Investigation of features for prediction modeling of nanoscale conduction with time-dependent calculation of electron wave packet

Abstract A model to predict the electron transmission probability from the random impurity distribution in a two-dimensional nanowire system by combining the time evolution of the electron wave function and machine learning is proposed. We have shown that the intermediate state of the time evolution calculation is advantageous for efficient modeling by machine learning. The features for machine learning are extracted by analyzing the time variation of the electron density distribution using time evolution calculations. Consequently, the prediction error of the model is improved by performing machine learning based on the features. The proposed method provides a useful perspective for analyzing the motion of electrons in nanoscale semiconductors.