Inline mixing efficacy evaluation using an image-based convolutional neural network combined with numerical simulation and mixture sampling

Abstract

An image-based Single-kernel Convolutional Neural Network (SK-CNN) was optimized to forecast the CFD-based inline mixing uniformity of chemical pesticides in direct injection systems (DIS), and its accuracy was validated through mixture sampling (MS) tests. Optimization results indicate that applying Xavier parameter initialization, SGDM training method, ReLu-type activation function, learning-rate decay, Moving-average, L2-regularization models, and the fully connected module of a single hidden layer with the theoretically optimal number of nodes for the SK-CNN establishment can make the verification accuracy higher, reaching 96.28%. With repeatedly captured images as test sets, the prediction results closely align those from CFD and MS, and the overall accuracies were 96.04% and 95.71% in comparison, respectively. Generalization tests presented that the model’s accuracy for typical mixers (roughly 94%) may drop slightly below the benchmark, highlighting not only its practicality, but also the necessity for further optimization by more abundant training sets and precise labels.