Graph convolution network-based surrogate model for natural convection in annuli

This work develops a model for natural convection in annuli with internal heat sources based on Graph Convolution Network (GCN), achieving rapid prediction by directly establishing the mapping relationship between the geometric features and the temperature field. The GCN learns node features and connections between nodes in data structures. Compared with traditional machine learning networks, GCN exhibits greater advantages in handling complex geometric shapes and unstructured nodes. In this study, numerous data samples were constructed with various geometric features, e.g., varying sizes, positions, and quantities of internal heat sources. The results indicate that a fully trained GCN was capable to adaptively predict temperature field distributions with arbitrary heat source sizes, positions, and quantities. The mean prediction errors of GCN for the temperature field are around 0.17 %, 0.32 %, and 0.85 %, for the single, double, and triple heat sources cases, respectively. Furthermore, the GCN prediction performance was also compared with surrogate models by the convolutional, and fully connected neural networks (CNN and FNN), which shows an improvement of 88.1 % and 84.8 %, respectively. Therefore, compared with numerical results and other surrogate models, the proposed model presents superior geometric adaptability, high prediction accuracy and a remarkable enhancement in computational efficiency (three orders of magnitude).