Recently, the first-principle methods have allowed materials database in terms of property to be quickly updated and screened. However, despite the large number of such databases, only a few data exist on Gibbs energies of formation, which predicts the synthesis potential and stability of materials in various industries operating at elevated temperature. DFPT calculation obtaining for vibration contribution on materials as a function of volume has operated to reduce time costs. However, computing entropy of the phonon quantum mechanically is still cost-effectively demanding. Approximately 10 k compounds that calculated G(T) can be obtained from the Phonon DB.<br/>Although Gibbs energy prediction of inorganic materials through machine learning has recently been developed, it is still struggling to predict or improve the accuracy of polymorphic structures. Here, we use 9,880 Phonon DB included in the Materials Project as input data and 275 descriptors to construct a model to predict about 40k of Inorganic Crystalline Solid Database (ICSD) materials. Descriptors considered in this study include chemical and structural features, band gap, the number of elements and atoms and space group number. To improve the accuracy of the model, we fixed the lower 40% rmse and mae with low accuracy for the training set, which showed high accuracy of rmse = 31.9kJ/mol, mae = 18.7kJ/mol, and R<sup>2</sup> = 0.99. Uncertainty was then measured using the Standard deviation-Average Gibbs energy value of the ICSD material as the representative value. A total of 15 DFT calculations were performed except for materials that were not calculated, excluding more than 100 atoms. 573 materials were calculated, and statistics of the Standard deviation-Average Gibbs energy value shows the model's uncertainty reduction and result test accuracy to show how advanced the model has been through DFT and machine learning.

Machine learning framework for predicting Gibbs free energy of inorganic crystalline solids

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.