High-to-low flow dynamics learning with deep convolutional encoder–decoder networks for randomly packed pebble-bed geometry

Randomly packed pebble-bed reactors are integral components in various engineering applications, in nuclear reactors where they offer inherent safety advantages through the use of tristructural isotropic coated fuel particles embedded in a graphite matrix. Predicting coolant flow and heat transfer within these packed beds presents significant challenges due to the complex, non-uniform arrangement of pebbles, resulting in intricate flow patterns and thermal fields. High-fidelity simulations like large Eddy simulation (LES) provide detailed insight but are computationally expensive, necessitating efficient alternatives for practical applications. This study introduces a machine learning-based approach for high-to-low flow field learning using deep convolutional encoder–decoder networks applied to randomly packed pebble-bed geometry. An end-to-end field-to-field regression framework is employed, utilizing a fully convolutional encoder–decoder architecture with DenseNet feature extraction. The model is trained on velocity fields derived from both coarse and fine mesh simulations across multiple Reynolds numbers. The proposed method significantly reduces computational cost while maintaining high accuracy in predicting detailed velocity flow fields. The model's performance is validated across different Reynolds numbers and flow configurations, demonstrating a strong ability to capture dominant flow structures and localized turbulence, especially near pebble surfaces. The results confirm that this deep learning model can effectively upscale coarse mesh flow fields to high-resolution outputs, offering a promising solution for efficient and accurate simulation of packed bed reactors in thermal-hydraulic applications. Furthermore, the model's robustness is validated through tests on different pebble bed configurations, ensuring its generalizability and potential for real-world applications.