flare
flare copied to clipboard
mir-group
mean absolute error too big. How to improve it?
The error is too big after training GP model from aimd. How to improve? Thanks
mean absolute error: 241.92 meV/A mean absolute dft component: 883.22 meV/A mae per species type O mae: 268.13 meV/A type Li mae: 116.08 meV/A type P mae: 435.55 meV/A type N mae: 326.57 meV/A
Hi kitpeng11,
It's possible that your system has many-body interactions that a 2+3-body potential can't capture. I recommend computing the learning curve (MAE on an independent test set as a function of training set size) to confirm that the error won't drop with additional training data.
Best, Jon
Just to check, could you also report the cutoff and likelihood gradient you're seeing?
The reporting result is not calculated by 2+3-body model ( which takes really long). The reporting result is from 2 body. The cutoff is 5 and the likelihood gradient various. Do you mean the last one? Thanks a lot for your reply.
Isn’t a 2-body kernel supposed to give good performance only for some special cases?
Hi kitpeng11,
As mkrompiec mentioned, pure 2-body potentials will only give acceptable performance in very simple settings, e.g. simple crystals at low temperature; most materials require at least a 3-body contribution.
Have you tried the 2+3-body kernel yet? What kind of MAE do you see for that?
Best, Jon
