UZr based metallic nuclear fuel is the leading candidate for next-generation sodium-cooled fast reactors in the United States. US research reactors have been using and testing this fuel type since the 1960s and accumulated considerable experience and knowledge about the fuel performance. However, most of knowledge remains empirical. The lack of mechanistic understanding of fuel performance is preventing the qualification of UZr fuel for commercial use. This paper proposes a data-driven approach, coupled with advanced post irradiation examination, powered by machine learning algorithms, to facilitate the development of such understandings by providing unpreceded quantified new insights into fission gas bubbles. Specifically, based on the advanced postirradiation examination data collected on a neutron-irradiated U-10Zr annular fuel, we developed a method to automatically detect, classify ~19,000 fission gas bubbles into different categories, and quantitatively link the data to lanthanide transpiration along the radial temperature gradient. The approach is versatile and can be modified to study different coupled irradiation effects, such as secondary phase redistribution and degradation of thermal conductivity, in irradiated nuclear fuel.