Exploring Error Expression Disparities Between Direct Training with DP and Finetune: Impact of Powerless Data

[JiangXiaoMingSan]

Discovering the difference in error expression between training potential functions directly with finetune and dp. For example, an isolated atom with only energy and zero force will exhibit a very small value (1e-17) in the error of finetune, while training the potential function directly on the same dataset dp will not result in this situation. Will these powerless data affect the results of finetune？
The lcurve.out files for both are located on GitHub：https://github.com/JiangXiaoMingSan/lcurve.out.git
The image inside is an example of a small training error in force, which is caused by a training set with zero force

[njzjz]

1e-17 is less than the accuracy of FP64, so it might be the floating error.