High training MSE

[psn417]

I use 170000 AIMD frames to train a model, but error (both training and testing) is high and the melting point is different from AIMD. I try to modify the number of neurons , but result is similiar. How can I make error lower?

My Input:
{
"_comment": " model parameters",
"use_smooth": false,
"sel_a": [40],
"sel_r": [60],
"rcut": 5,
"n_neuron": [240, 240, 240],
"axis_rule": [],

"_comment": " traing controls",
"systems":		["../deepmd_data/"],
"set_prefix":	"set",    
"stop_batch":	1000000,
"batch_size":	1,
"start_lr":		0.002,
"decay_steps":	5000,
"decay_rate":	0.95,

"start_pref_e":	0.02,
"limit_pref_e":	1,
"start_pref_f":	1000,
"limit_pref_f":	1,
"start_pref_v":	0,
"limit_pref_v":	0,


"seed":		1

,

"_comment": " display and restart",
"_comment": " frequencies counted in batch",
"disp_file":	"lcurve.out",
"disp_freq":	1000,
"numb_test":	50,
"save_freq":	1000,
"save_ckpt":	"model.ckpt",
"load_ckpt":	"model.ckpt",
"disp_training":	true,
"time_training":	true,
"profiling":	false,
"profiling_file":	"timeline.json",

"_comment":		"that's all"

}

[psn417]

When I train other phases (bcc, fcc), the error is much lower and melting point is good.

[jameswind]

Could you check the AIMD setup? If the model works for bcc and fcc phases I guess maybe the numerical error of DFT in this new case have some problems. You may pick up some snapshots in the training data, do DFT calculations with a more strict criterion (ecut, kpoint, etc.) and see the results are consistent with you training data or not. Best wishes, Linfeng

…

On Tue, Feb 25, 2020 at 10:08 AM psn417 ***@***.***> wrote: When I train other phases (bcc, fcc), the error is much lower and melting point is good. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/deepmodeling/deepmd-kit/issues/195?email_source=notifications&email_token=AEJ6DC5PLUSO6KFBF4LC2NTREUX5LA5CNFSM4K3LU4G2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEM4J7OA#issuecomment-590913464>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEJ6DCZGRWKPPGAFF2WHRZDREUX5LANCNFSM4K3LU4GQ> .