Multi-scale enhanced multiwavelet-based operator learning model for multiphase flow simulation

doi:10.1063/5.0257751

Multi-scale enhanced multiwavelet-based operator learning model for multiphase flow simulation

2025-03-25

0

OA

AI

DOI：10.1063/5.0257751

Original

Original request for help

Save

Abstract

En 中文

Numerical simulation of multiphase flow in porous media is critical for geoscience problems. However, numerical simulations are often computationally complex and inefficient due to multi-physics coupling and multi-scale characteristics of these problems. In recent years, several data-driven deep learning surrogate methods have been proposed to overcome these difficulties, but they still face limitations in fine-scale and global long-term evolution accuracy, data efficiency, and robustness. We propose a multi-scale enhanced multiwavelet-based model (MS-MWT) for multiphase simulation. MS-MWT learns complex dependencies across scales by training the projection kernel at multiple scales, while adding U-Net on each multiwavelet decomposition-reconstruction layer to compensate for information loss caused by decomposition, reconstruction, and Fourier truncation, and enhancing multi-scale feature extraction capabilities. This has significant advantages in improving accuracy and data efficiency for highly complex CO2-water multiphase flow problems with porosity heterogeneity, multiphase flow properties, injection configurations, and reservoir conditions. We also use a nonstandard multiwavelet representation improved by Fourier factorization to reduce model complexity and improve robustness. Compared with the best enhanced Fourier neural operator (U-FNO), MS-MWT achieves the best prediction, reducing errors by 29.03% in gas saturation prediction and by 23.53% in pressure buildup. Moreover, it demonstrates superior data utilization efficiency, achieving accuracy comparable to or better than U-FNO with only 2/3 of the data. MS-MWT also shows greater robustness on data with gradually increasing noise levels. The introduction of this model provides a more accurate and stable solution for solving complex multiphase flow problems that can be generalized to small datasets.

Keywords:

ENCODER-DECODER NETWORKS

CAPILLARY-PRESSURE

SUBSURFACE FLOW

OIL-RECOVERY

UNCERTAINTY

HETEROGENEITY

STORAGE

SCALE

Multi-scale enhanced multiwavelet-based operator learning model for multiphase flow simulation

Abstract

Journal

Organization

Cited Papers