Autonomous sonar target recognition suffers from uncertainty caused by waveguide distortions to signal, unknown target geometry, and morphing target features. Typical “black-box” neural networks do not produce physically interpretable features and, therefore, are not effective in meeting these challenges. The primary objective of our work is to harness signal processing with machine learning to extract braided features that allow such physical interpretation by a domain expert. In this work, we introduce a feature extraction method using graph neural networks (GNNs) that seeks to discover braid manifolds from sonar magnitude spectra data. The approach involves representing the sonar magnitude spectra as sparse, dynamic graphs. These dynamic graphs can then be fed into a GNN as sequences of timed events to produce feature dictionaries that are resilient to environmental uncertainty and agnostic to ping direction. The ability of GNNs to learn complex systems of interactions makes them a great choice for braid-like feature discovery. To handle the evolving dynamic features of the sonar spectra graphs, a variation of a GNN, called temporal graph networks (TGNs), is used. TGNs utilize memory modules and graph-based operators to outperform previous GNN-based approaches when handling dynamic graphs. We use TGNs to model the evolution of the sonar spectra graphs and ultimately perform graph-based classification. Preliminary results performed on the Malta Plateau field experiment are presented.