Demonstrating Sensitivity Analysis using Parametric Urban Design and a GPU-based Lattice-Boltzmann Solver

Abstract

In order to adapt cities to the changes in climate, it is essential to understand the impact of design decisions on the local urban climate early in the planning and architectural process. Our contribution to this goal is to demonstrate a method for coupling parametric design with a GPU-based lattice Boltzmann method (LBM) solver to calculate the sensitivity of the local urban climate with respect to architectural design decisions. As a proof of concept, we performed a sensitivity analysis of the cold airflow through a construction site, down a partially greened slope, and into the city of Bonn, Germany, so cooling the city during summer. Official guidelines in Germany state that keeping slopes free of buildings parallel to the slope leads to more cold air flowing into the valley. They also acknowledge that tree canopies and bushes can hinder air flow. The latter is true for our use case, while first low fidelity simulations indicate that building parallel to the greened slope may mitigate this effect. A sensitivity analysis of the impact of planning designs on turbulent flow is necessary to improve planning decisions and resolve the interaction of the effects. A CAD kernel supporting algorithmic differentiation (AD) provides sensitivities of surface point coordinates of an architectural design regarding parameter changes. Propagating these surface point sensitivities throughout the simulation yields the sensitivities of simulation results with respect to design parameters via forward-mode AD. The information gained from this process can be used to steer these design parameters, either manually or automatically, towards a parameter choice that is more beneficial to the local climate. The simulation is based on LBM, a mesoscopic computational fluid dynamics (CFD) method approximating microscale fluid dynamics as particle flows on an equidistant grid (lattice). The LBM solver lettuce models the airflow over the slope and buildings. Using GPU parallelization with the Python package PyTorch, sensitivies can be tracked throughout the simulation. The area in question can be resolved in a Large Eddy Simulation (LES) on high-performance GPUs. In future work, this method will also allow us to perform optimization tasks in urban design, in particular the optimization of curved architectural shapes with respect to their impact on local urban climate.