Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Pore-scale flow prediction using physics informed neural networks
Submission Author:
Pedro Calderano , MG
Co-Authors:
Pedro Calderano, Marcio CARVALHO, Helon Vicente Hultmann Ayala
Presenter: Pedro Calderano
doi://10.26678/ABCM.COBEM2023.COB2023-1690
Abstract
Numerical simulations are employed to study the behavior of different dynamical systems. They are usually based on the solution of a set of differential equations that models the system. Although numerical simulation may provide detailed information about the system of interest, the traditional approach consumes large amounts of computational resources. Surrogate models are a manner to circumvent the referred time-consuming characteristic and provide quick response solutions. Physical Informed Neural Networks (PINNs) is a numerical method recently developed that can attach information from a system of partial differential equations in the loss function of a neural network. As trained neural networks provide an almost instantaneous answer and PINN formulation includes information from well-established equations, PINNs can be used as accurate surrogate models. In the present work, we use PINNs to predict steady-state velocity field in the pore space of three different porous media geometry configurations. The proposed method solves the flow that results from a constant pressure differential applied to the porous media. First, we implement the basic PINN formulation to predict pressure distribution and the velocity of the flow passing through a couple of simple geometries. After that, we implement a PINN to a more complex geometry, which requires a Fourier transformation in its input space to provide an appropriate solution. The results show that PINNs can quickly predict the discussed flow cases with small errors when compared to the flow solutions obtained by the Finite Element Method. We also observe that the PINNs inference stage consumes less computational resources when compared to the direct numerical simulation.
Keywords
machine learning, Physics-informed neural networks, porous medium
