Fluid rheology prediction using interface detection and machine learning regression

This paper aims to introduce a prediction method for fluid viscosity based on fluid–air interface features in the pendant drop configuration. An automatic algorithm for interface detection, including an image processing pipeline is proposed, followed by a comprehensive study regarding the fitting method of the extracted pixels with a polynomial function. We propose interface detection based on k-means clustering algorithm followed by artifact removal and polynomial fitting and parametrization using curvature radii, minimum and maximum diameter of the droplet, droplet configuration, liquid type (viscous or viscoelastic), density, and surface tension. The radii of curvature were used to classify the liquids in viscous/viscoelastic, while all eight features were used to train a regression algorithm to quantitatively and qualitatively characterize weakly viscoelastic polymeric liquids. Viscous and viscoelastic liquids were involved in the training process, namely water–glycerin solutions and water–polyethylene oxide solutions of various concentrations. Algebraic as well as Chebyshev polynomials were comparatively implemented to provide the most accurate continuous and differentiable function of the interface.