Tese e Dissertação

Tese: Deep Learning for Oil Production Prediction in 2D Pore-Scale Geometries and Viscosity-Variant Flows

Two-phase flow in porous media is complex, as pore geometry and fluid properties strongly influence flow patterns. Macroscopic behavior is driven by pore-scale phenomena and typically described using parameters like absolute and relative permeability or capillary pressure. Determining these from pore-scale features is difficult, requiring extensive experiments or numerical simulations. Direct pore-scale simulation is challenging and computationally expensive. This work proposes machine learning-based tools to assess two-phase macro-scale porous media flow properties as a function pore space geometry and fluid properties. Recently, deep neural networks have been explored to predict properties of single-phase flow through porous media. However, only a few works have extended these models to infer two-phase flow behavior. In a first study, a data set is created with different pore space geometry based on Voronoi diagram patterns. The pore-scale two-phase flow in all geometries is solved using the finite element method. We propose as neural network architectures a Convolutional Neural Network and a Deep Operator Network to predict the evolution of produced oil volume that results from water injection. In a second study, models based on fully connected networks using a Fourier kernel and a Laplace-Beltrami eigenfunction kernel are proposed to predict the complex pore-scale flow patterns. The two-phase flow solution is also computed using the finite element method. Results indicate that the Laplace kernel-based architecture outperforms the Fourier-based network, which proved to be less reliable. We also explore training with a physics-informed loss in addition to a purely data-driven loss. However, the inclusion of the physics-based loss did not enhance model performance. The studies conducted in this thesis propose systems that are cheaper in terms of time and computational resources than the traditional methods due to the low cost of machine learning inference.

Link da defesa:
https://puc-rio.zoom.us/j/95144718618?pwd=oolOKYzsCb90fsaaYKvqDn0fIIf5B7.1

Ver todos