Machine-learning-based measurement of fluid properties through observation of particle train evolution in microchannels

The characterization of complex liquids is of great importance in many applications. Among the properties that can be measured, the relaxation time has great relevance, as it provides a measure of fluid elasticity. We propose a novel method to estimate the relaxation time of viscoelastic fluids by applying machine learning to microfluidics. Specifically, we train a Long-Short Term Memory (LSTM) neural network to identify the Weissenberg number (a dimensionless parameter connected to liquid relaxation time) that characterizes the dynamics of trains of rigid particles suspended in a viscoelastic liquid flowing in a cylindrical microchannel. We first study the effect of the Weissenberg number on the evolution of the microstructure through numerical simulations. An in silico dataset consisting of the distributions of the interparticle distances at different channel sections is built and used to train the network. The performance of the LSTM model is tested on both classification and regression problems. The proposed method is non-intrusive and requires a simple setup.