Also, I have additional question
I attempted to create an MTP (Machine Learning Interatomic Potential) for GeTe that could be applied to structures with different Ge and Te ratios.
For the ML-FF (Machine Learning Force Field) crystal training, I only used GeTe structures with a 1:1 ratio, as no other compositions were available.
However, for the ML-FF amorphous training, I generated amorphous GeTe structures using Packmol with varying Ge:Te ratios, including 1:1, 1:4, and 23:77, among others.
My question is:
Since I only used 1:1 ratio GeTe for crystal training, while incorporating various Ge:Te ratios for amorphous training, can the resulting MTP still be considered reliable?
I am currently validating the MTP, and the fitted results seem reasonable(figure posted). However, I am still uncertain whether the difference in training data composition could affect the generalization and accuracy of the MTP.
Would this discrepancy impact the reliability of my potential?
Thank you