Insights into the Effects of RbF‐Post‐Deposition Treatments on the Absorber Surface of High Efficiency Cu(In,Ga)Se₂ Solar Cells and Development of Analytical and Machine Learning Process Monitoring Methodologies Based on Combinatorial Analysis

Abstract

The latest record efficiencies of the Cu(In,Ga)Se2 (CIGSe) photovoltaic technology are driven by heavy alkali post‐deposition treatments (PDTs). Despite their positive effect, especially in the Voc of the devices, their underlying mechanisms are still under discussion. This work sheds light on this topic by performing a high statistics analysis on 620 high efficiency CIGSe solar cells submitted to four different PDT processes (different RbF source temperature) employing a combinatorial approach based on Raman and photoluminescence (PL) spectroscopies. This reveals with statistical confidence subtle differences in the spectroscopic data that relate to the redistribution of defects between the ordered vacancy compound (OVC) and the chalcopyrite phases at the absorber surface. In particular, there is an intertwined decrease of the OVC phase and increase of the so‐called “defective chalcopyrite phase.” Additionally, two industry‐compatible methodologies for the assessment of the RbF‐PDT process and prediction of the Voc of the final devices with a ±2% error and an efficacy of ≈95% are developed based on analytical and machine learning approaches. These results show that the combined Raman and PL spectroscopic techniques represent a powerful tool for the future development of the CIGSe technology at a research level and for more efficient industrial manufacturing.

Combinatorial Analysis Methodologies for Accelerated Research: The Case of Chalcogenide Thin‐Film Photovoltaic Technologies

“…Similarly, in the case of photoluminescence spectra, the changes in the intensity of the bands have to be determined as they can be associated with changes in the nonradiative recombination processes. [16,[106][107][108][109] All these pretreatments can be easily performed by applying several existing software solutions, although in most cases this has to be done only in manual mode which significantly increases the data pretreatment and treatment process times. Taking this into account, in the case of spectroscopic techniques (as Raman and PL), the group from IREC has developed a Python package spectrapepper, which includes several useful functions allowing to treat big amount of data in an easy and fast way.…”

“…[75] In addition to the materials study, combinatorial analysis of the CZTSe samples allowed to develop nondestructive methodologies based on Reproduced with permission. [16] Copyright 2014, John Wiley and Sons.…”

“…3) Automatized mappings: In parallel to the measurement's time limitations, it is essential to have the possibility to implement automated mapping measurement procedures; these allow not only to perform the measurement without the need for the permanent presence of a researcher, but also to align the measurement points used in the different techniques for better results correlation. 4) Combined techniques: Combining several techniques in parallel measurements can greatly decrease the total measuring time; however, this option is barely realized in commercial available systems, and these are rather customized solutions (e.g., see another study [ 16 ] ). 5) Spatial resolution should be high enough to resolve significant property variations in a graded sample.…”

“…This allows the achievement of a large number of samples, allowing to ensure the statistical validity of the performed study. [ 16 ] As in the conventional combinatorial approach, this high number of samples is studied by different characterization techniques that allow rapid measurements. In this case, the analysis of the massive dataset provides trends of the impacts of different physicochemical properties on the materials that would not be identified with the use of a low statistical set of data, thus enabling finer material optimization that otherwise would be impossible.…”

“…Right: comparison of the V oc predictive power for two different criteria with either separated Raman or PL data or a combination of both. Reproduced with permission [16]. Copyright 2014, John Wiley and Sons.…”