How to select an approperiate validation set (for train) and test set (for test)?

[Terrla]

Dear DP developers,
I am confused about the data selection between the validation set in the training and the input data in the test process. What is the difference between validation and test data selection?
For example, I have DFT data of the same system at 700 K, 800 K, and 1000 K, and I have set the data of 700 K and 1000 K as input trained data (extract part of the data as a validation set). After freezing the model, can I set the data of 800 K as input for training? Or, should I set the same data (700 K and 1000 K) as input of the test? @njzjz

Kind Regards,
Terrla

[njzjz]

If you are using the same or similar data for testing, you can see the training effectivity, and it's expected to see an accurate result. If you are using a different data, you can see the transferability and robustness of the model and it doesn't work in all time.

[Terrla]

Thanks for ur time. I am wondering what fluctuation range of force fluctuation is considered accurate?

[njzjz]

There isn't an absolute criterion as it varies with different systems.

[LiangMD-BGI]

I have a quick question: Does using a validation set help converge the training (i.e., achieving smaller RMSE values for energy, forces)?

[njzjz]

The validation set has no effect on training.