CFD and Machine Learning for performance analysis in 3D printed heat exchangers

In many industrial contexts, such as in the automotive, energetic, aeronautic and chemical sector, thermal management plays a fundamental role. As an example, in the automotive field the electrification is leading to a major requirement in terms of thermal management of the electronic components, like batteries or electronic converters. This leads to a necessary improvement of heat exchangers technology. In recent years, additive manufacturing (AM) has started to provide new design possibilities. Additive manufacturing allows indeed the production of more complicated parts than those with traditional production systems, for example, micro-lattice structures. The 3D printed lattice structures allow a reduction of material utilised, a lower product density and higher surface-volume ratios. These characteristics make lattice structures intriguing for heat exchangers development. Furthermore, in the industrial field, simulation is becoming more and more important, as it can partially substitute the experimental stages on early prototypes. In particular, Computational Fluid Dynamics (CFD) is pivotal in fluid mechanics studies. Therefore, in the first part of this thesis, CFD simulations were performed on seven different lattice structures, analysing their performances and comparing with those of traditional setup of heat exchangers. More precisely, the convective heat transfer coefficient and the pressure drop of each geometry are evaluated. These two parameters must indeed be considered together as, for performance enhancement, higher heat transfer coefficients should be obtained at no significant expense in terms of pressure drop increase. If, on the one hand, CFD allows lower costs for research activities, due to lower prototypes production and testing, on the other hand, the simulations may be time consuming and require high computational cost. This is the reason why Machine Learning (ML) may be adopted to reduce simulations. Machine learning is a branch of Artificial Intelligence, with the potential to reduce CFD simulations. In fact, ML algorithms can be used to recognise data patterns from previously obtained simulation data, and to extrapolate to different cases. This finally helps to obtain results in a faster way with respect to CFD simulations only. In the second part of this work, genetic programming, which is a branch of Machine Learning, is used to analyse simulated data and to obtain correlations between relevant variables, in order to predict performance variations as a function of geometric parameters or flow conditions.