Approximating Computational Fluid Dynamics for Generative Tall Building Design

Abstract

Background literature review, methodology, results, and analysis are presented for a novel approach to approximating wind pressure on tall buildings for the application of generative design exploration and optimisation. The predictions are approximations of time-averaged computational fluid dynamics (CFD) data with the aim of maintaining simulation accuracy but with improved speed. This is achieved through the use of a back-propagation artificial neural network (ANN) with vertex-based shape features as input and pressure as output. The training set consists of 600 procedurally generated tall building models, and the test set of 10 real building models; for all models in both sets, a feature vector is calculated for every vertex. Over the test set, mean absolute errors against the basis CFD are 1.99–4.44% (σ:2. 10–5.09%) with an on-line process time of 14.72–809.98s (0.028s/sample). Studies are also included on feature sensitivity, training set size, and comparison of CFD against prediction times. Results indicate that prediction time is only dependent on the number of test model vertices, and is therefore invariant to basis CFD time.