The model's RMS validation error of force, RMS training error of force did not decrease

[714036303]

I converted the data of water molecules calculated by cp2k (natom=768, nconfig=50000) into raw format with python, and then used the raw_to_set.sh command to make a data set. After the training, RMS validation error of energy, and RMS training error of energy decreased satisfactorily, but RMS validation error of force and RMS training error of force did not decrease, but fluctuated around the initial value.

When I train the model of water in the deepmd-kit example, everything works fine.
This is the codes when I convert the python file in raw format and training
import numpy as np
from numpy.linalg import inv

f1=open("forces.xyz", "r")
f2=open("force.raw", "w")

natom=768
nconfig=50000
convf=27.2114/0.529177

for k in range(0,nconfig):

O=[]
H=[]

for i in range(0,2):
f1.readline()

for i in range(0,natom):
c=f1.readline().split()
if c[0]=="O":
O.append(c)
else:
H.append(c)

for i in range(0,len(O)):
for j in range(1,4):
O[i][j]=float(O[i][j])
O[i][j]=O[i][j]*convf

for i in range(0,len(H)):
for j in range(1,4):
H[i][j]=float(H[i][j])
H[i][j]=H[i][j]*convf

for i in range(0,len(O)):
f2.write(str(O[i][1])+" "+ str(O[i][2])+" "+str(O[i][3])+ " ")
for i in range(0,len(H)):
f2.write(str(H[i][1])+" "+ str(H[i][2])+" "+str(H[i][3])+ " ")
f2.write("\n")

[njzjz]

convf should be 0.529177/27.2114 when you convert Hartree/Bohr to eV/A.

[714036303]

But the strange thing is that when I use the inverse to convert the units, the model's RMS validation error of force and RMS training error of force drop successfully.

[njzjz]

What is the unit you set in the cp2k configuration?

[714036303]

The cp2k configuration is normal, and now I am trying to increase the rcut_ smth to 5.5, then shrink the neural network to [5,10,20], and now the model can fit normally.

[714036303]

Unfortunately, it seems that modifying the JSON file cannot solve the problem. The model is still the same as before

[njzjz]

Can you show a plot where the x-axis is the reference forces, and the y-axis is the model forces?

[njzjz]

In your script, you reorder the atoms, making O atoms ahead of H atoms. Did you do the same thing for coordinates and atomic types?

[714036303]

yes, I did.