Comparative analysis of ensemble, supervised, and deep learning regression algorithms for parametric modelling of solid-liquid fluidization

Background: A comparative regression modelling of fluidization bed data parameters is performed in this work using different algorithms. Computational fluid dynamics (CFD) modelling of particle and fluid flow characters using two-fluid Eulerian-Eulerian model. RNG k-ε turbulence coupled with kinetic theory of granular flow was also combined. The developed numerical model is used for generating the fluidization related data of parameters like turbulent viscosity, turbulent dissipation rate, solid velocity, solid volume fraction, granular temperature, and turbulent kinetic energy. Methods: Comparative modelling and performance analysis between ensemble learning, supervised learning, and neural networks is performed for the mentioned fluidized bed parameters. Ensemble Regression algorithms: Gradient boosting regressor (GBR), Voting regressor (VR), and Random-forest regressor (RFR), supervised learning algorithm - Decision tree (DT), and Deep Artificial neural network (ANN) models are used for the data mapping of fluidization parameters. Performance metrices are accessed in details to compare the modelling results or the algorithms in details for each fluidization parameter. Findings: From the modelling of this data it is found that numerical data is highly non-linear. DT and RFR algorithms are the most accurate algorithms that predicted with >90 % of accuracy in each case. VT and GBR trained and tested with around 85 % accuracy in most cases but failed in prediction of granular temperature. ANN also sufficiently provided good accuracy while it also failed to predict granular temperature. Solid volume fraction, turbulent kinetic energy, turbulent viscosity, and turbulent dissipation rate were modelled perfectly with all the algorithms. Among all the parameters, turbulent viscosity during training and testing from each model is highly accurately modelled from each of the algorithm with prediction accuracy >90 %.