Hyperparameter optimization giving abnormally high energy noise

[MorAnyAn]

Hi, FLARE developer team,

I have been using FLARE (ACE descriptor + sparse GP) lately.

The on-the-fly training was successful and I am glad that the resultant model has provided a decent force validation compared with DFT results.

However, the total potential energy given by the GP model is observing a constant shift, and the amount of such shifting is consistent across validation sets with different energy ranges:

In other attempts of training, such shift may be different as well:

It seems to be a issue in hyperparameter optimization, since the hyperparameters for the noise of energy ended up relatively high:

As for the related setting, I have set the initial noise of energy at 0.063 as the recommended 0.001*atom numbers; hyperparameters are optimized after every 10 DFT calls and maximum iterations of optimization is 100.

Do you have any suggestion on this issue?