Airvox: efficient computational fluid dynamics prediction using 3D convolutional neural networks for building design

Abstract

Computational Fluid Dynamics (CFD) has long been a specialized field for addressing airflow and heat transfer problems in buildings. This field requires a deep understanding of the physics involved. Recently, Artificial Neural Networks (ANNs) have been introduced to predict airflow, allowing for quick adjustments in design by evaluating potential changes in openings or mass configurations. This paper discusses using 3D Convolutional Neural Networks (3DCNNs) for building-scale simulations. We examined the performance and efficiency of standard 3DCNNs compared to residual networks and convolutional-deconvolutional networks. We focused on optimizing different 3DCNN architectures to minimize Mean Squared Error (MSE) and fine-tuned hyperparameters to ensure accurate predictions across various geometries. Notably, convolutional-deconvolutional networks with skip connections in a 3DCNN achieved an MSE of 0.0058 in validation datasets. Our findings highlight the potential of 3DCNNs to enhance CFD modeling in early building design stages, potentially reducing computational time needed for these applications.