Estimating the band gap of materials with machine learning methods

Abstract

Methods of machine learning have shown significant progress in the last decade. The number and quality of applications of these methods in various fields of physics are also increasing. Machine learning algorithms are mathematical models that can learn patterns in a data set and estimate the values of the target label afterward. The selection and optimization of a learning algorithm depend on the problem and the structure of the used data set. Processing this data set and selecting features before training a model is also important. Predicting the band gap of different types of materials with machine learning methods while investigating methods that are used for model optimization is the main objective of this thesis. Knowing the band structures is especially important in environmental technologies, such as solar panels or light-emitting diodes. They consist of semiconductor devices, and like all semiconductor materials, their band gap determines their conductivity. Forecasting material properties in physics is challenging because of the long work hours and computational resources required to dedicate to related experiments or simulations. Machine learning can offer a solution to these problems and increase overall efficiency by decreasing workloads. The proposed random forest model was developed and optimized in Python programming language. A custom feature selector algorithm that utilizes multiple metrics as a feedback tool to select optimal features also improves the performance of the final model. Results show that the optimized random forest model can predict band gaps of the materials in the Citrine and Matminer data sets with less than the mean absolute error value of 0.500 eV and an R2 score higher than 0.800.