Efficient Quality Diversity Optimization of 3D Buildings through 2D Pre-Optimization.

Area of Science:

• Computational Engineering
• Machine Learning
• Fluid Dynamics

Background:

• Quality diversity (QD) algorithms efficiently generate diverse solutions but are computationally expensive for high-dimensional, costly problems.
• Existing QD methods require hundreds to thousands of evaluations, even with surrogate models, limiting their application in real-world engineering.
• Predicting complex phenomena like wind nuisance in 3D building design demands computationally intensive simulations.

Purpose of the Study:

• To develop a more efficient QD approach for computationally expensive problems by leveraging pre-optimization in lower dimensions.
• To demonstrate the method's effectiveness in predicting and minimizing wind nuisance for 3D building designs.
• To improve the accuracy of predictive models by using QD-generated training data compared to traditional sampling methods.

Main Methods:

• Implemented a pre-optimization strategy by solving a lower-dimensional problem and mapping solutions to a higher-dimensional design space.
• Utilized a quality diversity algorithm to sample 2D building footprints, generating diverse training data.
• Trained a predictive model using QD-sampled data to forecast 3D building wind nuisance from 2D footprint features.

Main Results:

• Achieved accurate prediction of 3D building flow features from 2D footprint data.
• QD-generated training data resulted in a more accurate predictive model than space-filling (Sobol sequence) sampling.
• Generated 1,024 building designs with low predicted wind nuisance by simulating only 16 buildings in 3D.

Conclusions:

• The proposed pre-optimization strategy enhances the efficiency of quality diversity algorithms for expensive optimization problems.
• Using quality diversity for training data generation significantly improves machine learning model performance in computational domains.
• This method enables generative design in computationally demanding 3D environments, aiding early-stage engineering design and analysis of wind nuisance.