Abstract

I will present a GENERIC-Informed Neural Network (GINN) framework to discover constitutive equations for the extra-stress in rheological models of polymer solutions. In this framework, the training of the neural network is guided by a meta-model for the conformation tensor that adheres to the GENERIC formalism (General Equation for Non-Equilibrium Reversible-Irreversible Coupling). The use of GENERIC enables a reduction of the parameter space by restricting the search to thermodynamically admissible fluid models – that is, only those that strictly satisfy the First and Second Laws of Thermodynamics. The framework operates in a diagonalized polymer conformation space, enabling generalized microstructure-based learning independent of shear, extensional, or mixed-flow kinematics data, therefore improving generalizability. Moreover, the discovery of different geometric blocks within GENERIC includes irreversible entropic contributions, anisotropic mobility terms related to the friction tensor, and specific choices of the objective derivative for the microstructural variables, helps to provide interpretability of the GINN architecture. We discuss the potential of data-driven GINN approaches to identify such models and compare various training strategies, including viscometric and complex flow data. Finally, GINN is incorporated into CFD tools for the simulation of polymeric flows using RheoTool-OpenFOAM. Fluid dynamics results will be presented to assess the accuracy of the aforementioned methodology.

Mathematical Physics and Dynamical Systems Seminar