Hybrid deep learning framework for high-precision prediction of transient thermal-flow in supercritical LNG heat exchangers

This study presents a deep learning approach for predicting eight transient thermally coupled physical quantities in supercritical LNG flowing through a teardrop-shaped printed circuit heat exchanger (PCHE) subjected to oceanic coupled motions. The quantities include the Nusselt number, Fanning friction coefficient, entropy generation rates, field synergy angles, and performance metrics. We evaluate four neural network architectures—CNN, LSTM, CNN-LSTM, and CNN-LSTM with self-attention—and demonstrate that the hybrid CNN-LSTM-Attention model significantly outperforms the others across all conditions. In particular, for a protrusion case under combined pitching and heaving motions, the model achieves an RMSE of 1.06024, about 81.5 % lower than the LSTM. The model effectively captures high-frequency oscillations and transient peaks with the highest precision in complex protrusions, and maintains the highest dynamic fidelity. The CNN-LSTM-Attention model exhibits minimal error dispersion and maximum robustness, with its error extremes strictly confined within a narrow range. By integrating local feature extraction, long-term dependency modeling, and dynamic attention weighting, this hybrid architecture offers a promising solution for high-precision prediction of complex transient thermal-flow phenomena in energy systems.