Issues with glassy solid dataset preparation and NPT ensemble in DP-GEN

[oakree1688]

Hello everyone, we constructed an initial training set for Li-P-S and Li-P-S-I glassy systems in the NPT ensemble. The interatomic potential fitted from the datasets generated by DP-GEN under NPT conditions performs well for simulating different systems.
Due to additional requirements, we added more NPT datasets with varying Li-P-S-I compositions. After iteration, when using the newly fitted potential from the updated DP-GEN dataset to run NPT simulations on the above glassy systems, we observed that the densities of the structural trajectories are generally higher than expected. Even the predicted densities of the original normal Li-P-S and Li-P-S-I datasets increased after incorporating the new data.
The densities in the newly added datasets are 0.1–0.2 g/cm³ higher than experimental values, which is still within a controllable range. However, this relatively small amount of new data appears to have a disproportionately large influence on the other datasets.
Furthermore, the new model, when performing NVT simulations at normal densities, shows significant aggregation/clustering and the formation of very large voids. In contrast, the original model produces normal structural trajectories under the same NVT conditions at normal densities.
May I ask what could be causing this issue? Why does a small quantity of higher-density data exert such a strong impact? Notably, data from the Li-P-S-I system even affects the simulation results of the Li-P-S system.
For reference, due to Linux version limitations, we are using the DPA-1 model with the se_atten_v2 descriptor.
I provide the void structures observed in the NVT simulations (new model) and the normal simulation structures (original model) for reference.