Predictive modelling of nanofluids thermophysical properties using machine learning

Nanofluid plays significant roles in different application areas as a result of its enhanced thermal properties. Thus, studying the thermophysical properties of nanofluids has enormous technological benefits. Traditionally, the evaluations of these properties have been undertaken by experimental approaches which can be time-consuming, laborious and costly. Consequently, many researchers have developed empirical models to predict nanofluids properties. Unfortunately, many of these models grossly underestimate or overestimate the experimental values of the thermophysical properties. Hence, there is a need to develop a better approach to overcome the stated problems with the empirical models. In this regards, there have recently been series of efforts aimed at developing machine learning (ML)-based models to address the above challenges. This thesis aimed to develop machine learning algorithms to estimate the thermophysical properties of commonly used nanofluids. The machine learning algorithms used in this thesis comprise support vector regression (SVR) and artificial neural network (ANN) developed in a MATLAB computing environment. The optimization of the machine learning parameters was conducted using the Genetic Algorithm or the Bayesian Optimization Algorithm techniques. The first part of the thesis deals with modelling and prediction of the viscosity of nanofluids while the second part deals with modelling the specific heat capacity of nanofluids. For the viscosity, a systematic study of various factors that affect the viscosity of nanofluids was conducted, the results showed that an accurate prediction of viscosity of nanofluids can be accomplished using the following input parameters; volume fraction of the nanoparticles, the fluid temperature, the size of the nanoparticles, and the density of the nanoparticles. Furthermore, the four-input BSVR model proposed in this thesis showed over 50 per cent improvement in results over the five-input ANN-based model already presented in the literature and at the same time exhibits significantly improved accuracy over the existing empirical models. For the specific heat capacity study, the following nanofluids were modelled; Al2O3-water, Al2O3-ethylene glycol (EG), CuO-water, nitrides-ethylene glycol (EG). The results of the machine learning models for each of the nanofluids were compared with simple mixing theory (model I) and thermal equilibrium based model (model II) to highlight the accuracy of the proposed techniques. For the Al2O3-water nanofluid, the model accuracy as measured by root mean square error (RMSE) obtained for the model I, model II, and the developed GA/SVR are 4.39 x 10-1 J/gK, 6.67 x 10-2 J/gK, and 1.4 x 10-3 J/gK, respectively. The GA/SVR results for Al2O3-water exhibits better accuracy than model I and Model II. In the case of Al2O3-EG nanofluids, the developed technique comprises of hybridization of Bayesian optimization algorithm with support vector regression (BSVR). The RMSE values obtained are 1.75 x 10-1 J/gK, 2.77 x 10-2 J/gK and 4.7 x 10-3 J/gK for the Model I, Model II and BSVR model, respectively. The BSVR exhibited at least an order(s) magnitude improvement for the prediction of Al2O3-EG nanofluids compared to both existing models. A similar improvement in accuracy was obtained using machine learning for the CuO-water and nitrides-ethylene glycol (EG) nanofluids. The machine models developed in this thesis are significantly better than the other existing theoretical models for all the classes of nanofluid modelled. In summary, this thesis demonstrates that machine learning-based approaches can provide more precise prediction results for specific heat capacity and viscosity of nanofluids than existing empirical/classical models. These results will be useful for experimentalists working on nanofluids design and applications.

Text FS 2021 31 - IR.pdf Download (1MB)

Actions (login required)

View Item