MBD-ML: ML for Many-Body Dispersion

• MBD-ML is a machine learning approach that predicts atomic static polarizabilities and C6 coefficients directly from molecular structures.
• It integrates with libMBD to compute dispersion corrections in energies, forces, and stresses, bypassing traditional electronic-structure calculations.
• Benchmark results show high accuracy and efficiency across diverse systems, highlighting practical improvements in many-body dispersion modeling.

Searching arXiv for the primary paper and closely related context papers on many-body dispersion and the SO3krates/SO3LR framework. Tool call: arxiv_search(query="(Moerman et al., 25 Feb 2026) MBD-ML Many-body dispersion from machine learning for molecules and materials", max_results=5) MBD-ML denotes a machine-learning realization of many-body dispersion (MBD) theory in which a pretrained message passing neural network predicts the atomic static polarizabilities and atomic $C_6$ coefficients required by the MBD formalism directly from atomic structure. The method is designed for molecules and materials, and its stated purpose is to make van der Waals interactions immediately available in total energies, forces, and stress tensors through integration with libMBD, without intermediate electronic-structure calculations (Moerman et al., 25 Feb 2026). In this sense, MBD-ML is a density-free route to MBD-inclusive modeling that preserves the underlying many-body dispersion machinery while replacing the usual upstream determination of atomic response parameters.

1. Conceptual setting

Van der Waals interactions are described as essential for molecules and materials, including applications ranging from drug design and catalysis to batteries. Within this landscape, the MBD method is identified as one of the most accurate and transferable approaches for capturing vdW interactions, provided that atomic $C_6$ coefficients and polarizabilities are available as input (Moerman et al., 25 Feb 2026).

MBD-ML targets precisely this input bottleneck. Rather than deriving atomic response properties from an electronic-structure pipeline, it predicts them from structure using a pretrained equivariant message passing network. The resulting workflow is explicitly presented as compatible not only with electronic-structure codes, but also with empirical force fields and machine-learned force fields. A plausible implication is that MBD-ML should be understood less as an alternative dispersion theory than as an interface layer between structural representations and the established MBD Hamiltonian.

The method is framed as a continuation of density-free dispersion modeling. Earlier work on DNN-MBD replaced Hirshfeld-type density partitioning by predicting atom-in-molecule volume ratios from local atomic environments, then using those volumes inside MBD@rsSCS (Poier et al., 2022). MBD-ML moves the surrogate target further downstream by predicting atomic polarizabilities and $C_6$ coefficients themselves (Moerman et al., 25 Feb 2026).

2. Many-body dispersion formalism

The underlying physical model is the standard MBD construction in which each atom $i$ is mapped to a quantum Drude oscillator characterized by a static polarizability $\alpha_{0,i}$ and an effective excitation frequency. The coupled system is represented by a quadratic Hamiltonian with long-range dipole-dipole coupling tensor $T_{ij}^{\rm lr}$ and short-range damping controlled by $\beta$ (Moerman et al., 25 Feb 2026).

Because the Hamiltonian is quadratic, the MBD correlation energy can be obtained by diagonalizing the $3N \times 3N$ dynamical matrix and summing the resulting collective eigenfrequencies. The equivalent ACFD-RPA expression is given as

$$E_{\rm MBD} = \frac{\hbar}{2\pi}\int_0^\infty d\omega\; \mathrm{Tr}\!\left[ \ln\!\left(1-\mathbf{A}(i\omega)\,\mathbf{T}^{\rm lr}\right) \right],$$

with block-diagonal atomic response

$$\mathbf{A}_{ii}(i\omega)=\alpha_i(i\omega)\,I_{3\times 3}, \qquad \alpha_i(i\omega)=\frac{\alpha_{0,i}}{1+(\omega/\omega_i)^2}.$$

Within this formulation, the central learned quantities are therefore the atom-resolved $C_6$0 and $C_6$1 that parameterize the oscillator model and its frequency dependence (Moerman et al., 25 Feb 2026).

This division of labor is important. MBD-ML does not replace the MBD energy expression, the dynamical-matrix diagonalization, or the computation of forces and stresses in libMBD. It supplies the atomic response data on which those steps depend. That architectural choice explains why the method is presented as transferable across molecules and materials while remaining tied to a physically explicit dispersion theory.

3. Neural-network architecture and predicted quantities

MBD-ML is built on the SO3krates/SO3LR framework. Its node-level inputs include atomic number $C_6$2, total charge $C_6$3, spin multiplicity $C_6$4, and radial and angular neighbor signals within a cutoff $C_6$5 (Moerman et al., 25 Feb 2026).

The network uses two rounds of equivariant message passing with a spherical-harmonic basis up to $C_6$6 and learned radial expansions. Neighbor information is mixed by an attention-like update that preserves $C_6$7 equivariance, and nonlinear transformations are implemented through gated equivariant blocks. The model therefore belongs to the class of geometric message passing architectures that preserve rotational and translational structure in the representation.

The output layer is split into two MLP heads that predict unit-less atomic ratios,

$C_6$8

from which the quantities used in MBD are reconstructed as

$C_6$9

No explicit clipping is enforced, and the reported behavior is that training naturally confines $C_6$0 and $C_6$1 to approximately $C_6$2 (Moerman et al., 25 Feb 2026).

This design makes the model explicitly atomic and response-centric. Instead of regressing a total dispersion correction directly, it predicts physically interpretable local inputs to the many-body formalism. A plausible implication is that the method can inherit some of the compositional and extensible properties of atom-based response models, provided that the target chemical environments are represented in training.

4. Training data, optimization, and software workflow

The training set is QCML, comprising 33.5 million single-molecule geometries with up to 8 non-hydrogen atoms and spanning 79 elements. The reference labels are $C_6$3 computed at PBE0+MBD-NL, from which the target ratios $C_6$4 are formed (Moerman et al., 25 Feb 2026).

Optimization uses mean-squared error on $C_6$5 and $C_6$6 without weighting between the two heads, together with Adam at initial learning rate $C_6$7, exponential decay, and early stopping on a hold-out QCML subset. On the QCML test set of $C_6$8 molecules, the reported prediction errors are:

• MAE$C_6$9, RMSE$i$0
• MAE$i$1, RMSE$i$2

These response-level errors propagate to MBD property errors of MAE$i$3 meV/atom and RMSE$i$4 meV/atom for $i$5, MAE$i$6 meV/\AA\ and RMSE$i$7 meV/\AA\ for $i$8, and RMSE$i$9 meV/\AA$\alpha_{0,i}$0 for $\alpha_{0,i}$1 (Moerman et al., 25 Feb 2026).

Integration is carried out through libMBD, exposed through the pymbd Python interface. The reported workflow is: the user provides positions $\alpha_{0,i}$2, species $\alpha_{0,i}$3, the cell for periodic systems, and total $\alpha_{0,i}$4; MBD-ML returns $\alpha_{0,i}$5; libMBD multiplies by free-atom references, constructs the MBD Hamiltonian, diagonalizes it, and computes energy, forces, and stress (Moerman et al., 25 Feb 2026). The practical significance is that the learned component is modular: the neural model predicts atomic response parameters, while libMBD remains responsible for the physics-based postprocessing.

5. Benchmark accuracy and computational scaling

The reported evaluation extends beyond QCML to noncovalent dimers, organic crystals, polymorph rankings, and large molecules. The benchmark numbers are as follows (Moerman et al., 25 Feb 2026):

System	Reported error	Note
QCML test set	$\alpha_{0,i}$6 RMSE $\alpha_{0,i}$7 meV/atom	$\alpha_{0,i}$8 molecules
DES370k	$\alpha_{0,i}$9 RMSE $T_{ij}^{\rm lr}$0 meV/atom	noncovalent dimers
DES370k	$T_{ij}^{\rm lr}$1 RMSE $T_{ij}^{\rm lr}$2 meV/\AA	noncovalent dimers
OMC25	$T_{ij}^{\rm lr}$3 RMSE $T_{ij}^{\rm lr}$4 meV/atom	organic crystals
OMC25	$T_{ij}^{\rm lr}$5 RMSE $T_{ij}^{\rm lr}$6 meV/\AA	organic crystals
OMC25	$T_{ij}^{\rm lr}$7 RMSE $T_{ij}^{\rm lr}$8 meV/\AA$T_{ij}^{\rm lr}$9	organic crystals

For polymorph rankings, seven of nine transitions are reported as correctly predicted within 0.3 kJ/mol per molecule. For OMol25 systems up to 350 atoms, the force-error median is given as approximately 2–3% versus MBD-NL, with angular error around $\beta$0 (Moerman et al., 25 Feb 2026).

The scaling study uses $\beta$1 clusters up to $\beta$2, corresponding to approximately 13,000 atoms, on a single 128-core node. Ratio prediction by the ML model is nearly constant time at $\beta$3 s up to 1000 atoms, then follows approximately $\beta$4 scaling and reaches roughly 200 s at 4321 molecules. By contrast, the libMBD energy/force stage scales approximately as $\beta$5 beyond 60 atoms and crosses over with the ML stage near 250 atoms. For large systems, the total MBD-ML workflow is reported as approximately $\beta$6 the cost of a single DFT SCF cycle (Moerman et al., 25 Feb 2026).

These figures indicate that the principal acceleration does not arise from changing the asymptotic cost of the MBD diagonalization itself, but from eliminating the upstream electronic-structure step used to obtain atomic response parameters. This suggests that MBD-ML is especially attractive in workflows where dispersion corrections must be evaluated repeatedly or embedded inside force calculations.

6. Chemical scope, limitations, and relation to earlier density-free MBD

The covered chemical space is defined by training on 79 elements in small molecules, with successful tests on organic, biomolecular, and molecular-crystal data sets (Moerman et al., 25 Feb 2026). This reported transfer across system classes is central to the method’s positioning as a pretrained component for both molecules and materials.

Two limitations are stated explicitly. First, molecular anions are excluded because electronically unbound states lead to unreliable reference ratios; the recommendation is to exclude anions until reliable ab initio benchmarks exist. Second, alkali and alkaline-earth elements together with inorganic solids are under-represented in QCML, leading to large errors of $\beta$7 and occasional negative MBD eigenvalues. The proposed remedy is to expand the training data toward periodic and inorganic environments (Moerman et al., 25 Feb 2026).

The relation to earlier density-free MBD clarifies the methodological shift:

Approach	Learned quantity	Reported scope
DNN-MBD (Poier et al., 2022)	atom-in-molecule volume ratios $\beta$8	ANI-1x restricted to C,H,N,O
MBD-ML (Moerman et al., 25 Feb 2026)	atomic $\beta$9 and $3N \times 3N$0 ratios	QCML, 79 elements

DNN-MBD couples DFT to MBD by predicting volume ratios from Behler-Parrinello-style local descriptors and then using standard TS-style rescaling and MBD@rsSCS (Poier et al., 2022). MBD-ML instead predicts the atomic response quantities needed by the MBD Hamiltonian directly, uses an equivariant message-passing model, and exposes a libMBD-based path to energies, forces, and stress tensors (Moerman et al., 25 Feb 2026). A plausible implication is that MBD-ML reduces the number of intermediate modeled quantities and makes the learned component more directly aligned with the MBD solver interface.

Potential extensions identified for MBD-ML include retraining or fine-tuning on periodic solids, such as OMat24, and the incorporation of explicit charge-state features or specialized anion and electron-affinity treatments (Moerman et al., 25 Feb 2026). In that sense, the method occupies an intermediate position between pure first-principles dispersion correction and end-to-end machine-learned force fields: it preserves an explicit many-body response model while using ML to supply the atomic inputs that had previously required separate electronic-structure computation.