Multi-system training got low energy accuraccy due to large energy std

[Franklalalala]

I prepared 14 systems for training. They have at least 2 Carbons in different, which induced large energy std. Going through the deepmd paper, I found that, the energy accuracy was measured with test dataset's std.
So I checked my dataset's distribution. The std is 172. The pred error is 1~2, which is roughly 1% of the std. However, the accuracy is still too low.
My systems↓

The distribution↓

Is there any tricks to lower training data's std? I was recommended to replace the energy with 'formation energy', however, i don't think it's a good idea, for the energy prediction is the sum of atoms' energy, it should be linear with atom number.

[njzjz]

The model has already had a bias for atomic energies, which is just like the formation energy.

…

On Oct 15, 2021, at 23:08, Franklalalala ***@***.***> wrote:  I prepared 20 systems for training. They have 2 Carbons in different, which induced large energy std. Going through the deepmd paper, I found that, the energy accuracy was measured with test dataset's std. So I checked my dataset's distribution. The std is 172. The pred error is 1~2, which is roughly 1% of the std. However, the accuracy is still too low. Is there any tricks to lower training data's std? I was recommended to replace the energy with 'formation energy', however, i don't think it's a good idea. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.