Loss Convergence

[alib-l]

Hello,
I am training a model using the "se_a" type. I am using "ReLu" as an activation function, and the "batch size" is 1. The problem is that the loss converges slowly and there is a significant difference between test and train loss (after 170,000 batches, the training loss is: 12, and the test loss is:13). Also, the force losses are increasing. What is your advice for better convergence?
I attached the "lcurve.out" plot for total and force losses. I would be thankful if you help me with this issue.

[njzjz]

I don't recommend to use ReLU, since ReLU function is not derivable at x=0, and its derivative is not continuous at x<0 and x>0. However, the force is the negative energy gradient.

[njzjz]

I am assuming your energy are inconsistent with forces. You may check it by only using energy or forces and see the loss again.

[njzjz]

I believe your forces are not correct. Can you provide more information, for example, how you take the data?

[alib-l]

I provided all of the data from the OUTCAR using my own python script, and using the below tags.
"TOTAL-FORCE" for "force.raw"
"POSITIONS" for "coord.raw"
"in KB" for "virial.raw"
"free energy TOTEN" for "energy.raw"

Also, I checked my raw files with "dpdata".

[WU2JING]

Hello, i also meet the problem.Have you solved the problem？

[alib-l]

@WJforu Not yet, HBU?