Insider Brief
- Researchers developed a molecular dynamics acceleration method called DMTS-NC that combines distilled neural networks, multi-time-stepping and nonconservative forces to speed simulations while maintaining accuracy.
- The approach achieved speedups of up to 5.6 times compared with conventional single-time-step simulations and delivered 15% to 30% additional performance gains over the team’s previous distilled multi-time-step framework.
- Tests on water systems, proteins and small molecules showed that the method largely preserved key physical properties and produced results consistent with more computationally intensive simulations.
A new machine-learning approach could significantly speed up molecular dynamics simulations, a development that may help researchers study biological systems, materials and drug candidates more efficiently without sacrificing much of the accuracy that has made neural-network-based models increasingly popular.
According to a study published in the Journal of Chemical Theory and Computation, researchers developed a method called Distilled Multiple Time-Stepping with Nonconservative Forces, or DMTS-NC, that accelerates molecular dynamics simulations powered by neural network potentials. The technique delivered simulation speedups of 15% to 30% over the team’s previous acceleration framework and as much as 5.6 times faster than conventional single-time-step simulations in some tests.
The work was led by researchers from Sorbonne Université and Qubit Pharmaceuticals.
Molecular dynamics simulations are widely used to model how atoms and molecules move and interact over time. The technique plays a central role in fields ranging from drug discovery and protein research to chemistry and materials science. However, the computational cost of accurately calculating molecular interactions often limits the size and duration of simulations.
In recent years, neural network potentials have emerged as an alternative to traditional force fields. These machine-learning models are trained on large databases of quantum-mechanical calculations and can often reproduce quantum-level accuracy at a fraction of the computational cost. Even so, they remain substantially more expensive than conventional empirical models.
The new study addresses that remaining performance gap. The DMTS-NC framework combines two established ideas. First, it uses a process known as knowledge distillation, in which a smaller and faster neural network learns to imitate the behavior of a larger and more accurate model. Second, it employs a computational strategy known as multi-time-stepping.
In molecular dynamics, simulations advance through a sequence of time increments. Smaller time steps generally produce greater accuracy but require more calculations. Multi-time-stepping methods reduce computational costs by allowing expensive calculations to be performed less frequently while cheaper approximations are used in between.
The researchers’ earlier work had already demonstrated that a distilled neural network could accelerate simulations in this manner. The new study extends the concept by replacing conventional conservative force calculations with so-called nonconservative forces.
In physics, conservative forces can be derived directly from an energy landscape. Nonconservative forces do not necessarily satisfy that requirement. Directly predicting forces instead of first calculating energies and then deriving forces eliminates computationally intensive steps and can improve performance.
Historically, however, nonconservative approaches have raised concerns because they can introduce simulation artifacts or numerical instabilities. To address those issues, the researchers designed the distilled model to preserve several important physical properties, including rotational symmetry and the cancellation of atomic force components.
The resulting neural network was dramatically smaller than the reference model. The distilled force model contained roughly 287,000 parameters compared with more than 9.5 million parameters in the larger FeNNix-Bio1 foundation model used as the reference.
According to the study, the smaller model also matched the larger model’s force predictions more closely than earlier conservative distilled models. The improved agreement allowed the researchers to use larger simulation time steps while maintaining stability.
The team evaluated the method on a variety of systems, including boxes of liquid water, solvated proteins and collections of small molecules used in hydration free-energy calculations.
For water simulations, DMTS-NC achieved acceleration factors ranging from roughly 2.9 to 4.5 times compared with standard single-time-step simulations. Relative to the team’s previous conservative distilled multi-time-step approach, performance improved by approximately 31% to 56%, depending on system size.
Tests involving two biologically relevant protein systems produced similar results. Simulations of a phenol-lysozyme protein-ligand complex and a solvated dihydrofolate reductase protein achieved speedups of roughly threefold relative to standard simulations. The method also outperformed both generic and specially fine-tuned versions of the earlier conservative framework.
Importantly, the researchers reported that key physical properties remained largely unchanged. Measurements of temperature distributions, structural properties and potential-energy distributions closely matched results obtained from conventional simulations.
The method also performed well on hydration free-energy calculations, a common benchmark used in drug-discovery research. Across 44 small molecules, differences between conventional simulations and DMTS-NC averaged approximately 0.11 kilocalories per mole, indicating that accuracy was largely preserved despite the increased speed.
The researchers further demonstrated that the approach is not limited to a single neural network architecture. In a proof-of-concept test, they applied the method to MACE-OFF23, another widely used neural network potential. Because MACE-OFF23 is computationally more expensive than FeNNix-Bio1, the acceleration benefits were even greater, reaching between 3.7 and 5.6 times faster than conventional simulations.
Several additional techniques were used to further improve performance. One approach, called hydrogen mass repartitioning, redistributes mass within molecules to slow the highest-frequency vibrations that often constrain simulation time steps. Another method, called high hydrogen friction, dampens hydrogen motion to reduce numerical instabilities.
These techniques allowed some simulations to use time steps as large as 10 femtoseconds while maintaining stability. The tradeoff was a reduction in diffusion rates, meaning particles moved somewhat more slowly through simulated systems. The researchers found that the loss in diffusion generally remained smaller than the gains in computational speed, though the effect became more pronounced when the most aggressive acceleration techniques were used.
The study also highlights ongoing challenges for multi-time-step molecular dynamics. As time steps become larger, simulations can experience resonance effects, numerical artifacts caused by interactions between physical molecular motions and the artificial timing structure imposed by the algorithm. These resonances ultimately limited the maximum stable acceleration that could be achieved.
The researchers indicated that future work will focus on improving performance for larger and more complex systems, refining sampling behavior and exploring additional techniques for mitigating resonance-related instabilities.
While the method does not eliminate the substantial computational demands of high-accuracy molecular simulations, the results suggest that carefully designed distilled neural networks can narrow the gap between machine-learning potentials and traditional force fields. If the approach scales successfully to larger biological and materials systems, it could enable longer simulations and broader use of neural-network-based molecular modeling in chemistry, drug discovery and materials research.
