Predicting solar cell efficiencies using historical data from a manufacturing process

Abstract The solar cell manufacturing data of a passivated emitter and rear cell solar cell manufacturing plant was studied to assess the effects of tool usage and the processing time spent on each tool on the solar cell efficiency. Since manufacturing processes involve several steps with multiple tools, tracing their quality parameters back to the tool usage is difficult—most plants measure parameters at the end of the manufacturing process. We used multilinear regression, partial least squares (PLS), kernel based PLS, random forest, gradient boosted decision trees (GBDTs), and 2D convolutional neural network (CNN) models to study the variation of the cell efficiencies with variations in tools and tool processing times. We evaluated our models' performance using mean squared error (MSE) and the coefficient of determination. The GBDT had the best coefficient of determination, but with a relatively higher MSE for efficiency prediction using batch processing times. Shapley analysis was used to study the effect of the frequency of tool usage on efficiency prediction and identify process steps and tools having maximum impact on efficiency. This work can be used as a basis to selectively utilize tools in the solar cell manufacturing process that will result in better solar cell quality (or reduce material wastage due to poor quality), thereby making solar cells more affordable and hence more readily adoptable. It can also be translated to other industries whose manufacturing processes involve multiple tools/steps.