Inferring viscoplastic models from velocity fields: a physics-informed neural network approach

TL;DR

This paper introduces a physics-informed neural network framework that infers complex viscoplastic rheological laws directly from velocity field data, overcoming traditional measurement limitations and enabling model-agnostic rheology identification.

Contribution

The work presents a novel PINN-based method for directly learning rheological laws from flow data, including strategies for model selection and handling noisy measurements.

Findings

Successfully infers rheological parameters from synthetic data

Demonstrates robustness to noise in velocity measurements

Highlights importance of shear rate distribution for inference accuracy

Abstract

Fluid-like materials are ubiquitous, spanning from living biological tissues to geological formations, and across scales ranging from micrometers to kilometers. Inferring their rheological properties remains a major challenge, particularly when traditional rheometry fails to capture their complex, three-dimensional, and often heterogeneous behavior. This difficulty is exacerbated by system size, boundary conditions, and other material-specific physical, chemical, or thermal constraints. In this work, we explore whether rheological laws can be inferred directly from flow observations. We propose a physics-informed neural network (PINN) framework designed to learn constitutive viscoplastic laws from velocity field data alone. Our method uses a neural network to interpolate the velocity field, enabling the computation of velocity gradients via automatic differentiation. These gradients are used to estimate the residuals of the governing conservation laws, which implicitly depend on the unknown rheology. We jointly optimize both the constitutive model and the velocity field representation by minimizing the physical residuals and discrepancies from observed data. We validate our approach on synthetic velocity fields generated from numerical simulations using Herschel-Bulkley, Carreau and Panastasiou models under various flow conditions. The algorithm reliably infers rheological parameters, even in the presence of significant noise. We analyze the dependence of inference performance on flow geometry and sampling, highlighting the importance of shear rate distribution in the dataset. Finally, we explore preliminary strategies for model-agnostic inference via embedded model selection, demonstrating the potential of PINNs for identifying the most suitable rheological law from candidate models.