ViscNet: Vision-Based In-line Viscometry for Fluid Mixing Process

TL;DR

ViscNet introduces a stand-off, vision-based viscometer that infers bulk viscosity from how a fixed background pattern is refractively distorted by a stirring-induced free surface. The approach uses a ViViT-inspired model with impeller speed conditioning and a three-component Gaussian Mixture Model to provide calibrated viscosity estimates, trained on a large real dataset augmented with extensive synthetic data. Key results show accurate regression in $\log m^2 s^{-1}$ space ($\text{MAE} \approx 0.113$) and up to 81% viscosity-class accuracy, with uncertainty quantified and calibrated to reflect predictive confidence. The work demonstrates data-efficient strategies via synthetic data pretraining and multi-pattern inputs, highlighting practical potential for autonomous, in-line mixing monitoring.

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 m2 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.

Figures (10)

• Figure 1: ViscNet Model Architecture, Regression units are in log $\mathrm{m}^2 \mathrm{s}^{-1}$.
• Figure 2: Experiment setup and Dataset Diversity Structure.
• Figure 3: (a) Confusion matrix of the classification model trained with all 4 patterns, (b) Class accuracy plotted for each single pattern models.
• Figure 4: t-SNE plots of (a) the all-pattern model and (b) the noise-scale-3 model, which achieved the highest performance.
• Figure 5: Spatial attention graphs of the encoder’s final layer, binned into five groups by (a) Reynolds number, (b) Capillary number, and (c) Weber number.