Paper Number
AM2 

Session
AI and ML Based Rheological Characterization

Title
From meta-modeling of complex fluids using physics-based machine learning to digital rheometer twins

Presentation Date and Time
October 10, 2022 (Monday) 10:10

Track / Room
Track 6 / Mayfair

Authors (Click on name to view author profile)

1. Mahmoudabadbozchelou, Mohammadamin (Northeastern University)
2. Kamani, Krutarth M. (University of Illinois at Urbana-Champaign, Department of Chemical and Biomolecular Engineering)
3. Rogers, Simon A. (University of Illinois at Urbana-Champaign, Department of Chemical and Biomolecular Engineering)
4. Jamali, Safa (Northeastern University)

Author and Affiliation Lines (in printed abstract book)
Mohammadamin Mahmoudabadbozchelou¹, Krutarth M. Kamani², Simon A. Rogers² and Safa Jamali^1
1Northeastern University, Boston, MA; ²Department of Chemical and Biomolecular Engineering, University of Illinois at Urbana-Champaign, Urbana, IL 61801

Speaker / Presenter 
Mahmoudabadbozchelou, Mohammadamin

Keywords
AI based; ML based

Text of Abstract
Accurate property predictions of complex fluids and soft materials are of interest across many disciplines, particularly in designing processes incorporating such materials. In this work, we present a hybrid approach built upon our earlier work on Rheology-informed Neural Networks (RhiNNs), to model, describe, and predict the behavior of complex fluids. We leverage advances in data-driven constitutive meta-modeling through Multi-Fidelity Neural Network (MFNN), as well as accurate solutions of differential equations that are commonly used in constitutive modeling of complex fluids. We then introduce Rheology-Informed Graph Neural Networks (RhIGNets), which are based on the graph mode of Tensor Flow, to learn the hidden rheology of complex constitutive models. In practice, identifying model parameters for a multi-variant thixotropic or viscoelastic constitutive model generally necessitates a long list of experimental testing. RhIGNets are found to be capable of learning these non-trivial model parameters for a complex material using only a handful of data points from a single flow procedure, allowing for accurate modeling with a small number of experiments. A multi-fidelity approach is then taken to combine limited additional experimental data with the RhiGNet predictions to develop “digital rheometers” that can be used in place of a physical instrument. Such digital rheometers are benefiting from a hybrid mode of physical inclusion, meaning the explicit and implicit inclusion simultaneously, allowing robust and precise predictions of a complex fluid to different flow protocols with unprecedented accuracy, beyond what has been made possible with any constitutive model regardless of their foundation. We finally discuss the outlook and opportunities, and more importantly the limitations of science-based ML platforms in rheology and non-Newtonian fluid mechanics.