Machine learning approach with Bayesian optimized neural network for non-Newtonian Eyring-Powell fluidic system in wire coating

This study introduces machine learning approach for mathematical models of wire coating analysis involving non-Newtonian Eyring-Powell fluidic system (NEPFS) by adopting the Bayesian regularization optimized neural network method (BRNNM) to increase the predictive performance of model. Two models of NEPFS namely the constant velocity model and the Reynolds Model are examined in this study by investigating the influence of Brinkman number, porous parameter, heat generation parameter, flow behavior index parameter and Eyring–Powel fluid parameters on temperature profile. Datasets for said stiff NEPFS models are generated by exploiting the strength of Adam Bashforth numerical method for each scenario to portray the evidence of heat generation parameters as an important factor that affected the uniformity and effectiveness of the coating. The adaptive parameter tuning allowed better control of the industrial processes, and less material was wasted while the quality of the products was enhanced. The Reynolds numbers were kept low to ensure fluid stability and a stable solution that can easily be applied to a real situation. It is observed that the BRNNM framework has a high level of predictive accuracy with mean squared error of 10−12to10−13, and regression coefficients close to unity (R ≈ 1) over training, testing and validation data. Increasing Brinkman number, heat generation parameter, flow behavior parameter and Eyring-Powell viscosity parameter enhanced temperature profiles, while porous parameter had minimal influence on thermal distribution. Moreover, the work demonstrates the possibility to apply non-Newtonian fluid mechanics in optimization of wire coating technologies in particular at thermal and dynamic operation. The best training performance is observed in relation to epoch number index at epoch 85, 881, 306, 806 and 1000 respectively to all the scenarios. Furthermore, observed mean squared errors (MSE) between target and output data of approximately 4.469 × 10- -12, 1.439 × 10–13, 1.391 × 10–12, 4.451 × 10–13 and 1.838 × 10–13, recorded at times of 1 s, 3 s, 2 s, 3 s and 4 s respectively. These quantitative results indicate only a very small margin of error, confirming a reliable agreement with the numerical data.