Accuracy ratio decreases after dpgen active learning

[ZKNB]

My training process is as follows:

• Initial dataset: AIMD simulation data of Fe (100) surface (54 atoms) at 1500K, totaling 10ps, 10,000 frame structures.

• Structure for active learning: Fe (100) surface (54 atoms)

• param.json
  param.json

The test results for ‘iter-0’ are as follows:

After one active learning session, the test results for ‘iter-1’ are as follows:

Active learning to supplement data should increase the accuracy of the dataset, but in my tests, this does not seem to be the case. Could you tell me why the accuracy has decreased after active learning to supplement data. Is there any training experience to share on what strategies can be used to increase the accuracy of the dataset? I hope you can give me some advice. Looking forward to your reply.

[ur1742]

think your 300000 steps of learning not enough for 10000 frames of data