A comparison of neural network architectures for data-driven reduced-order modeling

The popularity of deep convolutional autoencoders (CAEs) has engendered new and effective reduced-order models (ROMs) for the simulation large-scale dynamical systems. Despite this, it is still unknown whether CAEs provide superior performance over established linear techniques or other network-based methods in all modeling scenarios. To elucidate effect autoencoder architecture on its associated ROM studied through comparison against two alternatives: a simple fully connected autoencoder, novel graph autoencoder. Through benchmark experiments, shown that given application highly dependent size latent space structure snapshot data, with proposed demonstrating benefits data irregular connectivity when sufficiently large.