| Abstrakt: |
The effectiveness of various machine learning methods in predicting the quantitative properties of inorganic compounds was compared. To assess the accuracy, cross-validation was used in the mode leave-one-out cross-validation. It has been shown that the use of ridge, Bayesian ridge, automatic relevance determination regression, elastic net, and extra trees regressor programs and gradient boosting regressor and hist gradient boosting regressor, based on the methodology of ensembles of machine learning algorithms from the scikit-learn package, as well as simple syndrome regressor and recursive regressor, specially designed to solve problems in the field of inorganic materials science, allows us to obtain the most accurate estimates. Using selected machine learning programs, the melting point at atmospheric pressure of double halides of compositions ABHal3, ABHal4, A2BHal4, A2BHal5, and A3BHal6 (A and B are different elements; Hal is F, Cl, Br, or I) was predicted and the parameters of their crystal lattice under normal conditions were estimated. When estimating melting points, the average absolute errors determined by the cross-validation method were in the range of 29–52 K, depending on the composition of halides and the selected algorithm. Multiple determination coefficient R2 for models used for forecasting was not lower than 0.7. When estimating lattice parameters, the average absolute errors for linear parameters were in the range of 0.0068–0.2120 Å and were 0.1209°–0.1562° for angles. R2 was above 0.6. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink