Deep learning model for rapid temperature map prediction in transient convection process using conditional generative adversarial networks

This article presents a deep learning model for three-dimensional (3D) transient mixed convection in a horizontal channel with a heated bottom surface. Using Conditional Generative Adversarial Networks (cGAN), we successfully approximate temperature maps at arbitrary channel locations and time steps. The model is specifically designed for mixed convection at Reynolds number 100, Rayleigh number 3.9 × 10⁶, and Richardson number 88.8. To investigate the impact of the discriminator network architecture on model accuracy, we compare Convolutional Neural Network (CNN) based classifiers with PatchGAN classifiers, both with and without strided convolutions. Remarkably, the cGAN with PatchGAN based classifier (without strided convolutions) yields the highest clarity and accuracy in inferring temperature maps. Additionally, other factors such as image contrast, spatiotemporal variation rate of temperature, and the number of channels in the temperature image significantly influence cGAN accuracy. This work highlights the potential of deep learning in efficiently modeling complex transport processes.