Ultra-scaled deep learning temperature reconstruction in turbulent airflow ventilation

Abstract

A deep learning super-resolution scheme is proposed to reconstruct a coarse, turbulent temperature field into a detailed, continuous field. The fluid mechanics application here refers to an airflow ventilation process in an indoor setting. Large eddy simulations are performed from a dense simulation grid and provide temperature data in two-dimensional images. The images are fed to a deep learning flow reconstruction model after being scaled down to 100 times. Training and testing are performed on these images, and the model learns to map such highly coarse fields to their high-resolution counterparts. This computational, super-resolution approach mimics the process of employing sparse sensor measurements and trying to upscale to a dense field. Notably, the model achieves high performance when the input images are scaled down by 5–20 times their original dimension, acceptable performance when 30, and poor performance at higher scales. The peak signal-to-noise ratio, the structure similarity index, and the relative error between the original and the reconstructed output are given and compared to common image processing techniques, such as linear and bicubic interpolation. The proposed super-resolution pipeline suggests a high-performance platform that calculates spatial temperature values from sparse measurements and can bypass the installation of a wide sensor array, making it a cost-effective solution for relevant applications.