How to improve the speed of ASE-MD simulation based on trained PhysNet model?

[liudfgoo]

Hello, everyone! I have trained one PhysNet model on the solvated_protein_fragments.npz data. Now I want to conduct the MD simulations based on the model through ASE environment. The simulation system I used has 1681 atoms and satisfies PBC conditions. I set the lr_cut 8 Angstrom in the NNCalculator.py.

I found the ASE-MD simulation was running slowly. I submit the task on computation cluster and used one gpu and five cpu cores. However, the simulation of 5000 steps consumes about 6 hours. In the running process, the volatile GPU-Util almost keeps zero, but the memory usage is large.

I think the low speed arises from the ase.neighbor_list process, but I don't know how improve the situation effectively.

Looking forward to your help!

[LIVazquezS]

Hi,

There is no simple solution to this issue. ASE is not optimized to run extensive MD simulations, and as you mentioned, the ase.neighbor_list can be a problem. We are working on solutions bypassing that step to torch or CHARMM.

I am sorry we can't help you more.