A Machine Learning Approach to Defect Parameters Extraction: Using Random Forests to Inverse the Shockley-Read-Hall Equation

Bulk defects in silicon wafers are key contributors to solar cell efficiency loss. The identification and characterization of these defects are critical steps in the process of improving the efficiency and reliability of solar cells. In this study, we present the first successful application of machine learning for extraction of defect parameters from temperature- and injection-dependent lifetime spectroscopy (TIDLS). With approximately half a million simulated TIDLS curves, random forest regressors are trained to obtain the defect’s energy level and both capture cross-sections. The high correlation coefficient between predicted and simulated values highlights the usefulness of this novel approach. With no prior knowledge of the physical model at work, the regressor learns the physical limitations of that model. This work pioneers the use of machine learning for lifetime spectroscopy, bringing the newest prowess of artificial intelligence to material quality inspection. It opens a new era in the area of defect characterization, as it has the potential to overcome the limitations of current methods, such as dealing with the temperature dependence of various electrical parameters and provide new insights in the theory of lifetime spectroscopy.