Nuclear Engineering ETDs
Accurate prediction of CHF is still a challenging issue in the study of boiling heat transfer. Many factors contribute to the occurrence of CHF and the various trigger mechanisms are proposed to unravel physical phenomena behind CHF. However, those mechanisms cannot cover the multiple primary factors simultaneously and even some of them still remain controversially unresolved. In light of the complexity and difficulty of CHF modelling, hereby an ensemble-learning based framework is proposed to model and predict CHF based on the databank of CHF. Some prior trials have been done for three primary aspects of dominant factors, that is, surface morphology, geometrical dimension and operation condition. These three primary constituents are respectively analyzed though three different sub-models of the ensemble framework in Chapter 3, 4 and 5.
In Chapter Three, relevant experiments about micro-pillar enhanced CHF are reviewed and the corresponding databank of microstructure enhanced CHF is compiled based on those CHF experiments from published papers. Although the impacts of micro-pillars on CHF are still not clear, through qualitative analyses, the parametrical trends of CHF with respect to geometrical parameters of pillar array can be roughly foreseen. Meanwhile, this study also evaluates performance of prediction accuracy among four current physical models of microstructure-enhanced CHF. Comparative results show that two capillary wicking models have higher prediction accuracy. Particularly, a special terminology, zero-infinity convergence, is introduced to discuss the parametrical trends of CHF and qualitatively assess veracity of two capillary wicking models. Given the drawbacks of current physical models, the DBN is proposed to more accurately predict CHF and study parametric trends of CHF based on the microstructure enhanced CHF databank. Different from the training process of other regression modelling problems, constrained CHF points, which are artificially derived from the training data datasets, are required to be coupled with the raw training datasets for achieving the zero-infinity convergence of the DBN based CHF model, exhibiting accurate parametric trends of CHF and improving the prediction accuracy. This new training technique provides a new reliable solution to the similar constrained machine learning problems. Numerical results demonstrate that DBN can achieve the best performance of CHF prediction in terms of prediction accuracy. Through studying parametrical trends of CHF reveals that micro pillar arrays with the same parameters on heat transfer substrates with different dimensional sizes presents different CHF enhancement profiles. The presented methodology provides new insights for CHF modelling in pool boiling enhanced by other surface modification techniques, including porous layer coating, nanoparticle deposition, textured roughen, and nanowire fabrication.
The effects of dimensions and materials of boiling surfaces on CHF are correlated and studied through the GRNN modelling in Chapter Four. Instead of inputting all parameters that indicate the thermal properties of materials into the trained model, the aggregated parameters from the primitive parameters of thermal properties, thermal activity and thermal diffusivity, are utilized as the input parameters of the trained model. This technique not only could capture the effects of thermal properties of materials on CHF effectively but also helps reduce the computational loads. The trained model shows the similar parametric trends of CHF to that of the traditional empirical correlation with respect to the thermal activity. If the thermal activity of heat transfer substrate is beyond a certain value, the corresponding effect of thermal activity will be absent, which somehow implies that the thickness of heat transfer substrate will not impact CHF after the asymptomatic thickness is reached. On the other hand, thermal diffusivity still affects CHF occurrence even if the effect of thermal activity is negligible. When coming to the effect of dimension size on CHF, it was found that when the side length of square heat transfer substrate is 5 times greater than the capillary length of working fluid, the CHF will be independent on the side length. Otherwise, CHF will be affected by the side length, and the influence of side length on CHF reaches ultimate if the side length of square boiling surface is exactly equal to the Raleigh-Plateau instability wavelength. This instability wavelength is only dependent on the thermal properties of working fluids, meaning that the optimal side length for CHF optimization is only related to the thermal properties of working fluid, namely, the surface tension, and the liquid and vapor densities of working fluid.
In Chapter Five of this study, n-support vector machine is adopted to explore and study experimental strategies for the data-driven approaches of CHF look-up table construction, on the basis of sparingly-distributed experimental CHF data points. In the virtue of the CHF look-up table of Groeneveld et al (2007), those CHF data was used as the reference data of this research. In this data collection, CHF data of the subcooled flow boiling (Xe < 0) is chosen to concentrate on the PWR steady-state condition because the in the normal operation of PWR, the system is under the subcooled flow boiling. The numerical results have demonstrated that ν-SVM trained by well sparsely-distributed training data in the parameter region of interest (pressure and mass flux) can yield a fairly acceptable degree of CHF prediction accuracy. Procuring training data points that can imply the parametric behaviors of CHF with respect to pressure and mass flux for support vector machine is the essential key of machine learning to achieving a high level of CHF prediction accuracy. For capturing the pressure-variant CHF behavior, training data that are in the proximity of the CHF inflection point significantly contribute to the improvement of prediction accuracy. Hence, training data preparation physics-informed with knowledge of CHF inflection points definitely augments the prediction accuracy of CHF. How the parametrical trends of CHF with respect to pressure and mass flux are close to the linear trends determines the level of prediction accuracy when lacking of a good spread of training data points. Besides, it is found that CHF extrapolation to a higher pressure with many data points collected at different low pressures can be effectively achieved by SVM if a few CHF data points are available under the high pressure, especially for PWR pressure of 15.5 MPa. This announces a possibility of strategic integration experiments between high pressure and low pressure, reducing experimental costs associated with the high pressure testing in terms of efforts and money. The proposed methodologies provides engineers and experimentalists with useful strategies to construct the look-up table tabulation of advanced cladding materials of ATFs.
It is found out that there are multiple sub-problems that could be divided for CHF prediction and each sub-problem has its individual suitable machine learning model. Those prior work done by this study proves that the data-driven CHF modelling by sub-models can provide accurate CHF prediction under various scenarios and correct parametrical trends with respect to separate variables.
Last but not least, another contribution of this thesis to the field of boiling heat transfer is that two databanks of experimental CHF data are compiled for the CHF enhancement by microstructures. The compiled databanks provide useful information and guidelines to the future design of surface structures that will possibly be applied to heat exchanger and nuclear fuel rod.
boiling heat transfer CHF Machine learning
Level of Degree
First Committee Member (Chair)
Second Committee Member
Third Committee Member
He, Mingfu Mr. "APPLICATION OF MACHINE LEARNING TO CHF MODELLING." (2019). https://digitalrepository.unm.edu/ne_etds/83