Data-Driven Inference of Fluid Properties Using Structural Response Signature

Abstract

The non-invasive and accurate estimation of fluid properties using minimal sample volumes holds significant importance across industrial, biomedical, and scientific applications. This study presents a novel approach for predicting fluid properties by analysing the dynamic vibrational response of a cantilever beam interacting with liquids. The methodology employs advanced signal processing techniques, including non-dimensional feature engineering, to extract meaningful patterns from frequency response data. Two machine learning approaches - polynomial regression and the Random Convolutional Kernel Transform (ROCKET) - are utilized to model the relationship between vibrational characteristics and fluid parameters. Furthermore, symbolic regression is applied to derive a mathematical expression linking the system's damping ratio to the fluid's viscosity and density. The developed models demonstrate excellent predictive performance, as validated through the test liquid dataset. Beyond immediate applications in fluid characterization, this work provides a proof of concept to the broader field of lab-on-a-chip technologies by offering a compact, efficient, and high-precision diagnostic platform. The proposed framework combines the advantages of mechanical sensing with data-driven modelling, paving the way for real-time fluid analysis in resource-constrained environments. © Published under licence by IOP Publishing Ltd.