With the addition of 'virial' information, the training accuracy is rather worse

[ZKNB]

For the same system, adding stress information makes the accuracy worse instead, here are my test results:

trainingdata	validationdata	rcut_smth	rcut	embedding net	fitting net	decay_steps	start/limit_pref_e	start/limit_pref_f	start/limit_pref_v	sel	Energy RMSE eV	Energy RMSE/Natoms	Force  RMSE  eV/A	Virial RMSE/Natoms
21412	0	0.50	6.50	25 50 100	240 240 240	50000	200/200	15/15	no	auto	3.23E-01	1.05E-02	2.10E-01
21412	0	0.50	6.50	25 50 100	240 240 240	50000	200/200	15/15	1/1	auto	4.90E-01	1.49E-02	2.64E-01	3.23E-01

As I understand it, adding stress should make the accuracy higher, but it turns out to be the other way around, I want to train the accuracy to an order of magnitude of 1e-3 for energy and an order of magnitude of 1e-2 for force, how do I tweak the parameters, can you give some advice?

[wanghan-iapcm]

it is highly possible that the virial labels are not consistent with energies and forces. please check the convergence of virial with respect to the energy cutoff and k-spacing used in your dft calculations.