Surrogate Model Using A U-NET-ConvLSTM Architecture For A Macroscopic Pedestrian Flow Model

The aim of this paper is to demonstrate whether a deep neural network can replace a FEM simulation of a complex system and capture its non-linear dependencies on geometry. The system we are interested in is the macroscopic pedestrian flow in evacuation scenarios, modelled by two coupled partially differential equations (PDE) on 2D domains, possibly with holes, with mixed boundary conditions. Based on a literature review, we narrow down promising network architectures and study them with varying input variables. We develop a surrogate model based on a U-Net-ConvLSTM network. Our tests show the basic ability of such models to mimic the simulation, although further tests with larger training sets are needed.