Moerman, Kabylda, and Khabibrakhmanov (2602.22086) train a neural network to predict the inputs to many-body dispersion calculations — not to replace the physics, but to make the physics affordable.
Note (2026-11-15; read against the paper's full text): the essay says MBD-ML inherits the physical guarantees of the many-body method and gives DFT-quality dispersion at a fraction of the cost for systems too large for DFT. The paper says the many-body dispersion method scales as O(N^3), like Kohn-Sham DFT, with a smaller prefactor, and that MBD-ML removes the electronic-structure step; I did not find the word guarantee applied to inherited properties. It also reports limits the essay leaves out: a per-atom error of 0.8 meV can give a total-energy deviation above 80 meV in cells of over 100 atoms (the errors largely cancel in energy differences), and on a random subset of 203 materials from the OMat24 dataset the mean absolute errors of the predicted polarizability and C6 ratios (0.247 and 0.269) are 10 to 20 times higher than for the other chemical classes examined. The model is trained against MBD-NL at the density-functional level as reference.
Van der Waals interactions — the weak attractions between all molecules — matter everywhere: drug binding, crystal packing, surface adsorption, battery materials. The many-body dispersion (MBD) method captures these interactions accurately by computing how each atom's polarizability is screened by its neighbors, producing a coupled system where the whole is not the sum of pairwise parts. The problem: MBD requires atomic C_6 coefficients and polarizabilities as input, and computing these from first principles (typically density functional theory) is expensive.
MBD-ML replaces the expensive DFT step with a pretrained message-passing neural network. Given only atomic positions and species, it predicts the C_6 coefficients and polarizabilities that MBD needs. The physics of the many-body calculation itself remains exact — only the inputs are approximated. This is a different strategy from end-to-end learned force fields that replace the entire energy calculation with a neural network. By keeping the physics and only learning the inputs, MBD-ML inherits MBD's physical guarantees: correct long-range behavior, proper many-body screening, and the right scaling with system size.
The practical value is speed: DFT-quality van der Waals interactions at a fraction of the computational cost, applicable to systems too large for DFT. The philosophical value is in the division of labor — the neural network handles the part that's expensive but smooth (mapping structure to atomic properties), while the physics handles the part that's structured but cheap (computing interactions from those properties). Each component does what it's good at.