Reduced graph-based representations of detonation-containing flows

Abstract

A graph neural network (GNN)-based autoencoder is leveraged to produce reduced latent graph representations of detonation-containing flows. In a two-dimensional hydrogen-air channel configuration with a uniform fuel-air mixture, it was found that latent graphs with effective resolutions of twice the induction length are capable of reconstructing both the wavefront and the post-shock region of the cellular detonation wave structure. Extrapolation studies using stratified and three-dimensional detonation configurations revealed an intrinsic property of GNN-based reconstruction from reduced latent graphs -- a representational hierarchy, rooted in the local nature of message passing, that tends to prioritize features at the detonation front (i.e., the wavefront profile and triple points) over the downstream deflagrative regions. In particular, it was found that at a fixed ambient condition, the information contained in uniform two-dimensional detonations was sufficient to reconstruct the wavefront structure of both stratified and three-dimensional detonations. Overall, this work demonstrates the utility of GNNs for both reduction and information overlap extraction in detonation-containing flows, opening promising pathways for accelerated simulations of these flows.