Figure 4

The prediction results in different shear rates and bead sizes. The measurements in the gray band are used for training, and the measurements in the neighboring regions are used only for the testing. The ER equation underestimated the viscosity for most of the cases. All the other baseline machine learning methods returned flat prediction that goes through the training data points. The proposed method ERR, on the other hand, correctly reproduced the smooth and nondecreasing characteristic due to the usage of Einstein–Roscoe model as a prior knowledge.