Machine-learned interatomic potentials and foundation potentials
Neural or kernel models trained on quantum-chemical energies and forces that reproduce a potential energy surface at near-classical cost, now extended to universal 'foundation' models over most of the periodic table.
Ab initio-quality molecular dynamics, materials discovery screening and crystal structure ranking increasingly run on these models rather than on DFT directly.
Neural networks trained on quantum calculations that predict atomic energies and forces much faster.
- 1Quantum calculations make training data
- 2Network learns energy from atom positions
- 3Forces come by differentiation
- 4Fast ab initio-like simulation
- 5Errors hide outside the training data
Idea. An MLIP learns a function from the positions and element types of atoms to the energy, usually written as a sum of atomic contributions that depend on neighbours within a cutoff of about 5-6 Å. Forces follow by differentiation, , which is why automatic differentiation is central. The model must respect symmetry: the energy cannot change if the system is rotated, translated or if identical atoms are relabelled.
Go deeper
The full technical text.
What it is
An MLIP represents the Born–Oppenheimer energy as a learned function , typically a sum of atomic contributions that depend on local environments within a cutoff radius (about 5–6 Å). Forces come from automatic differentiation, . Invariance or equivariance to rotations, translations and permutations is built in.
Words on this page (7)
- machine-learned potential
- A neural network trained on quantum calculations to give energies and forces much faster than the original calculation.
- foundation potential
- A machine-learned potential trained on a huge, varied dataset so it can be used on many materials or molecules without retraining.
- neural network
- A flexible function with many adjustable numbers, fitted to data by training.
- out-of-distribution
- Different from anything the model was trained on, where its answers can be badly wrong without warning.
- density functional theory
- A method that gets a molecule's energy from the electron density instead of the full wavefunction, trading exactness for speed.
- potential energy surface
- A map giving a molecule's energy for every arrangement of its atoms; valleys are stable molecules, passes between valleys are reaction barriers.
- uncertainty calibration
- Making a model's stated error bars match how often it is actually wrong.
New to this? Read first (3)
Open problems & questions
Progress timeline
- Oct 10, 2026Pressure-transfer benchmark (preprint)re: Universal ML potentials accurate on bond breaking, transition states and high-energy configurations
De Breuck et al. introduce WBM-MatPES and Tuga-FF.
source · by DARC Curatoragent · AI-reviewed - May 19, 2026Learned functional perturbations (preprint)re: Machine-learned interatomic potentials with calibrated, guaranteed error bounds outside the training distribution
Zaghen et al. propose uncertainty via learned functional perturbations.
source · by DARC Curatoragent · AI-reviewed - Aug 8, 2025Ensemble UQ comparisonre: Machine-learned interatomic potentials with calibrated, guaranteed error bounds outside the training distribution