Executive Summary : | The project aims to model unsaturated flow through large granular packed beds as Graph Neural Networks, trained using data from simulations of flow through packings. This approach has several benefits, including strongly localized flow dynamics and the ability to capture the locality of unstructured data through aggregation operations on graphs like Graph Convolution Network. Once enough configurations are represented in the trained graph network, the model can be applied to any scale, making it a promising tool for predicting complex flow phenomena in industry-scale reaction engineering.
The major challenges in this project include identifying the decoder-encoder architecture choice, generating large sets of realistic particle configurations, and performing efficient flow simulations. Simulations will be performed on packings of the order of tens of particles, each forming a graph basis. These packings will be subjected to Voronoi tessellation, resulting in discretized units of space around each particle. This allows for the definition of fields such as the volume fraction of liquid and the wetting area spread on the particle domain. Critics in the project include the need for a single-phase free surface flow model, the use of cutting-edge message passing, graph encoding, and decoding techniques, and the need for adaptation to experimentally measured flow through in packed beds. The proposed project is novel in several ways, including the use of graph networks in flow problems in porous media, the emerging approach of using simulations as training data for machine learning, and the process of training graphs from piecemeal configurations in physical sciences. |