Approximating computational fluid dynamics for generative design

Approximating computational fluid dynamics for generative design

By Samuel Wilkinson

Supervisors:

Sean Hanna

Tadj Oreszczyn

Industrial Supervisor/Sponsor:

Lars Hesselgren (PLP Architecture)

Volker Mueller (Bentley Systems)

2010 to 2014

The partnership between Bentley, as software developer, and PLP, as designer, is focused on creating prototypical analysis tools for GenerativeComponents, a parametric CAD software integrated with Microstation. The scope of the EngD is on generative tall building design and wind modelling through computational fluid dynamics (CFD). 

A novel approach is explored whereby machine learning is used to approximate or reduce steady-state CFD results in order to enable faster and broader design space searches and optimisation. It is demonstrated that wind-induced surface pressure can be locally approximated by intrinsic shape features, and that the method dominates existing solver approaches in both time and accuracy. The work also test the scalability of the approach in three aspects: to test geometry of realistic complexity; to time-dependent peak pressure from large eddy simulation; and to the localised approximation of urban interference.