Aplicação de redes neurais artificiais no método de multi forçagem direta

Abstract

The combination of Computational Fluid Dynamics (CFD) methods with Artificial Intelligence (AI) tools has shown great potential to overcome computational limitations and provide more robust solutions for the study of particulate materials at the micro and nanoscale. In this work, we propose the application of Physics-Informed Neural Networks (PINNs) within the Immersed Boundary Method (IBM) to simulate Taylor-Couette flow. The main objective is to replace the traditional IBM distribution function with a machine learning-based model, improving the transfer of information between Eulerian and Lagrangian meshes while reducing computational cost. The methodology employs a PINN trained to predict the weights of the distribution function based on parameters such as the dimensionless distance between the Lagrangian and Eulerian meshes and the velocity field, eliminating the need to impose the weight function directly in the code. The simulation of Taylor-Couette flow (Ta = 100) using the distribution function generated by the neural network yielded results consistent with the traditionally used model, including the formation of Taylor vortices, and demonstrated a significant reduction in computational cost during the simulation. As a contribution, this work evaluates the feasibility of applying neural networks to model the distribution function in the immersed boundary method.