UC Berkeley

Viscosity and density are fundamental physical properties governing liquid behaviors. Traditionally, these parameters are tested by bulky, sample-dependent, and time-consuming benchtop equipment while miniaturized and in-situ monitoring systems have gained strong interests. The characteristics of ultrasonic waves traveling through the liquid medium are influenced by both viscosity and density but there are significant challenges by using physical models to decouple and detect density and viscosity. This work uses piezoelectric micromachined ultrasonic transducers (PMUTs) with the assistance of machine learning to simultaneously sense viscosity and density toward practical applications such as the detection of the degradation of engine oils. A PMUT array operating in the pulse-echo mode is utilized to collect acoustic signals, from which ten signal features are extracted. These data are fed into a neural network classification model for liquid classification and multivariable neural network regression models for sensing viscosity and density. Two experimental setups have been conducted. In the first configuration, the PMUTs sensor is immersed in the testing fluid such that the ultrasonic waves travel through the test fluid. Experimental results show a 99.95% classification accuracy for three distinct liquids (DI water, ethylene glycol, and a glycerol–water mixture), and the mean absolute errors (MAEs) for the ethylene glycol–water mixtures were 1.6 ± 1.57% for viscosity and 0.20 ± 0.17% for density. The second setup extends the platform to a non-contact based measurement scheme where the sensor operates entirely outside the testing liquid container with two approaches. The first utilizes the pulse-echo from the test liquid-air interface as signals propagates through the test fluid and reflects to the sensor. Results show a 98% classification accuracy across four liquids of DI water, ethylene glycol, mineral oil, and propylene glycol. Measurements also demonstrate MAEs of 0.79 cP and 12.6 kg·m⁻³ for propylene glycol–water mixtures. The second approach use the pulse-echo from the container wall–test liquid interface and the ultrasonic waves don’t travel through the test liquid. This method achieves an 87% classification accuracy across the tested four liquids with MAEs of 1.88 cP and 13.9 kg·m⁻³ for propylene glycol–water mixtures. Overall, this new platform of AI-assisted simultaneous characterization of fluid viscosity and density via PMUTs offers a practical alternative to address the challenges in monitoring physical properties of liquids.