insider brief
- The researchers developed a molecular dynamics acceleration method called DMTS-NC that combines distilled neural networks, multi-time steps, and nonconservative forces to speed up simulations while maintaining accuracy.
- This approach achieved speedups of up to 5.6x compared to traditional single-timestep simulations and an additional 15% to 30% performance improvement compared to the team’s previous distilled multi-timestep framework.
- Tests on aqueous systems, proteins, and small molecules show that this method largely preserves key physical properties and produces results consistent with more computationally intensive simulations.
New machine learning approaches have the potential to significantly speed up molecular dynamics simulations, and this development could allow researchers to study biological systems, materials, and drug candidates more efficiently without sacrificing much of the accuracy that has made neural network-based models popular.
According to a study published in the Journal of Chemical Theory and Computation, researchers developed a method called nonconservative distillation multiple time-stepping (DMTS-NC) that uses neural network potentials to accelerate molecular dynamics simulations. The technology delivered simulation speedups of 15% to 30% compared to the team’s previous acceleration framework, and some tests were as much as 5.6x faster than traditional single-timestep simulations.
The study was led by researchers from Sorbonne University and Cubit Pharmaceuticals.
Molecular dynamics simulations are widely used to model how atoms and molecules move and interact over time. This technology 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 potential has 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. Still, it remains significantly more expensive than traditional empirical models.
New research addresses remaining performance gaps. The DMTS-NC framework combines two established ideas. First, we use a process known as knowledge distillation. In this process, a smaller, faster neural network learns to mimic the behavior of a larger, more accurate model. Second, a computational strategy known as multi-timestepping is employed.
In molecular dynamics, the simulation progresses in a series of time increments. In general, smaller time steps are more accurate, but require more computation. Multi-timestep methods reduce computational cost by performing expensive calculations less frequently and using cheaper approximations in between.
The researchers’ previous work had already demonstrated that distilled neural networks could speed up simulations in this way. New research extends the concept by replacing traditional conservative force calculations with so-called non-conservative forces.
In physics, conservative forces can be derived directly from the energy situation. Non-conservative forces do not necessarily meet that requirement. Predicting force directly, rather than first calculating energy and then deriving force, eliminates computationally intensive steps and improves performance.
Historically, however, concerns have arisen with non-conservative approaches because they can introduce simulation artifacts and numerical instability. To address these issues, the researchers designed a distillation model to preserve several important physical properties, such as rotational symmetry and cancellation of atomic force components.
The resulting neural network was significantly smaller than the reference model. The distillation force model contained approximately 287,000 parameters, compared to more than 9.5 million parameters in the large-scale FeNNix-Bio1 foundational model used as a reference.
According to the study, the smaller model more closely matched the force predictions of the larger model than the previous conservatively sampled model. The improved match allows 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 to calculate hydration free energies.
For water simulations, DMTS-NC achieved acceleration factors ranging from approximately 2.9 to 4.5 times compared to standard single time-step simulations. Compared to the team’s previous conservative distillation multi-time-step approach, performance improved by approximately 31% to 56%, depending on system size.
Tests involving two biologically relevant protein systems yielded similar results. Simulations of the phenol-lysozyme protein-ligand complex and solvated dihydrofolate reductase protein achieved approximately 3x speedup compared to standard simulations. This method also outperformed both general-purpose and specially-tuned versions of previous conservative frameworks.
Importantly, the researchers reported that the main physical properties remained largely unchanged. Measurements of temperature distribution, structural properties, and potential energy distribution were in close agreement with results obtained from conventional simulations.
The method also showed good performance in hydration free energy calculations, a common benchmark used in drug discovery research. Across 44 small molecules, the difference between conventional simulations and DMTS-NC averages about 0.11 kcal per mole, indicating that accuracy is largely maintained despite the increased speed.
The researchers further demonstrated that this 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 possibility. Since MACE-OFF23 is computationally more expensive than FeNNix-Bio1, the speedup benefit is even greater, reaching speeds 3.7 to 5.6 times faster than traditional simulations.
Several additional techniques were used to further improve performance. One approach, called hydrogen mass redistribution, redistributes mass within the molecule to slow the highest frequency vibrations that often limit the time step of simulations. Another method, called high hydrogen friction, reduces the numerical instability by weakening the hydrogen motion.
These techniques have allowed some simulations to use time steps as long as 10 femtoseconds while maintaining stability. The tradeoff is a reduction in diffusion rate, which means that the particles move somewhat more slowly through the simulated system. The researchers found that diffusion losses generally remain smaller than the gains in computational speed, although the effects become more pronounced when the most aggressive acceleration techniques are used.
This study also highlights ongoing challenges to molecular dynamics over multiple time steps. As time steps become larger, simulations can introduce resonance effects, numerical artifacts caused by interactions between physical molecular motion and artificial timing structures forced by the algorithm. These resonances ultimately limited the maximum stable acceleration that could be achieved.
The researchers suggested that future research will focus on improving performance for larger, more complex systems, improving sampling behavior, and exploring additional techniques to reduce resonance-related instabilities.
Although this method does not eliminate the substantial computational demands associated with high-precision molecular simulations, our results suggest that carefully designed distilled neural networks can narrow the gap between the potential of machine learning and traditional force fields. If this approach can be successfully extended to larger biological and materials systems, it could enable longer simulation times and enable more widespread use of neural network-based molecular modeling in chemistry, drug discovery, and materials research.
