The trained DP (Deep Potential) force field lacks the ability to generate sufficient repulsive forces when atoms approach each other closely, which consequently leads to situations where atoms get excessively near during the MD (Molecular Dynamics) process.

[Zhou-tao-USTC]

When I attempted to use dpgen for exploration, I found that during the MD process conducted by the model trained in the first round of iterative training, the trajectory showed that the distance between an O atom and an Al atom gradually approached an unreasonably small value. This ultimately led to the failure of the dpgen exploration process (the distance between that O and Al atoms was excessively close, and the forces inferred by the four models varied significantly).
Upon further reflection, since my training dataset itself did not cover such scenarios with extremely close atomic distances, it seems that the DP model could not learn the significant repulsive forces that would exist in real situations when atoms are very close to each other from the training set. So, what settings does DP rely on to enable the model to learn the strong repulsive forces when atoms are relatively close and thereby prevent atoms from getting too close during the MD process?
(PS: It's also possible that my reasoning is not the cause of the problem.)
deepmd.zip
coord.xyz.0：

coord.xyz.11400：

coord.xyz.13600：

[wanghan-iapcm]

In the initial stages of the DP-GEN iterations, the model exhibits suboptimal performance, which often results in the failure of the molecular dynamics (MD) trajectory.

To mitigate this issue, the primary strategy is to shorten the MD trajectory length, thereby increasing the likelihood of successful explorations. Concurrently, the frequency of frame recording from the trajectory is increased to capture more detailed data.

When the model lacks sufficient understanding of the interactions between adjacent atoms, a specific frame will emerge where the model deviation begins to rise. This frame, identified by DP-GEN as exceeding the lower trust threshold, will be designated as training data to enhance the model's capability in accurately handling interactions between nearby atoms.

[Zhou-tao-USTC]

Can the training settings be used to prevent such situations where these atoms get too close to each other?
The scenario where these atoms are almost overlapping makes it impossible to proceed from the very beginning.