E-Thesis 65 views 60 downloads
AI Mesh Informed Techniques for Optimising the Design Process / CALLUM LOCK
Swansea University Author: CALLUM LOCK
DOI (Published version): 10.23889/SUThesis.71072
Abstract
This thesis presents a novel, data-driven framework for automatically generating near-optimal unstructured meshes for computational simulations. The primary objective is to reduce the manual effort and expert intervention typically required in mesh generation by leveraging historical simulation data...
| Published: |
Swansea
2025
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Hassan, O.; Sevilla, R.; and Jones, J. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa71072 |
| Abstract: |
This thesis presents a novel, data-driven framework for automatically generating near-optimal unstructured meshes for computational simulations. The primary objective is to reduce the manual effort and expert intervention typically required in mesh generation by leveraging historical simulation data and artificial neural networks (ANNs) to predict appropriate mesh spacing fields. The work is motivated by the growing availability of high-fidelity simulation data in industry and the need to streamline simulation workflows – particularly in the aerospace sector, where the mesh generation process remains one of the most resource-intensive steps.Three different strategies are developed and evaluated. The first approach predicts the properties of point sources used to define the mesh resolution. The second introduces a coarse background mesh, onto which spacing functions are conservatively interpolated and predicted by ANNs. The third and final approach extends the method to fully anisotropic spacing by predicting the components of the metric tensor, allowing for directionally aligned mesh refinement. All three techniques are trained on datasets de-rived from prior simulations and are shown to generalise effectively to unseen geometric and flow conditions.Extensive numerical experiments in three-dimensional compressible flow scenarios –including wings and full aircraft configurations, demonstrate that the proposed methods yield high-quality meshes capable of producing accurate solutions. Furthermore, an environmental impact analysis shows the potential for a substantial reduction in computational cost and energy usage, highlighting the ability of the methods outlined to be part of sustainable simulation practices.This work lays the foundation for integrating machine learning into the meshing pipeline, enabling intelligent, scalable, and more efficient simulation-driven design across a wide range of engineering applications. |
|---|---|
| Keywords: |
Mesh generation, Machine learning, Near-optimal mesh prediction, Computational fluiddynamics |
| College: |
Faculty of Science and Engineering |
| Funders: |
EPSRC, Airbus Defence and Space |