data or fine-tuning helps
Observation. Deng et al. showed that universal MLIPs (M3GNet, CHGNet, MACE-MP-0) underpredict the curvature of the potential energy surface: energies and forces on surfaces, defects, phonons, migration barriers and high-energy states are systematically too low by a roughly constant factor. They attribute this to training sets dominated by near-equilibrium configurations; fine-tuning with a modest number of high-energy structures largely removes it.
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- Machine-learned interatomic potentials with calibrated, guaranteed error bounds outside the training distributionComputational Chemistry · D60
- Neural-network wavefunctions at chemical accuracy for 100+ electron molecules with transferable, amortised trainingComputational Chemistry · D54
- Step 2
- Step 3
- A transferable exchange–correlation functional free of delocalization and self-interaction errorComputational Chemistry · D72
- Then
- Universal ML potentials accurate on bond breaking, transition states and high-energy configurationsThis problem
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Step 2
- Step 3
- Machine-learned interatomic potentials and foundation potentialsAlgorithm · host
- Then
- Universal ML potentials accurate on bond breaking, transition states and high-energy configurationsThe problem
State of the art
- Softening. Deng et al. (npj Comput. Mater. 2025; arXiv:2405.07105) show that M3GNet, CHGNet and MACE-MP-0 systematically underpredict curvature on surfaces, defects, migration barriers, phonons and high-energy states, which they trace to pre-training data dominated by near-equilibrium structures; modest fine-tuning on high-energy data largely corrects it.
- Larger and more diverse data. OMol25 (over 100 million ωB97M-V/def2-TZVPD calculations, Meta FAIR, arXiv:2505.08762) and UMA (arXiv:2506.23971) include reactive and distorted structures, and eSEN/UMA report strong results on reaction and conformer benchmarks. MatPES-type datasets and pressure-transfer benchmarks target the materials side; a WBM-MatPES pressure benchmark preprint appeared in October 2026 (arXiv:2610.14891).
- Reaction data. Transition1x (Schreiner et al., Sci. Data 2022) provides about 10 million DFT calculations along nudged-elastic-band paths for small organic reactions; generalisation of models trained on it to reaction classes outside the dataset's chemistry is limited.
- Open. To this entry's knowledge, no universal model has been demonstrated, on independent held-out benchmarks, at barrier-height errors near 1 kcal/mol across organic, organometallic and heterogeneous reaction classes, and reference functional error itself (several kcal/mol for DFT barriers) limits what "accuracy" means.
Realistic applications
Reactive MD for combustion, atmospheric and catalytic chemistry, battery interphase formation and mechanochemistry needs barriers accurate enough for rates (1 kcal/mol is about a factor of 5 at room temperature); transition-state searches and automated reaction-network exploration with MLIPs depend on it.
What counts as solved
A model with released weights evaluated on a pre-registered benchmark of at least 10,000 reactions and a materials benchmark (migration barriers, surface energies, high-pressure equations of state), achieving the stated error levels against the reference method in independent evaluation, plus comparison against fine-tuned baselines.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
A model with released weights must be evaluated on a pre-registered benchmark of at least 10,000 reactions and a materials benchmark (migration barriers, surface energies, high-pressure equations of state) and meet the stated error levels against the reference method in independent evaluation, with fine-tuned baselines compared.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Mainly data coverage and training at scale, plus reference-method accuracy for barriers.
Bottlenecks
ranked by votes- #1EmpiricalHigh-energy and transition-state training data
Near-equilibrium datasets cause softening; transition-state data at hybrid or better accuracy is scarce.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Oct 10, 2026Pressure-transfer benchmark (preprint)source
De Breuck et al. introduce WBM-MatPES and Tuga-FF.
- May 13, 2025OMol25source
A very large molecular dataset including distorted and reactive structures.
- May 12, 2024Softening identifiedsource
Deng et al. document systematic PES softening in universal potentials (arXiv date).
- Dec 1, 2022Transition1xsource
A dataset of reaction pathways for training reactive ML potentials (date approximate).