Microscopic spatial defects (inhomogeneities) in solar cells have a detrimental impact on the overall performance of the solar cells. These defects can be due to the polycrystalline nature of the photovoltaic absorber (general case) or on the other hand, the characterization method itself can induce such inhomogeneities (ex: local excitation with confocal system). Photoluminescence imaging in particular is the most attractive type of experimental character- ization technique which has been studied by many research groups. Since it is contactless, it allows a complete analysis of the photovoltaic material and it can actually be performed at each step of the solar cell fabrication process. To properly analyze the recorded images, one has to model the transport properties. In this work we model the transport properties in 2D using both a numerical and analytic approach. First,we model the global illumination of the sample, we then analyze the effects of grains and surface recombination. Second, we study lateral transport that can be influenced by recombination at the surface (passivation issues), grain boundaries (polycrystalline cells) or local artifacts (shunts, defects...). At the same time we are able to extract the lifetime knowing the generation rate and solving the excess carrier density from our model. And finally we implement our model to extract some cell parameters like the diffusion length from experimental data through data fitting.
In most corruption scandals, the use of front companies for money laundering is almost ubiquitous. This work proposes to apply image classification to detect such organizations, through the use of Convolutional Neural Networks (CNN), namely the AlexNet architecture. The images are obtained by address search in Google Street View API, and the resulting classification will be further used along with other features to detect front com- panies in order to help the auditors from the Ministry of Transparency and Office of the Comptroller General (CGU, in Portuguese). To this moment, we applied classification to almost 15 thousand suppliers scenes with active contracts with the Brazilian Government until September 2016, obtained through data matching between the Government Purchases database and the Brazilian Federal Revenue Office database (more recent scenes should be added as this work progresses). Preliminary results with a pre-trained AlexNet CNN show the need for developing new scene classes more suited to the Brazilian context. In order to do this, we propose to apply clustering algorithms in features extracted from the last fully-connected layer of this net. The classes obtained will be used to fine-tune the AlexNet CNN for future classification, through the use of training from scratch or fine tuning techniques.