[BUG] 训练出来的模型，lammps第一次结合正常，第二次结合有问题

[wangyi01]

Bug summary

DPGEN结合DPA2和fp标记样本。预训练模型+alloy_domains 获得初训练的模型后，并结合lammps来获得轨迹。然后dpgen2选择出进行fp计算的构型，fp计算获得标记的样本，然后重新训练出新模型，到此为止都是正常的，没有异常，但是在用这个新模型去再次和lammps结合时候就出问题了，这很可能是这一轮循环时候，新生成的模型和lammps的某个接口出了问题，出现的现象就是nvt 的时候，温度异常，导致原子丢失。第一轮和第二轮的区别：温度由1273K--》873K， 然后修改了train.json里面的numb_steps, disp_freq save_freq.

DP-GEN Version

dpgen2

Platform, Python Version, Remote Platform, etc

No response

Input Files, Running Commands, Error Log, etc.

https://workflows.deepmodeling.com/workflows/argo/sampling-titaalcrfenico-hk4e5

Steps to Reproduce

无

Further Information, Files, and Links

No response

[njzjz]

We don't accept non-English issues. Please use English for discussion.

[wangyi01]

We don't accept non-English issues. Please use English for discussion.

The process combines DPGEN with DPA2 and the 'fp' flag for sampling. After obtaining the pre-trained model + alloy_domains and obtaining the initially trained model, lammps is used to generate trajectories. Then, DPGen2 selects configurations for FP calculation, FP calculation obtains labeled samples, and then a new model is retrained. Up to this point, everything is normal without any exceptions. However, when using this new model to combine with lammps again, there is a problem. This is likely to occur in this round of the loop when the newly generated model encounters some interface issue with lammps, resulting in abnormal temperature during NVT simulation, leading to atom loss. The difference between the first and second rounds is that the temperature changed from 1273K to 873K, and then modifications were made to numb_steps, disp_freq, and save_freq in train.json.