Paper Number
ML19                         My Program 

Session
AI and ML in Rheology

Title
High-throughput viscometry via machine-learning using videos of inverted vials: The effects of process parameters on the inference accuracy

Presentation Date and Time
October 23, 2025 (Thursday) 9:45

Track / Room
Track 6 / Sweeney Ballroom C

Authors (Click on name to view author profile)

  1. Arretche, Ignacio (University of Illinois Urbana-Champaign, Beckman Institute for Advanced Science and Technology)
  2. Hossain, Mohammad Tanver (University Of Illinois Urbana-champaign, Mechanical Science and Engineering)
  3. Tiwari, Ramdas (University of Illinois Urbana-Champaign, Mechanical Science and Engineering)
  4. Mills, Mya G. (University of Illinois Urbana-Champaign, Department of Chemistry)
  5. Kim, Abbie J. (University of Illinois Urbana-Champaign, Department of Chemistry)
  6. Armstrong, Connor D. (University of Illinois Urbana-Champaign, Beckman Institute for Advanced Science and Technology)
  7. Lessard, Jacob J. (University of Utah, Department of Chemistry)
  8. Tawfick, Sameh H. (University Of Illinois Urbana-champaign, Mechanical Science and Engineering)
  9. Ewoldt, Randy H. (University Of Illinois Urbana-champaign, Mechanical Science and Engineering)

Author and Affiliation Lines (in printed abstract book)
Ignacio Arretche1, Mohammad Tanver Hossain2, Ramdas Tiwari2, Mya G. Mills3, Abbie J. Kim3, Connor D. Armstrong1, Jacob J. Lessard4, Sameh H. Tawfick2 and Randy H. Ewoldt2
1Beckman Institute for Advanced Science and Technology, University of Illinois Urbana-Champaign, Urbana, IL 61801; 2Mechanical Science and Engineering, University Of Illinois Urbana-champaign, Urbana, IL 61801; 3Department of Chemistry, University of Illinois Urbana-Champaign, Urbana, IL 61801; 4Department of Chemistry, University of Utah, Salt Lake City, UT 84112

Speaker / Presenter 
Arretche, Ignacio

Keywords
experimental methods; computational methods; artificial intelligence; methods; machine learning; networks; rheometry

Text of Abstract
Here, we explore the complex uncontrolled flow of the inverted vial—driven by gravity, surface tension, inertia, and initial conditions—to achieve scalable, simultaneous viscosity inference. To do this, we introduce the computer vision (CV) viscometer, a system that automates the test by inverting multiple vials at once and recording their flow with a simple camera. Given the flow's complexity, and the lack of approximate solutions, we approximate its inverse function for inference with a neural network. Rather than relying on velocity fields, we train the network to learn the inverse function directly from raw video footage, using the visual appearance of the flow as input. Surprisingly, from raw videos alone, we quantitatively infer the viscosity of Newtonian fluids across nearly five orders of magnitude (0.01–1000 Pa·s), achieving relative errors below 25%—improving to 15% above 0.1 Pa·s. By varying inversion speed and observation time, we delineate how inertial and capillary effects define the method’s accuracy limits. We show how high Deborah numbers and dimensionless stress amplitudes can also increase inference variability, outlining the method's practical boundaries when it comes to non-Newtonian behavior. Instead of depending on carefully controlled flows like traditional methods, we show that complex, uncontrolled flows can be used to infer viscosity. The CV viscometer represents this new approach, providing a simple, low-cost way to collect viscosity data at scale. It can easily fit into both automated and manual lab workflows, helping speed up material research and support robotic laboratories.