ViscNet: Vision-Based In-line Viscometry for Fluid Mixing Process Journal: Engineering Applications of Artificial Intelligence

Jongwon Sohn, Juhyeon Moon, Hyunjoon Jung, Jaewook Nam · 2025

Read ViscNet: Vision-Based In-line Viscometry for Fluid Mixing Process Journal: Engineering Applications of Artificial Intelligence by Jongwon Sohn, Juhyeon Moon, Hyunjoon Jung, Jaewook Nam (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-01268.txt.

Introduction

Jongwon Sohn, Juhyeon Moon, Hyunjoon Jung, and Jaewook Nam present ViscNet in a 2025 engineering paper. Viscosity, how thick a fluid is, must be measured for process monitoring and for autonomous laboratories. Conventional viscometers are invasive: they must touch the fluid and need controlled lab conditions unlike real process settings.

ViscNet is a vision-based viscometer that works in line, inside the process. It watches a fixed background pattern through the fluid's surface. Mixing deforms that surface, and refraction distorts the pattern in ways tied to viscosity. Reading thickness from a camera could let mixing processes self-monitor continuously without probes.

Catalog record

CREATORJongwon Sohn, Juhyeon Moon, Hyunjoon Jung, Jaewook Nam
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-01268

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ViscNet: Vision-Based In-line Viscometry for Fluid Mixing Process Journal: Engineering Applications of Artificial Intelligence Jongwon Sohn, Juhyeon Moon, Hyunjoon Jung, Jaewook Nam Abstract Viscosity measurement is essential for process monitoring and autonomous laboratory operation, yet conventional viscometers remain invasive and require controlled laboratory environments that differ substantially from real process conditions. We present a computer-vision-based viscometer that infers viscosity by exploiting how a fixed background pattern becomes optically distorted as light refracts through the mixing-driven, continuously deforming free surface. Under diverse lighting conditions, the system achieves a mean absolute error of 0.113 in log ⁡ m 2 ​ s − 1 \log\mathrm{m}^{2}\mathrm{s}^{-1} units for regression and reaches up to 81% accuracy in viscosity-class prediction. Although performance declines for classes with closely clustered viscosity values, a multi-pattern strategy improves robustness by providing enriched visual cues. To ensure sensor reliability, we incorporate uncertainty quantification, enabling viscosity predictions with confidence estimates. This stand-off viscometer offers a practical, automation-ready alternative to existing viscometry methods. 1 Introduction Viscosity is a fundamental property of fluids that quantifies the internal friction generated within the fluid flow. Because it strongly influences the characteristics of fluid motion, viscosity is measured across a broad range of laboratory and industrial settings in food processing [ 25 ] , coating operations [ 13 ] , pharmaceuticals manufacturing [ 38 , 42 ] , and petrochemical production [ 40 ] . In automated systems, viscosity functions not only as a diagnostic indicator of process conditions [ 10 , 34 ] but also as a key parameter guiding how fluids should be manipulated and controlled [ 39 ] . Therefore, the ability to measure viscosity rapidly, reliably, and with minimal operational burden has become increasingly critical. Modern viscometry is dominated by mechanical instruments that induce relative motion between a sensor component and the fluid, infering viscosity by measuring the resulting frictional force, velocity change, or other dynamic response of the internal component. Classical examples include falling-ball, capillary, and spindle viscometers [ 7 ] . However, these instruments require highly idealized, laboratory-specific environments that differ substantially from real process conditions. Hence, fluid samples must be extracted and transported to the device, a procedure that can be time-consuming and may compromise the sterility or purity of the original content [ 26 ] . In addition, sampling errors may arise from spatial inhomogeneity within the fluid volume or from changes that occur during the sampling interval [ 14 ] . To overcome these issues, in-line viscometers have been developed to be installed directly in pipelines and to measure viscosity during operations. While these instruments avoid the need for off-line sampling, they still require direct contact with the fluid, which presents several challenges. Sensor lifetime can be compromised by aggressive flow conditions, elevated temperature, or high viscosity [ 33 ] , and the measurement inherently reflects only a partial local state rather than the true bulk property. In addition, installation, removal, and periodic recalibration often require significant effort, further limiting their practicality [ 11 ] . These challenges highlight the need for stand-off viscometers that can infer viscosity without any direct interaction with the fluid. A promising alternative is to use computer-vision techniques. Recent studies have explored the idea that viscosity is reflected in the motion of the free surface of the fluid. It has been shown from a biological perspective that the shape of a fluid surface contains visually perceptible cues about its viscosity [ 37 ] . Various vision-based viscometry methods building on this insight have emerged, including the use of robotic interaction to generate surface motion in a bottle and estimate viscosity [ 27 , 39 ] , or using top-view videos of sloshing fluid in the cup to compute optical flow and infer viscosity [ 1 ] . A key challenge shared across these works is the need to effectively visualize the dynamics of the fluid surface. To obtain sufficiently clear observations, these methods rely on controlled or specialized experimental conditions. Consequently, although they offer stand-off measurements, they remain difficult to deploy in practical in-situ settings. In this work, we address the limitations of previous viscometers by extending computer-vision-based viscometry to fluid mixing processes. Mixing is a widely used operation in fluid-related processes [ 4 , 18 ] , and its free surface naturally exhibits rich fluctuation and decay behaviors that make it well suited for optical analysis [ 12 ] . These surface dynamics provide a physically meaningful and easily observable signal from which viscosity can be inferred without intrusive instrumentation. Motivated by these observations, we present ViscNet, a computer-vision-based, in-situ, stand-off viscometer designed specifically for mixing environments. The main contributions of this work are summarized as follows: 1. We demonstrate that bulk viscosity can be inferred by analyzing how viscous dissipation modulates the surface geometry of the stirring-induced flow. 2. We show that a background pattern acts as an optical probe whose refraction-induced distortion through the fluid surface encodes viscosity-dependent dynamics. 3. We constructed a large-scale video dataset of stirred fluids, combining real experimental recordings with simulation data generated via particle-based fluid simulations and photorealistic rendering. 4. We developed a modified Video Vision Transformer (ViViT) architecture that predicts viscosity directly from videos of vortex surface with impeller rotation speed embedded as an additional conditioning signal. The contents of the paper are organized as follows. Section 2 reviews previous work on stirring-flow physics and transparent free-surface visualization. Section 3 presents our methodology, including the model architecture, dataset construction, and training schemes. Section 4 reports viscosity estimation results, analyzes model behavior, and outlines additional experiments for improvement. Section 5 concludes with a summary of our findings, discusses the significance and limitations of our model, and highlights directions for future work. 2 Related Works 2.1 Viscous Effects on Free Surface of Mixing Vortex The critical influence of viscosity on the shape and motion of stirring vortex surfaces has been demonstrated across theoretical models, experimental investigations, and simulation studies. This influence is particularly evident during impeller-driven mixing, where the internal flow field separates into an inner forced-vortex region governed directly by the impeller and an outer free-vortex region shaped primarily by the container walls [ 12 ] . In the outer region, wall friction generates small turbulent eddies, producing corrugated surfaces whose size and morphology are governed primarily by viscous forces. Also, mixing vortices exhibit polygonal or multi-lobed surface motions whose size, shape, and rotation rate vary with viscosity [ 3 , 17 , 36 ] . Furthermore, when external forcing of rotation is removed, the dissipation of vortex rotation and free-surface curvature is likewise viscosity dependent [ 24 , 31 ] . 2.2 Visualization of Transparent Free Surface Visualizing transparent fluids has long been a challenging task due to their background-dependent appearance. Prior works primarily rely on segmentation-based approaches, using motion cues, appearance cues, or alternative imaging to isolate transparent liquids across various setups [ 9 , 22 , 30 , 41 ] . Our method takes more suitable and versatile approach by adopting the background-oriented schlieren (BOS) imaging principle, where the interface geometry of a transparent medium is inferred from optical distortion of a background pattern [ 28 ] . Mathematical models and, more recently, neural network methods have further enabled dynamic free-surface reconstruction

Read the complete text