The migration of atoms and small clusters is an important process in sub-nanometre scale heterogeneous catalysis, affecting activity, accessibility and deactivation through sintering. Control of migration can be partially achieved via encapsulation of sub-nanometre metal particles into porous media such as zeolites. However, a general understanding of the migration mechanisms and their sensitivity to particle size and framework environment is lacking. Here, we extend the time-scale and sampling of atomistic simulations of platinum cluster diffusion in siliceous zeolite frameworks, by introducing a reactive neural network potential of density functional quality. We observe that Pt atoms migrate in a qualitatively different manner from clusters, occupying the dense region of the framework and avoiding the free pore space. We also find that for cage-like zeolite CHA there exists a maximum in self diffusivity for the Pt dimer beyond which, confinement effects hinder intercage migration. By extending the quality of sampling, NNP-based methods allow for the discovery of novel dynamical processes at the atomistic scale, bringing modelling closer to operando experimental characterization of catalytic materials.