Enhancing Mesh-Based Simulation with Graph Neural Networks: Integrating Physics for Improved Accuracy

Supervisors: Christoforos Rekatsinas, Ilias Zavitsanos, Dimitris Kelesis

Description:

This thesis investigates advanced mesh-based simulations using Graph Neural Networks (GNNs), inspired by the MeshGraphNets framework. The objective is to enhance the predictive accuracy of physical simulations by incorporating additional physics-informed features. The student will begin by reproducing the results of the reference paper using PyTorch, leveraging real-world data relevant to the problem domain. The next step involves developing an enhanced model that integrates more comprehensive physical information, with the goal of outperforming the baseline approach in terms of accuracy and generalization.

Qualifications required: Python programming (PyTorch, TorchGeometric), Machine Learning and Deep Learning algorithms