Abstract
High-temperature heat pipes (HTHP) are key passive heat-removal components in heat-pipe-cooled microreactors, where their frozen startup behavior directly governs reactor safety margins. Existing analyses rely on either empirical system codes or computationally expensive CFD, making it difficult to obtain high-fidelity temperature-field predictions for startup assessment, digital-twin-assisted analysis, and rapid design iteration. This study proposes a deep learning approach for predicting temperature field evolution during HTHP frozen startup. Multiple architectures—U-Net, U-Net-LSTM, U-Net-GRU, Fourier Neural Operator (FNO), and implicit U-Net enhanced FNO (IU-FNO)—were systematically evaluated for data dependency, noise robustness, and prediction performance. Results show that IU-FNO achieves superior accuracy, with a mean absolute error (MAE) of 0.49 K and root mean squared error (RMSE) of 1.07 K, by capturing both global features and local details of temperature dynamics. Single-step inference completes within milliseconds on a workstation GPU, roughly four orders of magnitude faster than the reference CFD solver. A transfer learning strategy further enhances adaptability across HTHP configurations, reducing training time by over 90% while maintaining high accuracy (MAE 0.57 K, RMSE 0.90 K). This framework provides an efficient surrogate model for HTHP frozen-startup prediction and offers potential support for thermal-state evaluation and reactor design optimization.
| Original language | English |
|---|---|
| Article number | 131964 |
| Journal | Applied Thermal Engineering |
| Volume | 302 |
| DOIs | |
| State | Published - Aug 2026 |
Keywords
- Deep learning
- Fourier neural operator
- Frozen startup
- High-temperature heat pipes
Fingerprint
Dive into the research topics of 'Fourier neural operator-based approach for real-time prediction of high-temperature heat pipes frozen startup'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver