GPUMD-v4.8

[brucefan1983]

GPUMD-v4.8

• Updates for the nep executable:

  • Changed to also fine-tune the descriptor parameters for the fine_tune keyword, which has been tested to lead to higher accuracy than fine-tuning the neural network parameters only. fine-tune descriptor by default #1295
  • For qNEP models, added lambda_q and lambda_z to set the loss weights for the total charge (presently zero) and Born effective charge (BEC). lambda_q and lambda_ z #1293

• New gnep executable:

  • This has been in existence for quite a few months, but only became mature recently. It provides an alternative way to train certain versions of NEP potential models.

• Updates for the gpumd executable:

  • The fix keyword is in effective for the minimize sd keyword, which can be used to fix atoms during minimization. fix in minimize #1319

---

This discussion was created from the release GPUMD-v4.8.