During high-entropy alloy training, there are some discrepancies between DFT data and DP data.

[wangtongde]

I am conducting DP potential function training for a five-element high-entropy alloy. First, based on bulk and surface perturbations of the initial structure, I generated datasets for 10 different compositions of the high-entropy alloy. Then I further expanded these datasets using DP-GEN. After obtaining the datasets, I used 80% of them as the training set to train the DP potential function with a long training using DeepMD. The remaining 20% was used as a validation set to test the differences between DFT and the DP potential function. The results are as follows.

Obviously, the mechanical properties of some of the DFT data are underestimated by DP. I have tried many methods, including increasing the k-point mesh and changing the DeepMD version, but none of them worked.|The training files I use are as follows.
param.json

[dosubot[bot]]

Hi @wangtongde! I'm Dosu and I’m helping the deepmd-kit team.

The most common reason for DP underestimating mechanical properties, even when force predictions look good, is insufficient diversity in the training data—especially for challenging atomic environments and mechanical deformations. For high-entropy alloys, it's crucial that your dataset covers a wide range of configurations, including strained structures, defects, and surfaces. If the energy variance in your training data is low, or if you focus too much on force fitting (e.g., using a much larger force loss prefactor than energy), the model may not capture energy trends well, which impacts mechanical properties like elastic constants and moduli [source].

To improve accuracy:

• Use DP-GEN's active learning to add high-deviation or failed structures (e.g., those with large model deviation during MD) to your training set [source].
• Make sure your loss prefactors in param.json balance energy and force fitting; typical values are start_pref_f=1000, limit_pref_f=1 for force, but don't let energy prefactors be too small compared to force [source].
• Increase the number of training steps (4M–16M is common for production runs) and consider retraining from your current checkpoint with a smaller starting learning rate and higher energy loss prefactor [source].
• Relabel some AIMD frames with high-accuracy DFT if your initial dataset mixes low and high k-point meshes [source].
• Check that your force and energy labels are processed correctly, and monitor model deviation during training and MD simulations to identify gaps in your dataset [source].

If you can share which mechanical properties are underestimated (e.g., elastic constants, moduli), I can help with more targeted advice.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}