DoMiNO: Down-scaling Molecular Dynamics with Neural Graph Ordinary Differential Equations

Molecular dynamics (MD) simulations are crucial for understanding and predicting the behavior of molecular systems in biology and chemistry. However, their wide adoption is hindered by two main challenges: (1) computational cost, because fine-grained simulations often require millions of small timesteps, and (2) lack of flexibility, as existing machine-learning-based surrogates typically operate at either a single small or single large timestep. These approaches either accumulate significant rollout errors or lose the ability to produce finegrained results if the timestep is large. To address these issues, we propose DoMiNO: Down-scaling Molecular Dynamics with Neural Graph Ordinary Differential Equations, a hierarchical framework that models multi-scale dynamics. Specifically, DoMiNO performs down-scaling by progressively up-sampling the trajectory across multiple temporal resolutions, equipping each level with a Neural Graph ODE to capture that scale's dominant behavior. At inference, DoMiNO flexibly combines different timestep sizes to predict both short- and longrange dynamics with high fidelity. Empirical results on challenging MD bench-marks-ranging from small molecules to proteins-demonstrate the method's longterm stability, flexibility, and accuracy. Our implementation is available at https://anonymous.4open.science/r/domino-code-6B6A.