Fluid sensing using microcantilevers: From physics-based modeling to deep learning

Abstract

In-situ measurements of the viscosity and density of small volumes of liquids are required in several industrial applications. MEMS sensors deploying vibrating microstructures constitute an attractive alternative given the significant impact of the surrounding liquid on their dynamic behavior. In this work, we combine physics-based modeling approaches and deep learning techniques to simultaneously estimate the density and viscosity of liquids from the resonance frequencies and quality factors of immersed microcantilevers. The physics-based model is first validated by comparing the simulated resonance frequencies and quality factors of immersed microcantilevers to those obtained from experiments conducted on a large variety of liquids. Then, we use the simulations results to train deep neutral networks to learn the mapping from the data space to the parameter space. The deep learning method shows high prediction accuracy provided that there is enough independent input data, shows no bias in the predicted values, and provides the results instantaneously. The optimal accuracy in the estimation of the liquid viscosity and density is achieved when the first resonance frequency and corresponding quality factor are used as inputs.