DPgen： model_devi_e_trust_lo and _hi do not work

[MaterialDesign-USTS]

During my training using DPgen, the model_devi_e_trust_lo and _hi were set to 0.001 and 100, no warning or error was reported. After traing, the dptest shows the energy error was large. The accurate was 100% during the training and no first principles calculations were performed. There is no min_devi_e/max_devi_e/avg_devi_e in the model_devi.out. It seems the model_devi_e_trust_lo and _hi do not work.
dpgen.log:
2025-12-30 10:56:47,965 - INFO : =============================iter.000011==============================
2025-12-30 10:56:47,965 - INFO : -------------------------iter.000011 task 00--------------------------
2025-12-30 10:56:48,008 - INFO : -------------------------iter.000011 task 01--------------------------
2025-12-30 10:56:48,012 - INFO : -------------------------iter.000011 task 02--------------------------
2025-12-30 10:56:48,015 - INFO : -------------------------iter.000011 task 03--------------------------
2025-12-30 10:56:48,138 - INFO : -------------------------iter.000011 task 04--------------------------
2025-12-30 10:58:56,411 - INFO : -------------------------iter.000011 task 05--------------------------
2025-12-30 10:58:56,415 - INFO : -------------------------iter.000011 task 06--------------------------
2025-12-30 10:58:56,441 - INFO : system 000 candidate : 0 in 48 0.00 %
2025-12-30 10:58:56,441 - INFO : system 000 failed : 0 in 48 0.00 %
2025-12-30 10:58:56,441 - INFO : system 000 accurate : 48 in 48 100.00 %
2025-12-30 10:58:56,457 - INFO : system 000 accurate_ratio: 1.0000 thresholds: 1.0000 and 1.0000 eff. task min and max -1 0 number of fp tasks: 0
2025-12-30 10:58:56,460 - INFO : -------------------------iter.000011 task 07--------------------------
2025-12-30 10:58:56,466 - INFO : -------------------------iter.000011 task 08--------------------------
2025-12-30 10:58:56,748 - INFO : =============================iter.000012==============================
2025-12-30 10:58:56,748 - INFO : -------------------------iter.000012 task 00--------------------------
2025-12-30 10:58:56,786 - INFO : -------------------------iter.000012 task 01--------------------------
2025-12-30 10:58:56,790 - INFO : -------------------------iter.000012 task 02--------------------------
2025-12-30 10:58:56,793 - INFO : -------------------------iter.000012 task 03--------------------------
2025-12-30 10:58:56,918 - INFO : -------------------------iter.000012 task 04--------------------------
2025-12-30 11:01:05,262 - INFO : -------------------------iter.000012 task 05--------------------------
2025-12-30 11:01:05,265 - INFO : -------------------------iter.000012 task 06--------------------------
2025-12-30 11:01:05,292 - INFO : system 000 candidate : 0 in 48 0.00 %
2025-12-30 11:01:05,292 - INFO : system 000 failed : 0 in 48 0.00 %
2025-12-30 11:01:05,292 - INFO : system 000 accurate : 48 in 48 100.00 %
2025-12-30 11:01:05,310 - INFO : system 000 accurate_ratio: 1.0000 thresholds: 1.0000 and 1.0000 eff. task min and max -1 0 number of fp tasks: 0
2025-12-30 11:01:05,314 - INFO : -------------------------iter.000012 task 07--------------------------
2025-12-30 11:01:05,320 - INFO : -------------------------iter.000012 task 08--------------------------
2025-12-30 11:01:05,600 - INFO : =============================iter.000013==============================
2025-12-30 11:01:05,600 - INFO : -------------------------iter.000013 task 00--------------------------
2025-12-30 11:01:05,645 - INFO : -------------------------iter.000013 task 01--------------------------
2025-12-30 11:01:05,650 - INFO : -------------------------iter.000013 task 02--------------------------
2025-12-30 11:01:05,653 - INFO : -------------------------iter.000013 task 03--------------------------
2025-12-30 11:01:05,781 - INFO : -------------------------iter.000013 task 04--------------------------
2025-12-30 11:03:14,303 - INFO : -------------------------iter.000013 task 05--------------------------
2025-12-30 11:03:14,310 - INFO : -------------------------iter.000013 task 06--------------------------
2025-12-30 11:03:14,335 - INFO : system 000 candidate : 0 in 48 0.00 %
2025-12-30 11:03:14,335 - INFO : system 000 failed : 0 in 48 0.00 %
2025-12-30 11:03:14,335 - INFO : system 000 accurate : 48 in 48 100.00 %
2025-12-30 11:03:14,350 - INFO : system 000 accurate_ratio: 1.0000 thresholds: 1.0000 and 1.0000 eff. task min and max -1 0 number of fp tasks: 0

[dosubot[bot]]

Hi @MaterialDesign-USTS! I'm Dosu and I’m helping the dpgen team.

The parameters model_devi_e_trust_lo and model_devi_e_trust_hi are deprecated in current DPGEN versions and have no effect on triggering first principles (fp) calculations or on the contents of model_devi.out. That’s why you see 100% accuracy, no fp tasks, and no min_devi_e/max_devi_e/avg_devi_e in your outputs, regardless of the values you set for these parameters. Instead, you should use model_devi_v_trust_lo and model_devi_v_trust_hi (for value/virial-based thresholds) or model_devi_f_trust_lo and model_devi_f_trust_hi (for force-based thresholds) to control model deviation selection and trigger fp calculations. See this issue and maintainer comment.

Example snippet:

"model_devi_f_trust_lo": 0.05,
"model_devi_f_trust_hi": 0.15,
"model_devi_v_trust_lo": 1e10,
"model_devi_v_trust_hi": 1e10

Force-based thresholds are generally recommended for most workflows. If fp calculations are still not triggered, double-check your parameter names, DPGEN version, and job submission method (some users report issues when running training via SLURM instead of locally) source.

To reply, just mention @dosu.

---

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

  

[MaterialDesign-USTS]

Thanks for your reply. The forces predicted by the trained deep potential (DP) show excellent consistency with DFT-calculated results. Could you recommend practical methods to enhance the DP’s accuracy specifically in energy predictions?

[dosubot[bot]]

To improve energy accuracy in your DP model (when forces are already accurate), try these practical steps:

• Increase the number of training steps—production runs often use 4M–16M steps, which can help reduce constant energy shifts (source).
• Retrain from your current model checkpoint with a smaller starting learning rate (e.g., 0.0001) and set the energy loss prefactor higher (e.g., set both start_pref_e and limit_pref_e to 1). Example config:

  "learning_rate": {
      "type": "exp",
      "start_lr": 0.0001,
      "stop_lr": 3.51e-08
  },
  "loss": {
      "start_pref_e": 1,
      "limit_pref_e": 1,
      "start_pref_f": 1,
      "limit_pref_f": 1,
      "start_pref_v": 1,
      "limit_pref_v": 1
  }
  

  Then run: dp train --init-model model.ckpt second_round_input.json (source).
• Continue training from the last iteration instead of starting from scratch each time (source).
• Remove extremely high-energy structures from your training set—these can worsen energy accuracy even if force errors look small (source).
• If your dataset covers a wide energy range, consider requesting or experimenting with a logarithmic loss function (not standard yet) to reduce the impact of outliers (source).
• Use a pair-wise repulsive (tabulated) potential to avoid sampling unphysical, high-energy structures (source).
• Make sure you compare model deviations using the same trajectory/data for fair evaluation (source).
• Recent versions of deepmd-kit include normalization of energy configuration embeddings, which can further improve energy accuracy (source).

If you see energy or force RMSE increasing with training, check for overfitting or insufficient data diversity, and try increasing training steps or improving your dataset (source).

To reply, just mention @dosu.

---

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