We present a machine learning approach for evaluating thermodynamic properties within model compositional/configurational spaces (CCSs) of complex intermetallics with Mackay-type building units. Recently, by performing high-throughput density functional theory calculations, we determine and analyze the formation energies for the recently synthesized 1/1 approximant of icosahedral quasicrystals in the Sc-Pd system possessing structural disorder. Here, sequentially increasing training sets are used in regression models based on compositional and topological descriptors for predicting formation enthalpies. We investigate different learning strategies. The obtained R2 dependencies show higher learning rates for the models using real features (composition and contact counters of defects) with small training set sizes, whereas using categorical features (topological types of defect structure fragments) result in a slightly higher prediction quality when the training set size increases. In connection with the obtained results, the proposed approach is considered as a tool for investigations of disordering phenomena in complex intermetallics as well as for a subsequent searching for new stable phases of Mackay-type approximants of quasicrystals.
A machine learning approach for predicting formation enthalpy: A case study of Mackay-type approximants of icosahedral quasicrystals
Solokha P.
2019-01-01
Abstract
We present a machine learning approach for evaluating thermodynamic properties within model compositional/configurational spaces (CCSs) of complex intermetallics with Mackay-type building units. Recently, by performing high-throughput density functional theory calculations, we determine and analyze the formation energies for the recently synthesized 1/1 approximant of icosahedral quasicrystals in the Sc-Pd system possessing structural disorder. Here, sequentially increasing training sets are used in regression models based on compositional and topological descriptors for predicting formation enthalpies. We investigate different learning strategies. The obtained R2 dependencies show higher learning rates for the models using real features (composition and contact counters of defects) with small training set sizes, whereas using categorical features (topological types of defect structure fragments) result in a slightly higher prediction quality when the training set size increases. In connection with the obtained results, the proposed approach is considered as a tool for investigations of disordering phenomena in complex intermetallics as well as for a subsequent searching for new stable phases of Mackay-type approximants of quasicrystals.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
