Ever get tired of tiny timesteps bottlenecking your MD simulations?
We show how to train a model for large-timestep Hamiltonian dynamics directly on standard MLFF datasets. ๐ก๐ผ ๐ฟ๐ฒ๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ ๐๐ฟ๐ฎ๐ท๐ฒ๐ฐ๐๐ผ๐ฟ๐ถ๐ฒ๐, ๐ป๐ผ ๐๐ป๐ฟ๐ผ๐น๐น๐ถ๐ป๐ด, ๐ป๐ผ ๐๐ฒ๐ฎ๐ฐ๐ต๐ฒ๐ฟ needed!๐งต๐
@guforosso2@leonklein26 Unfortunately, there is no free lunch. For the system shown, our model trained about 7 days.
I think this approach is especially promising once it can be transferred. We also trained a model that can simulate all dipeptides, but this is not feasible for larger systems yet ...
@trendradar_app @leonklein26@DoctorYev We compare our results with the ground-truth simulation from the fast-folding protein paper and note close alignment between the two. What I find interesting is that the FES can look a bit oversmoothed, possibly due to the "simple" architecture we use that does not use any priors
Fantastic work by Michael Plainer and friends. Energy-based diffusion models to ensure that the denoising distribution equals exp(-u(x)), with the energy u(x). A keystone for connecting molecular dynamics, statistical mechanics and generative AI.
(1/n) Can diffusion models simulate molecular dynamics instead of generating independent samples?
In our NeurIPS2025 paper, we train energy-based diffusion models that can do both:
- Generate independent samples
- Learn the underlying potential ๐ผ
๐งต๐
https://t.co/TSurVY3YEl