MScE Defence - Irtiza Khan - Department of Mechanical Engineering-FR

Event date(s): August 26, 2026
Time(s): 10:30 AM - 12:30 PM
Category: Fredericton
Location: Fredericton

Abstract

High-fidelity computational fluid dynamics (CFD) resolves the viscous flow interactions essential for realistic engineering design, yet remains uncommon in direct three-dimensional (3D) geometry optimization. Two obstacles limit its use: individual evaluations are computationally expensive, and the surrounding workflow is difficult to execute without manual intervention. Each component of the pipeline is also non-trivial. This thesis develops an end-to-end workflow coupling automated geometry generation, CAD construction, meshing, and Reynolds-averaged Navier–Stokes (RANS) simulation with Gaussian process (GP)-based Bayesian optimization. The workflow is tested on the propeller of the Royal Canadian Navy’s ORCA-class patrol vessel at model scale in the hull wake. Radial chord and pitch distributions are parameterized using two-segment rational cubic Bézier curves, producing an 11-variable design space. Each candidate blade is evaluated at its self-propulsion point, where propeller thrust balances total resistance. Among 250 successful 3D RANS evaluations, the best candidate increased propulsive efficiency from 0.5093 to 0.5307, a gain of 2.14 percentage points. Structural and cavitation constraints were intentionally excluded because the objective was to validate the workflow rather than present a final ORCA propeller design. Approximately one-third of the candidates failed during automated meshing, hindering full convergence of the optimization. This result reflects the well-known difficulty of mesh automation in CFD-driven design. These failures motivated further investigation using a less expensive two-dimensional CFD benchmark. The resulting strategy retains failed mesh or CFD cases through penalized imputation. It also dynamically reallocates samples among complementary acquisition strategies based on recent performance. Against a particle swarm optimization (PSO) baseline, it matched the best PSO designs using approximately 42% fewer successful CFD evaluations. Penalized imputation also reduced the average failure rate from 12.9% to 4.4% while maintaining similar final performance. These benefits are expected to become more consequential when applied to expensive 3D problems. Overall, an automated, failure-aware, surrogate-assisted optimization workflow is presented that can make RANS-based design studies practical, provided that mesh robustness and physical constraints are handled with care.

Building: Head Hall, 15 Dineen Drive

Room Number: TME Room 224