Research on stirring process based on artificial neural network and multi-phase flow simulation technology

In response to the strong nonlinearity inherent in the stirred tank system, traditional research methods often struggle to accurately and rapidly infer the actual on-site conditions. To address this issue, this study proposed a combined approach of Artificial Neural Network (ANN) and Computational Fluid Dynamics (CFD) to construct an ANN-CFD turbulent state prediction model. Three training algorithms (Levenberg-Marquardt, Bayesian Regulation, and Scaled Conjugate) and optimization algorithms (genetic algorithm, particle swarm algorithm, and simulated annealing algorithm) were employed to investigate the model’s hyperparameters. A dual-layer genetic algorithm (GA-GA) was utilized to optimize the number of hidden layer nodes and initial weight thresholds in the neural network, resulting in an optimal neural network architecture. The ANN-CFD model was then used to predict the flow field status inside the stirred tank and then evaluate its accuracy. The results demonstrated that the BR algorithm attained high prediction accuracy when the number of neurons exceeded 9, with overall stability in accuracy variations. The genetic algorithm exhibited outstanding performance in terms of global convergence and prediction accuracy within this model. The combination of 11-10 hidden layer neurons achieved the best overall performance under the condition of a dual hidden layer. The optimized ANN-CFD based on the GA-GA model demonstrated excellent prediction accuracy, with all regression metrics surpassing 0.9. Compared to the traditional backpropagation neural network, the proposed model achieved more than a twofold improvement in fitting performance on the validation and test sets.