FLICK: A Lightweight Tool for Revolutionizing Urban-Scale Wind Field Simulation with Neural Networks

The FLICK project, jointly developed by the Barcelona Supercomputing Center and the Polytechnic University of Catalonia, uses neural networks to achieve fast and lightweight modeling of urban-scale wind fields. Addressing the pain point that traditional high-fidelity Computational Fluid Dynamics (CFD) simulations are extremely costly and cannot meet real-time decision-making needs, this project increases simulation speed by several orders of magnitude while maintaining reasonable accuracy, providing an efficient tool for climate research and urban planning.

Urban wind field distribution is affected by complex surface features such as buildings and streets. Traditional CFD simulations (e.g., Large Eddy Simulation, LES) require significant computational resources to solve the Navier-Stokes equations, and a single simulation may take hours or even days. This leads to practical application dilemmas: urban planners cannot quickly evaluate the wind environment impact of different building layouts; climate researchers struggle to conduct large-scale parametric studies; emergency response teams cannot predict pollutant diffusion paths in real time. The FLICK team explores using machine learning surrogate models to replace traditional numerical simulations to solve this problem.

FLICK is a Python toolkit whose core is a trained Convolutional Neural Network (CNN) that learns the mapping from urban geometric features to wind field distribution. The input is urban geometric representations (e.g., building height maps, terrain data), and the output is the predicted wind speed field and wind direction distribution. The training data comes from high-fidelity CFD simulation results (especially datasets generated by the SOD2D solver). The model can infer in areas larger than the training domain, but when encountering geometric configurations that differ significantly from the training data, the average error is about 40%, and the team is expanding the dataset to improve robustness.

FLICK's advantage lies in speed: calculations that take traditional CFD hours to complete can be done by FLICK in seconds. This supports real-time urban planning (instant evaluation of wind environment for building schemes), large-scale parametric studies (quick scanning of design parameters), and emergency response (fast prediction of pollutant diffusion). However, FLICK is positioned as a preprocessing/quick screening tool for CFD, not a complete replacement; key decisions still require verification by traditional CFD.

The main limitation of FLICK currently is insufficient generalization ability; when encountering geometric configurations that differ significantly from the training data, the error increases significantly. Future directions include: expanding the training dataset to include more diverse urban geometries and wind conditions; integrating physics-constrained neural networks (introducing basic fluid mechanics equations into the loss function); extending to time-varying wind fields, thermal buoyancy effects, and multi-scale coupled simulations.

FLICK demonstrates a typical application paradigm of machine learning in scientific computing: using high-fidelity simulations to generate training data, training lightweight neural network surrogate models, and achieving order-of-magnitude acceleration. For practitioners in urban planning, architectural design, and climate research, FLICK provides a path for rapid evaluation and iterative design. With the expansion of datasets and model optimization, FLICK is expected to become one of the standard toolkits for urban wind environment analysis.