Questions on the influence of the force error on the motion of atoms

[Ding-Hairui]

Dear developers and users,
I am interested in a recent work using DP to simulate superionic ices (PNAS. 119 e2203397119, 2022). After I run a NVT with their DP model and ice structure, I found the MSD (mean squared displacement) of H atoms can not satisfy the AIMD result (I run a 3000 K NVT, corresponding to the green line in the first picture).


Then I noticed the force RMSE of the trained model is about 0.4 eV/A. I am not sure if it is accurate, but the difference between MSDs of DP and AIMD shows that the description of this model to H atoms seems to be unreliable. So I have some questions:
1. The MSD of AIMD is nearly a straight line. But in the situation of DP, there is a curve at the front. What causes the change of the slope?
2. Although the MSDs of H atoms are distinguishing, the MSDs of O atoms are consistent. I have checked the force RMSE on H and O atoms (about 0.35 eV/A and 0.58 eV/A, respectively), showing that the model has worse description to O atoms. Is it because H atoms are lighter than O atoms?
3. Based on question 2, if I want to accurately simulate the motion of H atoms, do I need to train a model with very small force RMSE, such as less than 1e-2 eV/A? According to my experience, it's not easy to train force well, especially in situations of high temperatures. Or are there any other methods to achieve this goal?
Thank you for sharing your opinions!

[Ding-Hairui]

I found a mistake in my MSD calculation code. Now the corrected MSD of DPMD looks much better. Sorry for disturbing.