Pulsating Heat Pipe, Gradient Boosting, Machine Learning

Kurzfassung

Anthropogenic climate change increases the need for research and development aimed at reducing greenhouse gas emissions. In the transportation sector, and particularly in aviation, decarbonization requires innovative approaches to ensure the safe and efficient operation of alternative propulsion systems. Proton exchange membrane fuel cells (PEMFCs) are a promising option, but they require stringent temperature control. Pulsating heat pipes (PHPs) could represent a versatile and passive thermal management solution. Due to their high degree of design flexibility and compact construction, they are wellsuited for aerospace applications and can be integrated into existing components, such as the bipolar plates of fuel cells. Their complex two-phase flow behavior, however, including small liquid films, capillary effects, and phase change processes, makes accurate performance prediction difficult when relying solely on conventional CFD simulations, which either incur high computational cost, limiting their practical applicability, or achieve lower cost at the expense of predictive accuracy. This work introduces a machine learning (ML) framework designed to predict the thermal behavior of PHPs. This data-driven approach is benchmarked against conducted CFD simulations as an alternative approach to performance prediction and experimental results, showing that while CFD simulations can provide physical insight into PHP behavior, achieving computationally feasible simulations requires compromises in accuracy, which limits their applicability for design tasks on a large scale. By utilizing the ML approach, this study provides an efficient alternative for exploring broad PHP design spaces. Key PHP parameters, such as the heat flux, effective area, fill ratio, inclination angle, geometric features (e.g., hydraulic diameter, number of turns), and dimensionless numbers, are included as input parameters in the ML model. A total of 3,473 data points were compiled through past publications on PHP experiments. XGBoost, an algorithm based on gradient boosting, is chosen for the predictive task due to its accuracy, interpretability, and efficiency when working with complex datasets. The ML model is built to predict the key indicator of the heat transfer capabilities of PHPs, which is thermal resistance. Model optimization and performance are assessed using cross-validation, Bayesian optimization, and SHAP analysis, which quantifies the relative contributions of the input features to the model predictions, revealing the heat flux as the most impactful feature. The results indicate high predictive accuracy with an RMSPE = 20.2% and R2 = 0.941 and solid generalization across varying geometries and operating regimes. The results of the CFD simulations confirm the high computational effort required for simulating PHPs. Overall, while CFD contributes to an improved understanding of PHP behavior and the underlying challenges in its numerical simulation, its practical applicability remains constrained by limited accuracy and high computational requirements. In contrast, the ML model offers the necessary speed and reliability for design optimization, enabling rapid exploration of the design space and facilitating the design and integration of PHPs into complex aerospace thermal management systems