Design and evaluation of generative deep learning models for electrochemical device microstructure image augmentation

Abstract

The issue of making more efficient batteries is ever present in today’s world. Researchers often have a problem with having a variety of samples to study the inner structure of the batteries. At the same time, the development of generative AI became rapid. It was important to consider the integration of such systems into the energy sector to use its benefits by having opportunities of saving material production and increase the learning efficiency. An effective tool turned out to be Generative Adversarial Network (GAN) that is able to mimic the physical structure of the models to recreate the CT slice scans of the model, to produce an efficient output for further battery improvement. The main concept is to train the model based on the real segmented CT scan dataset, to then generate a 3D model that looks accurate both visually and in terms of physical and chemical properties. The correctness will be evaluated with Tau factor that is responsible for tortuosity calculation, and generated model will be vigorously compared with the original dataset on a pixel level.