This study presents a deep learning framework designed to accelerate Large Eddy Simulation (LES) of airflow in urban environments. The framework leverages physics constrained conditional Generative Adversarial Networks (GANs) trained on instantaneous velocity snapshots from a synthetically generated dataset comprising 130 high-fidelity CFD simulations of simple building configurations. By learning the mapping from early-stage flow fields to their statistically steady-state counterparts, the framework allows the simulation to bypass the lengthy transient averaging phase and predict the final time-averaged fields directly. Two GAN-based architectures are explored: a conventional convolutional model operating on structured uniform grids (Grid-GAN), and a graph-based model (Graph-GAN) that utilizes Graph Neural Networks (GNNs), specifically Graph Attention Networks (GATs), to process unstructured CFD mesh data while preserving native spatial connectivity. Both approaches are integrated into a fully automated pipeline built exclusively on open-source tools, including OpenFOAM for CFD simulations, FreeCAD and ParaView for preprocessing, and PyTorch for deep learning model development and training. Results demonstrate that the proposed models can significantly reduce LES computational costs while retaining accuracy in predicting turbulent flow characteristics. The Graph-GAN, in particular, shows enhanced adaptability and physical consistency due to its ability to exploit mesh refinements in critical regions. This work lays the foundation for the development of robust, physics-informed surrogate models and supports the growing integration of deep learning with scientific simulations in fluid mechanics.

This thesis explores the intersection of deep learning (DL) and computational fluid dynamics (CFD) to improve the modeling and analysis of built environmentaerodynamics. As urbanization accelerates and sustainability challenges intensify, accurate and efficient tools for airflow prediction in cities and buildings are increasingly vital. Traditional CFD methods, while powerful, are computationally demanding and limited by model assumptions, especially in turbulence modeling. This work investigates whether deep learning techniques can enhance both the speed and generalizability of aerodynamic simulations, and whether they can support experimental measurements such as those obtained from wind tunnels. The thesis presents a comprehensive framework that integrates convolutional neural networks (CNNs), graph neural networks (GNNs), and generative adversarial networks (GANs) to accelerate large eddy simulations (LES), reconstruct flow fields, and improve experimental data processing. Notably, GNN-based models are used to operate directly on unstructured CFD meshes, preserving geometric and topological information critical for urban flow predictions. Hybrid approaches that combine physics-based knowledge with data-driven models are also introduced. Applications span both simulated and experimental domains, including a case study on wind tunnel shape optimization using reinforcement learning. While deep learning models showed strong potential for improving both simulation accuracy and speed, the work also highlights important challenges, including the need for better generalization, model interpretability, and the lack of publicly available CFD datasets.The findings suggest that combining deep learning with traditional fluid dynamics offers a promising path forward, especially when supported by open data, physical constraints, and collaborative research efforts. The thesis concludes by outlining directions for future research in physics-informed learning, dataset curation, and real-time integration of predictive models into sustainable urban design.

Denna avhandling utforskar skärningspunkten mellan djupinlärning (DL) och beräkningsströmningsdynamik (CFD) med avsikten att förbättra modellering och analys av aerodynamiken i byggda miljöer. I takt med att urbaniseringen accelererar och hållbarhetsutmaningarna blir större, blir ökar behovet av noggranna och effektiva verktyg för luftflödesprognoser i städer och byggnader. Traditionella CFD-metoder, även om de är kraftfulla, är beräkningskrävande och begränsade av modellantaganden, särskilt inom turbulensmodellering. Detta arbete undersöker om djupinlärningstekniker kan användas för att förbättra både hastigheten och generaliserbarheten hos aerodynamiska simuleringar, och om de kan stödja experimentella mätningar som erhålls från vindtunnlar. Avhandlingen presenterar ett omfattande ramverk som integrerar faltningsneurala nätverk (CNN), grafneurala nätverk (GNN) och generativa adversariella nätverk (GAN) för att accelerera simuleringar av stora virvelsimuleringar (LES), rekonstruera flödesfält och förbättra experimentell databehandling. GNN-baserade modeller används särskilt för att arbeta direkt på ostrukturerade CFD-nät, vilket bevarar geometrisk och topologisk information som är avgörande för förutsägelser av luftflöden i stadsmiljöer. Hybridmetoder som kombinerar fysikbaserad kunskap med datadrivna modeller introduceras också. Tillämpningarna spänner över både simulerade och experimentella områden, inklusive en fallstudie om optimering av en vindtunnels form med hjälp av förstärkningsinlärning (reinforcemnt learning). Djupinlärningsmodeller visade stark potential för att förbättra både simuleringsnoggrannhet och hastighet. Samtidigt belyser arbetet också viktiga utmaningar, såsom behovet av bättre generalisering, modelltolkningsbarhet och även bristen på tillgängliga CFD-data. Resultaten tyder på att kombinationen av maskininlärning med traditionell fluiddynamik erbjuder en lovande väg framåt, särskilt när det finns tillgång till av öppna data, behov med fysiska begränsningar och även gemensamma forskningsinsatser. Avhandlingen avslutas med att skissera riktningar för framtida forskning inom fysikinformerat lärande (PINNs), datasetkurering (dataset curation) och realtidsintegration av prediktiva modeller för en hållbar stadsplanering.

This study aims to improve the accuracy and speed of predictions for thermal comfort and air quality in built environments by creating a coupled framework between computational fluid dynamics (CFD) simulations and deep learning models. The coupling approach is showcased by the development of a data-driven turbulence model. The new turbulence model is built using a deep learning neural network, whose mapping structure is based on a zero-equation turbulence model for built environment simulations, and is coupled with the CFD software OpenFOAM to create a hybrid framework. The neural network is a standard shallow multi-layer perceptron. The number of hidden layers and nodes per layer was optimized using Bayesan optimization algorithm. The framework is trained on an indoor environment case study, as well as tested on an indoor office simulation and an outdoor building array simulation. Results show that the deep learning based turbulence model is more robust and faster than traditional two-equation Reynolds average Navier-Stokes (RANS) turbulence models, while maintaining a similar level of accuracy. The model also outperforms the standard algebraic zero-equation model due to its superior ability to generalize to various flow scenarios. Despite some challenges, namely the mapping constraint, the limited training dataset size and the source of generation of training data, the hybrid framework demonstrates the viability of the coupling technique and serves as a starting point for future development of more reliable and advanced models.

This study develops a deep learning framework to speed up Large Eddy Simulation (LES) of airflow in built environments. The framework employs a convolutional neural network trained on instantaneous velocity snapshots of computational domains for one indoor environment case study. It accelerates LES simulations by reaching statistically steady state faster, resulting in significant speedups in computational cost and simulation run-time while maintaining similar accuracy levels to standard LES simulations. This approach can greatly reduce the computational effort required for LES simulations in large scale environments. Together with promising results, limitations in the current framework exist. The main one is the failure to extrapolate properly to different flow fields than the training set and understand the 3D turbulence structures. The current framework poses a base for the development of more trusty and robust accelerator models in the future.

The study and control of the airflow in the indoor environment is crucial. Computational fluid dynamics (CFD) allows detailed flow analysis, but faces challenges mainly in terms of computational efficiency. Modern data driven tools have lately received much attention for inexpensive airflow simulations. In this work we develop a coupled CFD–deep learning framework by employing a neural network to replace a classical zero-equation eddy viscosity turbulence model. A standard multi-layer perceptron is trained on TensorFlow and subsequently transferred and applied to a CFD solver in OpenFOAM. Training and testing were performed by collecting data from validated Reynolds-averaged Navier–Stokes (RANS) simulations of indoor airflow from literature. The new coupled framework enables accurate results with significantly faster prediction speed. A primary challenge remains in being able to train the neural network over a larger dataset and provide higher generalizability to the model with sufficient accuracy.

Fast and accurate airflow simulations in the built environment are critical to provide acceptable thermal comfort and air quality to the occupants. Computational Fluid Dynamics (CFD) offers detailed analysis on airflow motion, heat transfer, and contaminant transport in indoor environment, as well as wind flow and pollution dispersion around buildings in urban environments. However, CFD still faces many challenges mainly in terms of computational expensiveness and accuracy. With the increasing availability of large amount of data, data driven models are starting to be investigated to either replace, improve, or aid CFD simulations. More specifically, the abilities of deep learning and Artificial Neural Networks (ANN) as universal non-linear approximator, handling of high dimensionality fields, and computational inexpensiveness are very appealing. In built environment research, deep learning applications to airflow simulations shows the ANN as surrogate, replacement for expensive CFD analysis. Surrogate modeling enables fast or even real-time predictions, but usually at a cost of a degraded accuracy. The objective of this work is to critically review deep learning interactions with fluid mechanics simulations in general, to propose and inform about different techniques other than surrogate modeling for built environment applications. The literature review shows that ANNs can enhance the turbulence model in various way for coupled CFD simulations of higher accuracy, improve the efficiency of Proper Orthogonal Decomposition (POD) methods, leverage crucial physical properties and information with physics informed deep learning modeling, and even unlock new advanced methods for flow analysis such as super-resolution techniques. These promising methods are largely yet to be explored in the built environment scene. Unavoidably, deep learning models also presents challenges such as the availability of consistent large flow databases, the extrapolation task problem, and over-fitting, etc.

This study presents a deep learning framework designed to accelerate LargeEddy Simulation (LES) of airflow in urban environments. The framework leverages physics constrained conditional Generative Adversarial Networks (GANs) trained on instantaneous velocity snapshots from a synthetically generated dataset comprising 130 high-fidelity CFD simulations of simple building configurations. By learning the mapping from early-stage flow fields to their statistically steady-state counterparts, the framework allows the simulation to bypass the lengthy transient averaging phase and predict the final time-averaged fields directly. Two GAN-based architectures are explored: a conventional convolutional model operating on structured uniform grids (Grid-GAN), and a graph-based model (Graph-GAN) that utilizes Graph Neural Networks (GNNs), specifically Graph Attention Networks (GATs), to process unstructured CFD mesh data while preserving native spatial connectivity. Both approaches are integrated into a fully automated pipeline built exclusively on open-source tools, including OpenFOAM for CFD simulations, FreeCAD and ParaView for preprocessing, and PyTorch for deep learning model development and training. Results demonstrate thatthe proposed models can significantly reduce LES computational costs while retaining accuracy in predicting turbulent flow characteristics. The Graph-GAN, in particular, shows enhanced adaptability and physical consistency due to its ability to exploit mesh refinements in critical regions. This worklays the foundation for the development of robust, physics-informed surrogate models and supports the growing integration of deep learning with scientific simulations in fluid mechanics.

The growing capabilities of deep learning and artificial intelligence (AI) have revolutionized numerous fields, including indoor airflow simulations. Computational Fluid Dynamics (CFD) plays a crucial role in simulating airflow, heat transfer, and pollutant dispersion, which are essential for healthier indoor climates. However, the progress in this domain is hampered by the lack of publicly available CFD datasets that are consistent and standardized. With the urgent need to solve the recent COVID-19 crisis there has been a new wave of published studies to inspect airflow patterns in confined spaces. By reviewing many recent CFD applications in indoor environment, this study shows a perspective on the critical importance of establishing a robust, open-access repository of CFD data for indoor airflow simulations. Such a resource would not only accelerate research by providing a rich dataset fortraining AI models but also ensure that findings are replicable and comparable across different studies. This paper advocates for larger efforts to create accessibleand standardized CFD datasets, which will be instrumental in leveraging AI for thedevelopment of smarter, healthier, and more sustainable cities.