PERC, silicon solar cells, Manufacturing and Process, Machine Learning
Summary / Abstract:
We show how at Q CELLS, interpretable machine learning algorithms are used to understand the energy conversion efficiencies of mass-produced p-type Cz-Si solar cells. For this end, we use the data of a one week ramp-up of over half a million cells acquired utilizing a single-wafer tracking system. These consist of over 300 single information per cell and feature inline measurements, path- and tool related information, process data as well as the final I-V characteristics. In the analysis using state-of-the-art machine learning algorithms, we focus on understanding how the many different input features influence the solar cell efficiency over time by using additive feature impacts. As sources of variation are spread over several features, we pay special attention to correlated features in a hierarchical clustering approach. Finally, after ensuring the model validity at the hands of a hold out test data set (Pearson’s coefficient of correlation of 0.84 between true and predicted values), subtle hourly cell efficiency fluctuations in the order of 0.01%abs are explained. We identify features which are relevant for short-term changes in the efficiency and other which influence the efficiency on a general level and do not show temporal changes.