Machine Learning to speed up Computational Fluid Dynamics engineering simulations for built environments: A review

Abstract

Computational fluid dynamics (CFD) represents a valuable tool in the design process of built environments, enhancing the comfort, health, energy efficiency, and safety of indoor and outdoor applications. Nevertheless, the time required for CFD computations still needs to be reduced for engineering studies. Recent advances in machine learning (ML) techniques offer a promising avenue for developing fast-running data-driven models for physics-related phenomena. As scientific machine learning (SciML) research actively engages in efficiently coupling ML and CFD techniques, this literature review indicates that applications are multiplying in the built environment field to accelerate CFD simulations. This work aims to identify emerging trends and challenges in incorporating ML techniques into built environment flow simulations to foster further advancements in this domain. The prevailing approaches are direct surrogate modeling and reduced-order models (ROMs). Both approaches increasingly rely on deep learning architectures based on neural networks. The reviewed studies reported computational time gains of several orders of magnitude in specific scenarios while maintaining reasonable accuracy. However, several challenges remain to be addressed, including the models’ generalizability and interpretability, the methodology scalability, and the computational cost of deriving the models. Efforts are underway to address more complex cases with advanced SciML techniques. Among these, the incorporation of physics into the learning process and the hybridization of CFD solvers with data-driven models are worthy of further investigation. The exploration of these approaches represents a crucial step toward the deployment of reliable models that enable fast design for built environment engineering studies. • Machine learning can significantly speed up CFD simulations for the built environment. • Data-driven models need to be more generalizable, interpretable, and scalable. • Incorporating physics is essential to improve machine learning model performance. • Built environment research must adopt and foster scientific machine learning advances.

Topics & Concepts