Paper Number
ML3                         My Program 

Session
AI and ML in Rheology

Title
Rheo-Former: A generative platform for reliable rheological and non-Newtonian fluid mechanical modeling

Presentation Date and Time
October 22, 2025 (Wednesday) 10:30

Track / Room
Track 6 / Sweeney Ballroom C

Authors (Click on name to view author profile)

1. Saberi, Maedeh (Northeastern University, Mechanical and Industrial Engineering)
2. Jamali, Safa (Northeastern University, Mechanical and Industrial Engineering)

Author and Affiliation Lines (in printed abstract book)
Maedeh Saberi and Safa Jamali
Mechanical and Industrial Engineering, Northeastern University, Boston, MA 02115

Speaker / Presenter 
Saberi, Maedeh

Keywords
artificial intelligence; machine learning; non-Newtonian fluids

Text of Abstract
Rigorous and accurate prediction of complex fluids’ flow behavior in different scenarios has been a central objective of rheological sciences. Rheologically-relevant fluids often exhibit strong dependence on the history of flow, anisotropic stress responses, and nonlinearities that are nontrivial to capture using conventional numerical solvers, and thus non-Newtonian fluid mechanical platforms have been constantly evolving to alleviate such challenges. With unprecedented progress in development of data-driven techniques, reliable ML-based platforms that are capable of accurately modeling non-Newtonian fluids are generally lacking, or suffer from limited generalizability and computational inefficiencies when applied across varying flow conditions and geometries.
In this work, we develop a novel modeling platform based on neural operators—specifically, a Transformer-based architecture called OFormer—designed to learn the full operator mapping from system inputs (e.g., shear rate or velocity fields) to stress responses in complex fluids. This method integrates attention-based mechanisms with a latent time-marching propagator, enabling efficient and accurate prediction of both spatial interactions and temporal evolution within a reduced latent space. We evaluate our OFormer on multiple constitutive models (Thixotropic Elasto-Viscoplastic, Giesekus, and Oldroyd-B) and assess its performance on two categories of problems: (1) rheometric experiments involving prediction of scalar and tensorial stress responses from applied shear rates, and (2) full-field flow simulations in complex domains such as contraction channels and flow past obstacles. OFormer results suggest that it can serve as a robust surrogate for high-fidelity simulations, accurately resolving the entire flow of different complex fluids using minimal training data.