Li-ion batteries, magnets and fuel cells are of high importance for the adoption and success of renewable energy and e-mobility. These material systems show a high complexity and must be protected against aging in the field. Imaging techniques enable ever greater high-resolution data acquisition. Quality control tasks are therefore becoming increasingly complex, to the point where manual inspection is no longer possible.

In this work we present an overview of different deep learning methods and how they can be applied for automatic evaluation of large amounts of image data of energy materials. Examples include object detection and semantic segmentation for the classification and quantification of different defect types in Li-ion batteries and fuel cells. These supervised learning approaches are applicable if defect classes (like foreign particles and inclusions, cracks, delaminations, deformations, etc.) are already known, well defined and not too rare to build representative datasets. For magnets, defect detection focuses more on the identification of irregularities which are less specific and manifest themselves as deviations from an ideal microstructure. Unsupervised deep learning approaches are used to implement anomaly detection methods for these kinds of defects, as they don´t depend on labeled images of defects and can be trained on regular micrographs only. In addition, semantic segmentation is used to implement quantification of specific phases, which allows for localization of sample areas that contain high phase fractions of unwanted phases in the magnet.