Skip to content

Navigation Menu

Sign in
Sign up

ZBL will increase computational cost significantly #5348

Unanswered
MoseyQAQ asked this question in Q&A
Discussion options

Dear community members,

I used DeePMD-kit v3.1.2 to train two models: w/ ZBL and w/o ZBL.
I find that using ZBL will significantly slow the training and MD process (~10X slower). Is it the expected behavior?

  1. Related training JSON script:
 "use_srtab": "zbl.txt",
 "smin_alpha": 0.08,
 "sw_rmin": 0.5,
 "sw_rmax": 0.7
  1. Training:
    using ZBL:
[2026年03月25日 20:59:07,153] DEEPMD INFO batch 918000: total wall time = 43.29 s
[2026年03月25日 20:59:50,462] DEEPMD INFO batch 919000: trn: rmse = 1.81e-01, rmse_e = 4.80e-04, rmse_f = 3.95e-02, rmse_v = 7.10e-04, lr = 8.39e-08
[2026年03月25日 20:59:50,462] DEEPMD INFO batch 919000: total wall time = 43.31 s
[2026年03月25日 21:00:34,124] DEEPMD INFO batch 920000: trn: rmse = 1.42e-01, rmse_e = 6.52e-05, rmse_f = 3.09e-02, rmse_v = 1.74e-03, lr = 7.97e-08
[2026年03月25日 21:00:34,124] DEEPMD INFO batch 920000: total wall time = 43.66 s
[2026年03月25日 21:01:17,953] DEEPMD INFO batch 921000: trn: rmse = 1.85e-01, rmse_e = 1.94e-04, rmse_f = 4.01e-02, rmse_v = 2.34e-03, lr = 7.97e-08
[2026年03月25日 21:01:17,953] DEEPMD INFO batch 921000: total wall time = 43.83 s
[2026年03月25日 21:02:01,387] DEEPMD INFO batch 922000: trn: rmse = 1.04e-01, rmse_e = 1.62e-04, rmse_f = 2.26e-02, rmse_v = 1.01e-03, lr = 7.97e-08
[2026年03月25日 21:02:01,387] DEEPMD INFO batch 922000: total wall time = 43.43 s

w/o ZBL:

[2026年03月17日 04:30:05,399] DEEPMD INFO batch 1978000: total wall time = 11.19 s
[2026年03月17日 04:30:16,533] DEEPMD INFO batch 1979000: trn: rmse = 1.36e-01, rmse_e = 1.01e-04, rmse_f = 2.87e-02, rmse_v = 2.37e-03, lr = 4.09e-08
[2026年03月17日 04:30:16,533] DEEPMD INFO batch 1979000: total wall time = 11.13 s
[2026年03月17日 04:30:27,666] DEEPMD INFO batch 1980000: trn: rmse = 1.84e-01, rmse_e = 1.05e-04, rmse_f = 4.02e-02, rmse_v = 2.17e-03, lr = 3.89e-08
[2026年03月17日 04:30:27,666] DEEPMD INFO batch 1980000: total wall time = 11.13 s
[2026年03月17日 04:30:38,815] DEEPMD INFO batch 1981000: trn: rmse = 1.08e-01, rmse_e = 2.78e-04, rmse_f = 2.33e-02, rmse_v = 8.48e-04, lr = 3.89e-08
[2026年03月17日 04:30:38,815] DEEPMD INFO batch 1981000: total wall time = 11.15 s
  1. MD in LAMMPS:
    using ZBL:
Performance: 11.853 ns/day, 2.025 hours/ns, 137.187 timesteps/s, 2.489 Matom-step/s

w/o ZBL:

Performance: 1.367 ns/day, 17.552 hours/ns, 15.826 timesteps/s, 287.147 katom-step/s
You must be logged in to vote

Replies: 1 comment 1 reply

Comment options

Could you please provide a smallest training zip file that contains the whole data that required to do your test? Since in my situation, zbl do not lead to any more expense(or very few).

You must be logged in to vote
1 reply
Comment options

OutisLi Mar 30, 2026
Collaborator

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Category
Q&A
Labels
None yet
2 participants

AltStyle によって変換されたページ (->オリジナル) /