Data-driven selection of constitutive models via rheology-informed neural networks (RhINNs)

A myriad of empirical and phenomenological constitutive models that describe different observed rheologies of complex fluids have been developed over many decades. With each of these constitutive models' strength in recovering different rheological responses, algorithms that allow the data to automatically select the appropriate constitutive relations are of great interest to rheologists. Here, we present a rheology-informed neural network (RhINN) that enables robust model selection based on available experimental data with minimal user intervention. We train our RhINN on a series of experimental data for different complex fluids and show that it is capable of finding the appropriate model with the lowest number of fitting parameters for each data set. Finally, we show that uniform selection of a handful of data over the entire accessible shear rates does not affect the RhINN's accuracy, while providing a specific range of data (and omitting the rest) results in an erroneous model determination.