Machine Learning

Machine learning is useful to identify rules hidden in given data and to predict unknown data using the identified rules. It has been increasingly used in many real-world applications such as speech recognition, image recognition, and machine language translation. Recently machine learning methods have been applied to analyze microscopic measurement data and theoretical simulation data in materials science. This chapter introduces the basic concepts and methods of supervised problem settings, including linear models, tree-based models (decision tree and random forest), and neural networks. In order to apply these methods to real problems, overfitting problems and some techniques for evaluating prediction performance of trained supervised models are introduced. As examples of machine learning methods in materials science, machine learning potentials, which can be used for predicting formation energies and the forces of atoms in material structures, are introduced. Once machine learning potentials are constructed using training datasets obtained from theoretical computation based on quantum mechanics, they can predict materials properties related to atom dynamics with much lighter computational cost than the methods adopted in obtaining training datasets. Overview of such machine learning potentials is given together with the application examples of one of such potentials, high-dimensional neural network potential, to materials property predictions.