Energy and Force RMSE goes up with iteration

[changxiaoju]

system:H
init data: generated from init_bulk, 2000 configurations
training:

Iter	DPMD	Length[ps]	Ensemble	Temperature[K]	Pressure[Gpa]	Trust range[ev/A]	Accurate[%]	Candidate[%]	Fail[%]
00	189	0.2=0.0002*1000	NPT	200	50,100,200,300,400,500,600	[2.0,2.5]	25.87	13.21	60.92
01	189	0.8=0.0002*4000	NPT	200	50,100,200,300,400,500,600	[0.25,0.35]	40.3	33.90	25.80
02	189	4=0.0002*20000	NPT	200	50,100,200,300,400,500,600	[0.1,0.125]	46.7	38.11	15.19
03	189	4=0.0002*20000	NPT	200	50,100,200,300,400,500,600	[0.1,0.125]	87.68	10.66	1.66
04	189	4=0.0002*20000	NPT	200	50,100,200,300,400,500,600	[0.1,0.125]	82.7	3.24	14.05
05	189	8=0.0002*40000	NPT	200	50,100,200,300,400,500,600	[0.1,0.125]	85.21	0.44	14.35
06	189	16=0.0002*80000	NPT	200	50,100,200,300,400,500,600	[0.1,0.125]	85.58	0.13	14.29

but force error goes up with iteration:

iter	00	01	02	03	04	05
std(F_DP - F_DFT)(ev/A)	0.0812	0.1323	0.1474	0.1631	0.1629	0.1733
std_Force(ev/A)	0.8658	0.8658	0.8658	0.8658	0.8658	0.8658
std(F_DF - F_DFT)/std_Force	0.0938	0.1528	0.1702	0.1884	0.1881	0.2001

also can be seen by detail of the error:
iter00:

iter05:

training error:
iter00:

iter05:

init data force distribution:

Energy and Force RMSE goes up. Is this an overfitting problem? It seems that configuration dpgen learned are too close to each other, without H-H bond formation.
Init data generated by dpgen init_bulk
02.md : 100steps*0.2fs=0.02ps,200K,160Gpa
POSCAR:

CONTCAR:

however, in model_devi, the configuration is like:

[jameswind]

What’s the magnitude of forces? What is the training error? Does increasing the training steps help?

…

On Wed, Sep 29, 2021 at 8:42 PM changxiaoju ***@***.***> wrote: system:H init data: generated from init_bulk, 2000 configurations training: Iter DPMD Length[ps] Ensemble Temperature[K] Pressure[Gpa] Trust range[ev/A] Accurate[%] Candidate[%] Fail[%] 00 189 0.2=0.0002*1000 NPT 200 50,100,200,300,400,500,600 [2.0,2.5] 25.87 13.21 60.92 01 189 0.8=0.0002*4000 NPT 200 50,100,200,300,400,500,600 [0.25,0.35] 40.3 33.90 25.80 02 189 4=0.0002*20000 NPT 200 50,100,200,300,400,500,600 [0.1,0.125] 46.7 38.11 15.19 03 189 4=0.0002*20000 NPT 200 50,100,200,300,400,500,600 [0.1,0.125] 87.68 10.66 1.66 04 189 4=0.0002*20000 NPT 200 50,100,200,300,400,500,600 [0.1,0.125] 82.7 3.24 14.05 05 189 8=0.0002*40000 NPT 200 50,100,200,300,400,500,600 [0.1,0.125] 85.21 0.44 14.35 06 189 16=0.0002*80000 NPT 200 50,100,200,300,400,500,600 [0.1,0.125] 85.58 0.13 14.29 but force error goes up with iteration: iter 00 01 02 03 04 05 std(F_DP - F_DFT)(ev/A) 0.0812 0.1323 0.1474 0.1631 0.1629 0.1733 std_Force(ev/A) 0.8658 0.8658 0.8658 0.8658 0.8658 0.8658 std(F_DF - F_DFT)/std_Force 0.0938 0.1528 0.1702 0.1884 0.1881 0.2001 also can be seen by detail of the error: iter00: [image: image] <https://user-images.githubusercontent.com/44628664/135198569-7339347a-b390-436d-b049-288008d74a5f.png> iter05: [image: image] <https://user-images.githubusercontent.com/44628664/135199039-651285bb-259b-4f0c-baaf-19fda64dca01.png> Energy and Force RMSE goes up. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#557>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEJ6DC42RWM4M5QQYLLCMLTUEMCUZANCNFSM5E7ZFFWA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[changxiaoju]

thank you, i added training error and force distribution above. Increasing steps and "limit_pref_f" did not solve this problem.

[changxiaoju]

Maybe something wrong with my origin data?

[Chenghao-Wu]

any updates in this issue?