Morphing Building Profile under Winds

Conventional aerodynamic shape design of tall buildings relies on a trial-and-error procedure in which wind tunnel experiments for a limited number of building forms are carried out to select the best performing one. Now the burgeoning digital revolution offers opportunities to navigate through the labyrinth of form design space. In response to this pressing need, a computational platform for aerodynamic shape tailoring of structures is developed, which takes full advantage of computational fluid dynamics (CFD) to assess wind effects on a variety of parameterized structural profiles and surrogate modeling approaches to emulate the responses from CFD simulations.

In particular, multi-fidelity surrogate modeling through fusing CFD simulation data of multiple fidelities is applied to aerodynamic shape optimization. Another important aspect in multi-fidelity surrogate modeling is its sequential design strategy. In this work, a multi-fidelity sequential design framework is proposed, in which a probabilistic classification model is introduced to avoid the low-fidelity sampling in the regions of low accuracy.

Uncertainties in CFD modeling could largely impact the reliability of CFD simulations. An uncertainty quantification framework is developed to enable inflow and model-form uncertainties to be quantified in a coupled fashion to guide the shape optimization of buildings. In particular, synthetic inflow velocity generation is extended through harnessing the power of GPUs to speed up data processing.

To venture beyond static shape optimization, autonomous morphing of a structural form is proposed in order to mitigate dynamic wind loads. This study introduces reinforcement learning agents to learn to predict the optimal state of the building’s profile in near real-time, which are prototyped on a computational platform for the proof of the concept.

A major challenge in the use of CFD as a tool to model wind effects on structures lies in the steep learning curve of CFD software posed to general CFD users. To address this issue, Jupyter Notebooks with an interactive computing environment is developed to alleviate the sophistication in running CFD models.