Wind conditions can be not only uncomfortable but also unsafe for pedestrians and residents of urban areas. A new building has a significant impact on the wind behaviour. Due to tall buildings and wind tunnel effects between buildings, strong winds can occur at street level. To counteract this, it is important to carry out wind simulations before new buildings are built. This is considered an important issue by many municipal authorities when planning new buildings in a city. For this reason, the use of generative design software in the early stages of planning new buildings in urban areas is becoming increasingly important. In order for a planning tool to be practical in the early design phase, the performance feedback of such software is an important, but still critical factor.
In the proposed work we assume that a new building complex consisting of multiple segments is constructed at a fixed location, and present an AI-based workflow to predict and optimize wind comfort (LDDC criterion). A Deep Learning model trained on simulation results by the immersed-boundary flow solver IBOFlow to predict total velocities is used as a surrogate model for the optimization process. For the calculation of the LDDC criterion, it is necessary to simulate different wind directions and evaluate them with respect to a wind rose statistic. We propose an augmentation strategy to train the surrogate model independently of a specific wind rose statistic. Using the surrogate model, we can predict the LDDC criterion with an average F1–score of 84.39 % based on only 167 simulations.
In the optimization process, the objective is to maximize the wind comfort in a given use case area so that the constraints (minimum/maximum number of floors) are respected for each segment of the building complex. We use the gradient-free solver COBYLA for this task, which takes about 200 steps to solve the optimization problem. If we had performed the optimization with the CFD solver, we would have had to run 200 individual simulations. Each simulation takes about 5 days for the specific use case, resulting in a total time saving of 1000 days (about 2.74 years). In reality, a designer of a new building would never be able to explore the solution space as it takes too much time to run the simulations. Therefore, the use of a surrogate model in the planning process seems promising.