Optimization result

[ZhunyunTang]

Hello,
I encountered the following problem when fitting the ace potential using pacemaker (both the Cu-I example in the pyace package and the AlLi case in pace-tutorial-main). The fitting task was always interrupted and did not reach the set maximum number of iteration steps. The log file indicates that Optimization result(success=False, status=2, message=Desired error not necessarily achieved due to precision loss., nfev=909, njev= 897).

I'm not sure if this is normal or not and would appreciate some advice from you.

Best regards,
Zhunyun

[yury-lysogorskiy]

you have super-low metrics (RMSE, MAE, etc.) maybe you dataset is very "narrow" and thus can be fitted quickly.
Also, check this

[Nick19-bot]

Hi! i'm having the same issue,

my fitting process reaches 850 iterations for a 50 Cu structure dataset before stopping and giving the same "error" message as op,

(Optimization result(success=False, status=2, message=Desired error not necessarily achieved due to precision loss., nfev=887, njev=875)).

I also tried changing the gtol parameter to 1e-10 and 1e-2 in two separate training runs for my 50 Cu structure dataset input file as the documentation suggests but it doesn't seem to do anything,

also my final RMSE energy is around 30-40 meV/at so it's pretty high which leads me to believe the fitting process ended prematurely.

you have super-low metrics (RMSE, MAE, etc.) maybe you dataset is very "narrow" and thus can be fitted quickly. Also, check this

Yuri in your reply it says the dataset can be fitted quickly since it's "narrow" but right on the last iteration before ending the fitting process it says that the success state is false,
so that leads me to believe that the output potential isn't fully fitted and ready for use, could you correct me if i'm wrong?

Also is it viable to use the partially fitted output potential and use it in another fitting cycle as the beginning state of my potential and fit from there? i know pacemaker has the option but i don't know if this situation calls for it or if it's a reliable way to have a potential properly fitted

Thank you in advance :3

[yury-lysogorskiy]

1. Optimization result(success=False is from optimizer. You should ignore it, as it always says so. You only look on the meaningfull parameters, like error metrics of energies/forces. if they are low, then it is ok.

2. Yes, partially fitted potential can be used as initial for another fitting. At the end of the day you potential is just a point in the parameters space and optimizer is the "dynamic" of this point. So any combinations are possible, yes