BoostMD

Machine learning force fields have closed most of the accuracy gap with quantum chemistry, but they remain far slower than the classical force fields they aim to replace. Molecular dynamics advances in femtosecond steps, so capturing a conformational change in a peptide can require hundreds of millions of force evaluations — arithmetic that keeps these models out of routine biomolecular simulation.

BoostMD, from InstaDeep and the University of Cambridge, attacks the cost rather than the accuracy. Its starting observation is that the expensive part of an equivariant force field is not the energy readout but the construction of many-body, atom-centered node features, and that those features barely change from one MD step to the next — they oscillate with minimal drift unless the molecular geometry changes qualitatively. Rather than recomputing them from scratch at every step, BoostMD caches the node features from a recent reference evaluation and predicts only the change in energy that follows from atoms having moved since.

That reframing shrinks the learning task enough that a much smaller network suffices. In a simulation, the full reference force field is evaluated every N steps while BoostMD handles the intermediate ones, so the expensive model amortizes across a whole window. Classical MD has long used multiple-time-step integration for the same purpose, but BoostMD's speedup does not depend on separating short- from long-range interactions, so it composes with those schemes rather than competing with them.

#Key Features

• Feature reuse across time steps: The model consumes the current positions, the reference positions, and the reference node features of neighbors within a cutoff, and predicts a per-atom energy difference relative to the reference configuration.
• Kabsch-aligned local reference frame: Each atom's environment is re-centered and rotated by the optimal Kabsch rotation between reference and current neighbor positions, which is what makes the prediction translationally and rotationally equivariant.
• Shallow by design: All reported results use single-layer models built from MACE product-basis blocks, giving a 5 Å receptive field — small enough to parallelize across GPUs by domain decomposition.
• Correct equilibrium sampling: Driving metadynamics with BoostMD reproduces the reference model's free energy surface, so the accelerated trajectory samples the ground-truth Boltzmann distribution rather than merely tracking forces.
• Two inference topologies: The reference and surrogate can alternate serially on one GPU, or run as a producer/consumer pair on separate GPUs where the surrogate never waits.

#Technical Details

BoostMD is built from MACE equivariant message-passing blocks. Edge features combine spherical harmonics of the current displacement, of the positional change since the reference step, and of the rotated reference node features through learnable tensor products; summing over neighbors gives a permutation-invariant two-body basis, which the MACE product construction raises to higher body order. A readout MLP emits either the energy change directly or an updated node feature that the reference model's own readout converts to energy.

Training used a dipeptide molecular dynamics dataset generated with MACE-OFF23-M, initialized from the dipeptide subset of SPICE. The reference model itself carries an 0.85 meV/atom error against DFT; the best BoostMD configurations sit at or below it. The fastest variant achieves an 8.6× speedup at 0.84 meV/atom energy and 75.3 meV/Å force RMSE; including equivariant reference features trades speed for accuracy, reaching 0.63 meV and 56.9 meV/Å at 2.3×. A single-layer MACE model trained from scratch for comparable speedup was markedly worse (1.80 meV, 121.1 meV/Å), which isolates feature reuse as the source of the gain. Simulations stayed stable to 10 ns (10⁷ steps), and 5 ns metadynamics on alanine dipeptide — a molecule absent from training — reproduced the reference Ramachandran free energy surface with the reference evaluated only every tenth step.

#Applications

BoostMD targets the long-timescale simulations that biomolecular questions actually require: peptide conformational sampling, catalytic reaction pathways, and free energy estimation via enhanced sampling, where the acceleration compounds with the biasing scheme. It is not a standalone force field — it runs in tandem with a reference model and inherits that model's chemistry and accuracy ceiling. Demonstrated transfer is to unseen dipeptides rather than to whole proteins or broad chemical space, so the practical envelope is small biomolecules within the reference model's organic-chemistry domain.

#Impact

The contribution is architectural rather than chemical: it shows that the hidden representations of an atomistic foundation model are a reusable asset within a trajectory, not something to be discarded and rebuilt every step. Because BoostMD conditions on a frozen backbone, it inherits improvements to that backbone without retraining, and the same construction should apply to other equivariant force fields whose cost is dominated by feature generation. The work is a preprint awaiting peer review, and no code, weights, or license accompany it — a real barrier for a method whose value is measured in wall-clock time on someone else's hardware.