Integration of neutron radiography and CNN for hidden performance behavior analysis in dual-evaporator loop heat pipes

• CNN classification reached 97.8% accuracy in identifying steady-state phase. • Lowest RMSE of 0.0026 achieved at 10° tilt in vapor fraction prediction. • Neutron radiography revealed hidden liquid carryover and vapor backflow. • DE-LHP tilt angle enhanced condensation efficiency and CNN model accuracy. Emissions from gasoline-powered vehicles remain a major challenge to achieving carbon neutrality, accelerating the transition toward electric vehicles (EVs) powered by lithium-ion batteries. However, the thermal sensitivity of these batteries demands efficient passive cooling systems to maintain safety and performance. This study investigates the use of a Dual-Evaporator Loop Heat Pipe (DE-LHP) as a passive thermal management solution, focusing on the visualization of vapor fraction distribution using neutron radiography. To enhance image interpretation, Convolutional Neural Networks (CNNs) were applied for both classification and regression tasks using image datasets at four DE-LHP tilt angles (0°, 5°, 7.5°, and 10°), each containing 300 frames. Logarithmic intensity-based analysis revealed that the average vapor fraction decreased with increasing tilt angle, with the 10° configuration achieving the highest condensation efficiency. The CNN classification model effectively distinguished startup and steady-state phases, reaching 97.8% accuracy for steady-state and 76.7% for startup. The regression model achieved a minimum RMSE of 0.00326 at 10°, with all RMSE values lying within their 95% confidence intervals, confirming statistical reliability. Both models were trained and tested using a 70%–30% data split, identified as the most effective configuration among seven scenarios. Additionally, transient behaviors such as liquid carryover and vapor backflow were detected, highlighting the ability of neutron radiography to capture dynamic two-phase phenomena. The results confirm that tilt angle significantly affects both thermal performance and model accuracy. Overall, integrating neutron radiography with CNN-based analysis provides a robust framework for intelligent monitoring and optimization of two-phase thermal management in DE-LHP systems.