RheologyNet: A physics-informed neural network solution to evaluate the thixotropic properties of cementitious materials

Thixotropic behaviors can be predicted by rheological partial differential equations (PDEs) of cementitious materials. The ability to solve the rheological PDEs of viscous fluids accurately and efficiently has become an emerging interest in research. However, due to the growing number of parameters in rheological constitutive equations and the non-ideal behavior of materials from experiments, solving the rheological PDEs becomes computationally costive and error-prone. We propose a physics-informed neural network (PINN)-based framework, RheologyNet, as a surrogate solution to predict the general thixotropic behavior of cementitious materials. The complex PDEs are embedded in the well-designed RheologyNet architecture to link macroscopic viscous flow behaviors and microstructural changes. Numerical experiments suggested that RheologyNet can accurately and efficiently predict the rheological properties of cementitious materials compared to the traditional Fully-connected Neural Network (FNN) and mechanistic Finite Element Analysis (FEA). Particularly, RheologyNet demonstrated great promise for simulating history-dependent thixotropic behaviors.