Please select To the mobile version | Continue to access the desktop computer version

BBS-DeePMD

 Forgot password?
 Register
Search
View: 113|Reply: 1
Collapse the left

training strategy problem

[Copy link]
Post time: 2020-09-11 12:43:24
| Show all posts |Read mode
my strategy is that:
T=300K
P=1bar
sys_idx=4
nstep=1000->2000->4000->8000
ensemble=npt
  "model_devi_dt": 0.002,
  "model_devi_skip": 0,
  "model_devi_f_trust_lo":  0.05,
  "model_devi_f_trust_hi":  0.15,

but the result is that:
INFO:dpgen:system 000 candidate :      0 in    306   0.00 %
INFO:dpgen:system 000 failed    :    306 in    306 100.00 %
INFO:dpgen:system 000 accurate  :      0 in    306   0.00 %
INFO:dpgen:system 000 accurate_ratio:   0.0000    thresholds: 1.0000 and 1.0000   eff. task min and max   -1  500   number of fp tasks:      0

INFO:dpgen:system 001 candidate :      0 in   2424   0.00 %
INFO:dpgen:system 001 failed    :   2424 in   2424 100.00 %
INFO:dpgen:system 001 accurate  :      0 in   2424   0.00 %
INFO:dpgen:system 001 accurate_ratio:   0.0000    thresholds: 1.0000 and 1.0000   eff. task min and max   -1  500   number of fp tasks:      0

INFO:dpgen:system 002 candidate :      0 in   3015   0.00 %
INFO:dpgen:system 002 failed    :   3015 in   3015 100.00 %
INFO:dpgen:system 002 accurate  :      0 in   3015   0.00 %
INFO:dpgen:system 002 accurate_ratio:   0.0000    thresholds: 1.0000 and 1.0000   eff. task min and max   -1  500   number of fp tasks:      0

INFO:dpgen:system 003 candidate :      4 in   6015   0.07 %
INFO:dpgen:system 003 failed    :   6011 in   6015  99.93 %
INFO:dpgen:system 003 accurate  :      0 in   6015   0.00 %
INFO:dpgen:system 003 accurate_ratio:   0.0000    thresholds: 1.0000 and 1.0000   eff. task min and max   -1  500   number of fp tasks:      4

no accurate part, very few candidata part, but allmost 100% failed part.

so what's the problem of my training process?

Reply

Use magic Report

Post time: 2020-09-11 15:38:18
| Show all posts
First you should check your training accuracy, i.e. energy and force error in `00.train/000/lcurve.out`.
There are some reasons that may cause bad training accuracy, like wrong training parameters, inadequate training steps, or even bad quality of DFT data.

If you are sure that there's no problem within current training scheme, you could increase the threshold of error indicator.
Reply Support Not support

Use magic Report

You have to log in before you can reply Login | Register

Points Rules

Dark room|DeePMD

2020-10-02 02:41 GMT+8 , Processed in 0.098539 sec., 19 queries .

Quick Reply To Top Return to the list