Transferable ML potentials with non-local charge transfer, variable charge/spin states and long-range electrostatics
Difficulty 52: several architectures exist, but transfer across charge states and to interfaces at scale needs data and validation not yet available. Application value 56: electrolytes, electrochemistry and biomolecules. Leverage 45.5: it extends one method class and enables a subfield, not a discipline.
The problem
Most machine-learned potentials only see nearby atoms, so they miss forces between distant charges and charge moving between atoms.
- 1Local models see only nearby atoms
- 2Charges act at long range
- 3Electrons can move far
- 4Add learned charges plus Ewald
- 5Need broad, transferable tests
Limitation. A local MLIP writes with 5-6 Å. Coulomb interactions decay as , and polarisation, charge transfer and applied potentials act over larger distances, so cutoff models cannot reproduce them in principle.
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
- 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
- Step 4
- Transferable ML potentials with non-local charge transfer, variable charge/spin states and long-range electrostaticsThis problem
- Then
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
- Transferable ML potentials with non-local charge transfer, variable charge/spin states and long-range electrostaticsThe problem
State of the art
- Charge-equilibration and fourth generation. Ko et al. (Nat. Commun. 2021; arXiv:2009.06484) introduce a fourth-generation high-dimensional neural network potential with non-local charge transfer; Kocer et al. (arXiv:2502.07907) extend with iterative charge equilibration. SpookyNet (Unke et al., Nat. Commun. 2021) includes explicit charge, spin and nonlocal interactions.
- Long-range kernels. Latent Ewald summation (Cheng, arXiv:2408.15165) learns latent charges and evaluates them by Ewald, avoiding supervised charge labels; Gao, Yam and Mao (arXiv:2410.13820) report a foundation-style potential with polarisable long-range interactions for materials.
- Foundation models. UMA (arXiv:2506.23971) accepts total charge and spin as conditioning information but is built on a local, cutoff-based message-passing architecture, so explicit long-range effects beyond the cutoff are not handled by construction; the demo notes the training data are aperiodic and periodic systems should be treated cautiously.
- Gap. Evidence for transferable accuracy on electrolytes and electrified interfaces at scale, with non-local charge transfer correct across charge states, is limited to specific systems.
Realistic applications
Battery electrolytes and interphases, supercapacitors, electrocatalysis under potential, ionic liquids, proteins with charged residues and membranes all depend on long-range electrostatics and charge transfer; current local MLIPs cannot represent them faithfully.
What counts as solved
Publicly released model(s) achieving the stated accuracy on a pre-registered benchmark set of electrolytes, charge-transfer complexes and electrochemical interfaces (including an out-of-element test), with demonstrated linear-scaling inference on at least atoms, verified independently.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
Publicly released models must reach the stated accuracy on a pre-registered benchmark of electrolytes, charge-transfer complexes and electrochemical interfaces (including an out-of-element test), with linear-scaling inference demonstrated on at least 1e5 atoms and independent verification.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
New architectures plus implementation for linear-scaling long-range terms and systematic data.
Bottlenecks
ranked by votes- #1EmpiricalTraining data with charge transfer and applied potential
Datasets for electrolytes, interfaces and variable charge states are scarce and expensive.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Jun 30, 2025UMA conditioned on charge and spinsource
Universal models trained on OMol25 take charge and spin as inputs (arXiv date).
- Aug 27, 2024Latent Ewald summationsource
Learned latent charges with Ewald evaluation (arXiv date).
- Dec 1, 2021SpookyNetsource
Explicit charge, spin and nonlocal interactions (date approximate).
- Jan 20, 20214G-HDNNPsource
Non-local charge transfer introduced into neural-network potentials.