Life of Research
Appearance

Physics-Informed Machine Learning for Aerodynamic Flow Prediction and Optimisation of Next-Generation Aircraft

  • Dominic Namachila
  1. University of Cape Town

Thesis / Dissertation

LOR-00209

Download PDF (973.7 KB) Cite

Cite

Physics-Informed Machine Learning for Aerodynamic Flow Prediction and Optimisation of Next-Generation Aircraft

Read

Abstract

The design of next-generation aircraft—blended-wing-body configurations, truss-braced wings, distributed-propulsion concepts and hydrogen-powered airframes—requires the exploration of design spaces that are far larger and less well understood than those of conventional tube-and-wing transports. High-fidelity computational fluid dynamics (CFD)
remains the workhorse of aerodynamic analysis, but its cost limits the number of designs that can be evaluated, and purely data-driven surrogates trained on CFD databases generalise poorly beyond their training distribution and offer no guarantee that predictions respect conservation laws.
This thesis develops a physics-informed machine learning (PIML) framework that embeds the governing equations of compressible, turbulent flow directly into the training of neural network surrogates, and couples these surrogates to gradient-based aerodynamic shape optimisation. Three contributions are pursued. First, a hybrid physics-informed neural operator is formulated that maps aerofoil and wing geometry, together with flight conditions, to the full Reynolds-averaged Navier–Stokes (RANS) flow field, using a composite loss in which sparse CFD data are regularised by PDE residuals, boundary conditions and integral force constraints. Second, training pathologies that are known to afflict physicsinformed networks in aerodynamic settings—stiff loss landscapes, spectral bias towards lowfrequency solutions, and conflicting gradients between loss terms—are addressed through adaptive loss balancing, Fourier feature embeddings, boundary-layer-aware sampling and hard enforcement of wall boundary conditions. Third, the trained surrogate is embedded in a multi-fidelity, trust-region shape optimisation loop in which design sensitivities are obtained by automatic differentiation and predictive uncertainty, estimated with deep ensembles, governs when the surrogate is refined with new high-fidelity CFD samples.
The framework is developed on two-dimensional aerofoil benchmarks and extended to three-dimensional transonic wings representative of next-generation transport aircraft.
A potential-flow panel method implemented as part of the data-generation pipeline is verified against classical thin-aerofoil theory. The thesis concludes with a critical discussion of the accuracy, cost, robustness and certification implications of using physics-informed surrogates in the conceptual and preliminary design of aircraft.

Files (1)

Cite this record

Namachila, D. (2024). Physics-Informed Machine Learning for Aerodynamic Flow Prediction and Optimisation of Next-Generation Aircraft [Thesis]. Life of Research. https://lifeofresearch.com/theses/209

APA 7th edition
Export to a reference manager
BibTeX
@phdthesis{namachila2024physics,
  author       = {Namachila, Dominic},
  title        = {{Physics-Informed Machine Learning for Aerodynamic Flow Prediction and Optimisation of Next-Generation Aircraft}},
  year         = {2024},
  month        = dec,
  date         = {2024-12-15},
  type         = {Thesis},
  url          = {https://lifeofresearch.com/theses/209},
  abstract     = {The design of next-generation aircraft—blended-wing-body configurations, truss-braced wings, distributed-propulsion concepts and hydrogen-powered airframes—requires the exploration of design spaces that are far larger and less well understood than those of conventional tube-and-wing transports. High-fidelity computational fluid dynamics (CFD) remains the workhorse of aerodynamic analysis, but its cost limits the number of designs that can be evaluated, and purely data-driven surrogates trained on CFD databases generalise poorly beyond their training distribution and offer no guarantee that predictions respect conservation laws. This thesis develops a physics-informed machine learning (PIML) framework that embeds the governing equations of compressible, turbulent flow directly into the training of neural network surrogates, and couples these surrogates to gradient-based aerodynamic shape optimisation. Three contributions are pursued. First, a hybrid physics-informed neural operator is formulated that maps aerofoil and wing geometry, together with flight conditions, to the full Reynolds-averaged Navier–Stokes (RANS) flow field, using a composite loss in which sparse CFD data are regularised by PDE residuals, boundary conditions and integral force constraints. Second, training pathologies that are known to afflict physicsinformed networks in aerodynamic settings—stiff loss landscapes, spectral bias towards lowfrequency solutions, and conflicting gradients between loss terms—are addressed through adaptive loss balancing, Fourier feature embeddings, boundary-layer-aware sampling and hard enforcement of wall boundary conditions. Third, the trained surrogate is embedded in a multi-fidelity, trust-region shape optimisation loop in which design sensitivities are obtained by automatic differentiation and predictive uncertainty, estimated with deep ensembles, governs when the surrogate is refined with new high-fidelity CFD samples. The framework is developed on two-dimensional aerofoil benchmarks and extended to three-dimensional transonic wings representative of next-generation transport aircraft. A potential-flow panel method implemented as part of the data-generation pipeline is verified against classical thin-aerofoil theory. The thesis concludes with a critical discussion of the accuracy, cost, robustness and certification implications of using physics-informed surrogates in the conceptual and preliminary design of aircraft.},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}

Using Zotero or Mendeley? Their browser connector saves this record from this page, with its PDF when there is one. On search and browse pages it lists every result so you can save many at once.

Guides What a complete citation needs What CC BY 4.0 allows