Stephan Heise

In light of the accumulation of characterization measurement data in the industrial production of solar cell devices, the investigation of a large amount of samples by statistical means lends itself to be a useful tool to gain further insights into how the data correlate with performance parameters. However, due to the multicollinearity among high-dimensional input parameters of compositional data, revealing the underlying patterns may prove to be a difficult endeavor. In this work, we present statistics consisting of 280 thin-film solar cell samples based on Cu(In, Ga)(S, Se)2 absorber layers whose depth-resolved composition was assessed by glow-discharge optical emission spectroscopy (GDOES). After parameterization of the features of [Ga]/([Ga] + [In])and[S]/([S] + [Se]) gradings, we employ two-way clustering in order to group samples and features by their similarity. In addition, using principal component analysis, information in the dataset, which is irrelevant to the problem, is removed by dimensionality reduction. In this way, it is possible to create a map that provides an overview of the GDOES data of all samples in their entirety, including correlations among features. More importantly, it also opens up a more precise way to plan further improvements in the compositional gradings by unveiling a path along which the experimenter can read the feature changes concerned with an improvement in the open-circuit voltage deficit or any other target parameter of interest. New samples can then be assigned to existing cluster centroids to predict what target parameter value they would assume.