Transfer learning of interatomic potentials: from molecules to crystals

Strojno učenje ima mogućnost promijeniti paradigmu izračuna svojstava kompleksnih molekula i kristala. Dosadašnje precizne metode bazirane na teoriji funkcionala gustoće su računalno preskupe za složene zadatke kao što je predviđanje kristalnih struktura. U ovom radu je dan pregled primjene suvremenih metoda strojnog kao i dubokog učenja na problem određivanja interatomskih potencijala i sila. Osim toga, testiran je prijašnji strojno naučeni potencijal na strukturama sedmog slijepog testa predviđanja kristalnih struktura te su mu prijenosnim učenjem značajno popravljene performanse.Machine learning has the ability to change the paradigm of calculating the properties of complex molecules and crystals. Previous precise methods based on density functional theory are computationally too expensive for complex tasks such as the prediction of crystal structures. This paper provides an overview of the application of modern methods of machine as well as deep learning to the problem of determining interatomic potentials and forces. In addition, the previous machine learning potential was tested on the structures of the seventh blind prediction test of crystal structures, and its performance was significantly improved by transfer learning