Finding New Wide-Bandgap Semiconductor Point Defects With Machine Learning : Estimation of Formation Energies and Zero-Phonon Lines for Point Defects in Silicon Carbide

Semiconductors are the cornerstone of modern technology due to its use in com-putation as a transistor. Introducing point defects in semiconductors can comewith performance increases or new applications such as quantum computing.Finding point defects with interesting properties can, however, be demandingsince it requires either experiments or computer simulations which tend to beresource-heavy. Machine learning could potentially serve as a shortcut, needingless resources and time to identify interesting point defects. This study aimsto use machine learning for predicting the zero-phonon line and the formationenergy properties of point defects consisting of vacancies, substitutions, andinterstitials, or combinations of these, in the predominant 4H silicon carbide(4H-SiC). The ADAQ database were used to train the linear ridge model, thekernel ridge model, and the multi-layer perceptron neural model. These modelswere evaluated using mean absolute error paired with a variation of 10-foldcross-validation. Two descriptors were introduced to represent the point defects:the No Symmetry Descriptor and the Symmetry Descriptor. Both consist of anatom number representing each individual point defect, the distance betweenthe point defects, and the charge and the spin state of the defect cluster. TheSymmetry Descriptor also used the Wyckoff position of the point defect or thestacking sequence of the atom layer where it resides. Due to time constraintsthe neural model was not used together with the No Symmetry Descriptor. Thedata was separated into two parts, one consisting of vacancy and substitutionaldouble defects and the other of interstitial double defects. The mean absoluteerror averages of the zero-phonon line and formation energy predictions for thevacancies and the substitutions were 0.202 eV and 1.057 eV. These values were0.230 eV and 1.729 eV for the interstitials. These results show that machinelearning could replace high-throughput calculations of point defect properties ifits accuracy is increased, either through more data acquisition or by optimizingthe models better.