Machine-Learned Interatomic Potentials

Density functional theory gives you forces accurate enough to trust and scales roughly as \(O(N^3)\) in the number of electrons, which caps routine molecular dynamics at a few hundred atoms for a few tens of picoseconds. Classical force fields scale linearly and are fast enough for millions of atoms, but their fixed functional forms cannot describe bond breaking, and they must be hand-parameterised per chemistry.

A machine-learned interatomic potential aims to take the accuracy of the first and the scaling of the second: train a model to predict the potential energy surface from DFT reference data, then run dynamics on the learned surface.

Symmetry is architecture, not data

The energy of a configuration must be invariant under translation, rotation and permutation of identical atoms. Forces, being \(-\nabla_{\mathbf{r}} E\), must be equivariant: rotate the molecule and the force vectors rotate with it.

Early potentials achieved invariance by hand-designing descriptors (symmetry functions, SOAP) that are invariant by construction, then regressing on them. Modern models build the symmetry into the network. NequIP represents features as geometric tensors indexed by spherical harmonic degree and constrains message passing to the tensor products allowed by \(E(3)\), so equivariance holds exactly at every layer (Batzner et al., Nature Communications 13, 2022, arXiv:2101.03164).

The payoff is data efficiency, and it is dramatic. Because the model never has to learn from data that rotating a molecule does not change its energy, NequIP reached competitive accuracy on training sets orders of magnitude smaller than earlier invariant-descriptor models needed. Symmetry you enforce is symmetry you do not have to pay for in samples.

MACE extended this with higher body-order messages: constructing four-body messages via the Atomic Cluster Expansion means two message-passing layers suffice where earlier models needed five or six, which is both more accurate and far more parallelisable (Batatia et al., NeurIPS 2022, arXiv:2206.07697).

Energy first, forces by gradient

A design choice worth naming: these models predict a scalar energy as a sum of atomic contributions, and obtain forces by differentiating that energy with respect to atomic positions, rather than predicting force vectors directly.

This guarantees the force field is conservative, so energy is conserved in an NVE simulation. A model that regresses forces independently has no such guarantee and will heat or cool a simulation spuriously. It costs a backward pass per force evaluation, and it is worth it.

Foundation potentials

The current direction is universal models trained across the periodic table on large DFT corpora, then used zero-shot or fine-tuned per system. MACE-MP-0 was trained on Materials Project relaxation trajectories and applied across a broad range of materials without per-system fitting. The same shift underpinned large-scale screening: GNoME used graph networks to predict stability and reported 2.2 million new crystal structures, of which more than 380,000 were assessed as stable (Merchant et al., Nature, 2023, Scaling deep learning for materials discovery).

That number came with a caution attached from the materials community: predicted thermodynamic stability against a computed convex hull is not synthesisability, and a large fraction of the entries were later argued to be duplicates, trivial substitutions, or already-known compounds. The screening is real and the count is a screening count.

When it breaks

Extrapolation is invisible. The model is an interpolator over its training manifold. Push it to a transition state, a high-pressure phase, or a chemistry absent from training and it produces a smooth, confident, wrong energy. Uncertainty estimation via ensembles is the standard mitigation and it is not free.

The reference method is the ceiling. A potential trained on PBE-level DFT reproduces PBE, including PBE's known errors on van der Waals interactions and band gaps. "Near-quantum accuracy" means near the accuracy of the specific functional used, not near the accuracy of nature.

Errors compound over a trajectory. Small force errors integrate. A potential with excellent held-out force RMSE can still produce an unphysical radial distribution function after a nanosecond, which is why validation on dynamical observables rather than on force error is essential.

Speed comparisons are configuration-dependent. Speedups over DFT of four to six orders of magnitude are commonly reported and depend entirely on system size, cutoff radius and hardware. The scaling argument (linear versus cubic) is the durable claim; the specific multiplier is not.