Low dimensional fragment-based descriptors for property predictions in inorganic materials with machine learning

In recent times, the use of machine learning in materials design and discovery has aided to accelerate the discovery of innovative materials with extraordinary properties, which otherwise would have been driven by a laborious and time-consuming trial-and-error process. In this study, a simple yet powerful fragment-based descriptor, Low Dimensional Fragment Descriptors (LDFD), is proposed to work in conjunction with machine learning models to predict important properties of a wide range of inorganic materials such as perovskite oxides, metal halide perovskites, alloys, semiconductor, and other materials system and can also be extended to work with interfaces. To predict properties, the generation of descriptors requires only the structural formula of the materials and, in presence of identical structure in the dataset, additional system properties as input. And the generation of descriptors involves few steps, encoding the formula in binary space and reduction of dimensionality, allowing easy implementation and prediction. To evaluate descriptor performance, six known datasets with up to eight components were compared. The method was applied to properties such as band gaps of perovskites and semiconductors, lattice constant of magnetic alloys, bulk/shear modulus of superhard alloys, critical temperature of superconductors, formation enthalpy and energy above hull convex of perovskite oxides. An advanced python-based data mining tool matminer was utilized for the collection of data. The prediction accuracies are equivalent to the quality of the training data and show comparable effectiveness as previous studies. This method should be extendable to any inorganic material systems which can be subdivided into layers or crystal structures with more than one atom site, and with the progress of data mining the performance should get better with larger and unbiased datasets.