Prediction of bubble departing diameter in pool boiling of mixtures by ANN using modified ReLU

In this research, bubble departure diameter in pool boiling have been measured in aqueous amine and ethylene glycol solutions for various concentrations. The experimental data have been compared with major existing predictive correlations. It is shown that the effect and identity of the independent variables on bubble diameter proposed in the previous studies are inconsistent. The predictions of different correlations have on average a deviation of about 40% from the experimental data. This is mainly due to the complicated interactions between bubbles on the heterogeneous boiling medium, which provides a complex condition. This complexity limits any mathematical modelling of the forces acting on the developing bubbles. Particularly in liquid solutions, where mass transfer by back diffusion through micro-sub-layers adds further complexity. In this work, the classical artificial neural network, ANN, with rectified linear unit, ReLU, activating function, AF, has been modified. This modification is based on adding a numerical matrix to each layer to adjust the slope of AF for each neuron independently. The addition of this parameter, together with the adjustment of the bias matrix, makes the activation function more flexible than the classical ReLU. To find the tuning parameters, a genetic algorithm was implemented instead of the back-propagation technique. It is shown that the predictions of the trained ANN with modified ReLU AF agree within an absolute average error of 10%, which is equal to the total uncertainty of the measurements. Prediction of bubble departing diameter in boiling phenomena is a key parameter for accurate design, operation and optimisation in many industrial systems.