Questions regarding the low boundary of model deviation

[jaejae9804]

Thanks in advance for checking out my discussion to everyone.

Ive seen in the papers and the some documents of roughly setting the low boundary of model deviation.
What Ive read is "In principle, low boundary should be higher than the lowest error that DP can achieve" and also
"the low boundary should not be set arbitrarily small, but a value slightly higher than the training accuracy achievable by DP."
What I want to ask is, I dont really understand what "training accuracy" or "lowest error" indicates. What I can think of is as follows

1. Say that Ive went through 3 iterations and sampled some DFT data of system1. Then I test all the collected datasets and test them to an ensemble by using dp test mode to calculate the force_RMSE. By referring to the corresponding force_RMSE, I can modify the low boundary of model deviation to a value that is slightly higher than the force_RMSE.

2. Use the force_RMSE of lcurve.out at the third iteration and modify the low boundary.

Any suggestions or ideas would be very appreciated