Van Allen radiation belt electron dynamics are governed by a wide range of physical processes that can simultaneously drive acceleration, transport and loss. However, each individual process can be linked to a specific energy-dependent pitch angle distribution (PAD). We employ a new, unsupervised machine learning technique on 7-years of Van Allen Probe Relativistic Electron-Proton Telescope data and discover that six PADs, two each of: pancake, butterfly, and flattop, successfully describe >70% of classified relativistic PADs. We investigate the occurrence and storm-time evolution of each PAD through 45 geomagnetic storms. We find new populations of PADs, including: “shadowing-like” and wave-particle interaction signatures at low-L, and radial diffusion and substorm injections at higher-L, as well as determining that wave-particle interaction dominated PADs are swamped by radial diffusion processes through geomagnetic storms. Our results clearly demonstrate that PAD characterisation is a key component of understanding Van Allen radiation belt electron dynamics