The dynamics of neuron populations during diverse tasks often evolve on low-dimensional manifolds. However, it remains challenging to discern the contributions of geometry and dynamics for encoding relevant behavioural variables. Here, we introduce an unsupervised geometric deep learning framework for representing non-linear dynamical systems based on statistical distributions of local phase portrait features. Our method provides robust geometry-aware or geometry-agnostic representations for the unbiased comparison of dynamics based on measured trajectories. We demonstrate that our statistical representation can generalise across neural network instances to discriminate computational mechanisms, obtain interpretable embeddings of neural dynamics in a primate reaching task with geometric correspondence to hand kinematics, and develop a decoding algorithm with state-of-the-art accuracy. Our results highlight the importance of using the intrinsic manifold structure over temporal information to develop better decoding algorithms and assimilate data across experiments.