An attention-based generative adversarial network architecture for heat transfer characteristics estimation in internally finned tubes

Abstract The efficient transfer of heat is critical for optimizing thermal systems used in various industries, such as power generation, chemical processing, and automotive engineering. Traditional numerical methods, while accurate, can be computationally intensive and time-consuming, posing challenges for rapid design, iterations, and scalability. By incorporating machine learning (ML), this research bridges the gap between high accuracy and reduced computational load. This study introduces a novel hybrid approach that integrates a generative adversarial network (GAN) with a dilated micro-inception unit (DMIU) and convolutional neural network (CNN) to estimate heat transfer characteristics in internally finned tubes under laminar flow conditions (50 ≤ Re ≤ 300). The proposed DMIU-CNN architecture effectively captures complex spatial and thermal patterns through its asymmetric convolution layers with varying dilation rates, enhancing feature extraction capabilities. The GAN is utilized for data augmentation, addressing the challenge of limited data availability and enhancing model generalization. This combination results in a model that reduces simulation time while maintaining high accuracy. It has been shown through numerical simulations that the GAN-DMIU model can accurately predict outlet temperatures with a maximum error of less than 0.5 K. This is 15% better than regular computational fluid dynamics (CFD) simulations. The mean absolute error (MAE) recorded was 0.991, validating the robustness and reliability of the method. The results show that combining GANs with advanced deep learning architectures can make thermal analysis faster and more accurate. This opens the door for future uses in improving heat transfer systems in many engineering fields.