MACHINE LEARNING - ENHANCED MASS FLOW MEASUREMENTS USING A CORIOLIS MASS FLOW SENSOR

Aug 3, 2026

Uploaded by Eliana

Machine learning-enhanced mass flow measurements are presented for microfabricated Coriolis mass flow sensors with integrated pressure and temperature sensing. Linear regression and support vector regression are trained on four features extracted from raw sensor signals to improve mass flow estimation. Using information from all integrated sensors, the proposed regression-based approach increases full-scale accuracy by a factor of 4 for trained fluids, compared with conventional mass flow detection that relies on fewer signals.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

Abstract

Introduction

Sensor Theory and Design

Regression Based Approach

MACHINE LEARNING-ENHANCED MASS FLOW MEASUREMENTS USING A CORIOLIS MASS FLOW SENSOR R. Zubavicius, D. Alveringh1 , M. Poel2 , J. Groenesteijn1,3 , R. G. P. Sanders1 , R. J. Wiegerink1 , J. C. Lo¨tters1,3 1 MESA+ Institute, University of Twente, Enschede, The Netherlands 2 Data Management and Biometrics, University of Twente, Enschede, Netherlands 3 Bronkhorst High-Tech BV, Ruurlo, The Netherlands ABSTRACT We present machine learning-enhanced mass ﬂow measurements based on microfabricated Coriolis mass ﬂow sensors with integrated pressure and temperature sensors. Two machine learning techniques have been applied: linear regression (LR) and support vector regression (SVR) on four features extracted from the raw sensor data to improve mass ﬂow estimation. In contrast to conventional mass ﬂow detection, LR and SVR use information from all integrated sensors to estimate the mass ﬂow, which results in a full-scale accuracy improvement of a factor 4 for trained ﬂuids. KEYWORDS Coriolis, mass ﬂow, machine learning, linear regression, support vector regression, sensors, microﬂuidics. INTRODUCTION Miniaturization of ﬂow sensors using microfabrication techniques has enabled the measurement and control of very low ﬂows, i.e., in the order of a few micro or nano liters per minute [1] . Accurate ﬂow sensing and control has applications in pharmaceutical, biomedical and many other ﬁelds [2, 3, 4] . Thermal ﬂow sensors are capable of measuring down to a few nano liters per minute. However, thermal ﬂow sensing principles are highly dependent on liquid properties; the sensors need to be calibrated per liquid [5] . In contrast to thermal ﬂow sensors, Coriolis mass ﬂow sensors (CMFS) are capable of measuring the true mass ﬂow without the need for calibration per ﬂuid. Last decades, a few different fabrication techniques have been developed to miniaturise these type of ﬂow sensors [6, 7, 8] . The latter has been used to integrate multiple ﬂuid sensors into a single chip [9] and to ﬁnd accuracy limits for CMFS [10, 11] . Recent work has focused on applying machine learning for ﬂuid classiﬁcation [12] and extraction of ﬂuid parameters [13] . In the work presented in this paper, the improvement of the accuracy for mass ﬂow sensing by applying machine learning to the raw sensor signals has been investigated. Figure 1: Illustration of the proposed principle using a CMFS, temperature and pressure sensors, with (a) the chip,(b) the swing and twist mode of the CMFS,(c) the pressure sensor, (d) different ﬂuids, (e) ﬁve output signals, (f) preprocessing and (g) the regression-based machine learning. Adapted from [12]. SENSOR THEORY AND DESIGN The sensor chip consists of a CMFS, two pressure sensors (upstream and downstream with respect to the CMFS) and a temperature sensor, see Figure 1a. The 2 0 2 4 The 5th Conference on MicroFluidic Handling Systems , 21–23 February 2024 , Munich , Germany 65 CMFS utilizes the Coriolis principle, where inertia caused by a ﬂuid ﬂow ('m) in a twisting microchannel induces a swing mode (Figure 1b) . The swing mode is proportional to the mass ﬂow and is theoretically independent of ﬂuid properties. However, temperature and pressure inﬂuence the movement of the channel [11, 14], which includes the magnitude of the swing mode. Therefore use of the signals from the already integrated temperature and pressure sensors potentially improves the accuracy. The CMFS has two capacitive readout structureson each side with respect to the twist mode axis. The phase shift at the fundamental frequency between these electrodes is proportional (for low ﬂows) to the mass ﬂow [8] . The pressure sensors consist of a membrane that deforms proportional to the pressure in the channel. To measure the pressure, there are golden resistive tracks deposited on top of the channel and connected in Wheatstone bridge conﬁguration. On top of the CMFS, there are also a resistive track deposited to enable temperature

FAQs

Which machine learning methods are used to enhance mass flow estimation?

The study applies linear regression (LR) and support vector regression (SVR) to improve mass flow estimation from extracted sensor features.

What sensor signals and features are used as inputs for the regression models?

Four features are selected: phase shift Δ', resonance frequency f0, temperature-dependent track resistance R, and upstream pressure P1, derived from raw CMFS signals.

How much accuracy improvement is reported compared with conventional mass flow detection?

For trained fluids, the regression approach yields a full-scale accuracy improvement by a factor of 4.

Aug 3, 2026

Uploaded by Eliana