Quantitative trends in 8 physical properties of 115000 inorganic compounds gained by machine learning

TL;DR

This study uses machine learning to analyze and predict eight physical properties of nearly 115,000 inorganic compounds, revealing periodic patterns and enabling structure-property predictions, thus creating a materials 'periodic table.'

Contribution

Introduces a machine learning approach to predict multiple properties of inorganic compounds and uncovers periodic trends across a vast dataset, advancing materials informatics.

Findings

Identified periodic patterns in physical properties across compound types.

Successfully predicted properties using only crystalline structures.

Reversed the prediction task to infer structures from properties.

Abstract

We applied the decision trees (random forest) machine-learning technique for the large experimental materials dataset PAULING FILE, compiled from the world's peer-reviewed literature. The training and validation data were extracted from the hundreds of thousands of publications in materials science (1891-2017). Then, for the nearly 115'000 distinct inorganic compounds we predicted 8 thermodynamic, mechanical, and electronic properties, using the only crystalline structures as an input. For the predicted physical properties we observed certain periodical patterns in all unary, binary, ternary, and quaternary compounds. We also solved a reversed task of predicting the possible crystalline structure based on a given combination of values of the 8 mentioned properties. Therefore our observations may play a role of the periodic table, formulated not for the chemical elements, but for the entire set of materials.