Learning macroscopic parameters in nonlinear multiscale simulations using nonlocal multicontinua upscaling techniques

Maria V. Vasilyeva, Wing Tat Leung, Eric T. Chung, Yalchin Efendiev, Mary F. Wheeler

TL;DR

This paper introduces a novel machine learning approach to accelerate complex nonlinear simulations in heterogeneous and fractured porous media. By training a deep neural network to predict upscaled transmissibilities, the method achieves significant speedups while maintaining high accuracy compared to traditional upscaling techniques.

What Problem Does This Paper Address?

Simulating fluid flow and transport in heterogeneous and fractured porous media is computationally demanding. These systems are often modeled using complex multiscale and multicontinua approaches, requiring fine-grained grids to capture small-scale heterogeneities. Traditional upscaling methods aim to reduce computational cost by creating coarser grid models, but these often struggle with nonlinear flow behaviors and the complex interactions between different continua (e.g., matrix and fractures). Existing methods for nonlinear upscaling can be computationally expensive, requiring numerous local simulations or complex iterative procedures. This paper addresses the challenge of developing a fast and accurate method for computing macroscopic parameters (upscaled transmissibilities) in these nonlinear multiscale simulations.

Methods

The core of this work is a machine learning-based approach to predict nonlinear nonlocal upscaled transmissibilities. The authors leverage a convolutional neural network (CNN) architecture. The process involves:

  1. Dataset Generation: Creating a dataset of input features (local heterogeneous properties, fracture positions, and coarse grid solution values) and corresponding output targets (upscaled nonlinear transmissibilities). This is done by solving local problems on fine grids.
  2. Neural Network Training: Training a multi-input CNN using this dataset. The network learns the complex, nonlinear relationships between local fine-scale information and the macroscopic upscaled parameters.
  3. Coarse Grid Simulation: Once trained, the neural network is used to predict the upscaled transmissibilities on-the-fly during coarse grid simulations. This replaces the computationally intensive calculation of these parameters in traditional methods.

The authors compare their machine learning approach against a classic upscaling method based on numerical homogenization.

Datasets

The paper describes the generation of datasets for training, validation, and testing of the neural networks.

  • Global Datasets: Local information is extracted from fine-grid calculations performed on the global domain.
  • Local Datasets: Datasets are generated by solving local problems up to fine-grid resolution with varying boundary conditions.
  • Snapshots: For training, multiple solution snapshots are generated by varying source terms in the global fine-grid model.
  • Test Datasets: Additional solution snapshots are generated for testing.

The datasets are divided into training, validation, and test sets, with specific ratios mentioned (e.g., 80:20 for train/validation). The input data for the neural networks consists of local heterogeneous permeabilities, local fracture position markers, and coarse grid mean values of pressure and saturation in oversampled local domains.

Key Insights and Ideas

  • Machine Learning for Upscaling: The paper demonstrates that machine learning, specifically deep CNNs, can effectively learn the complex, nonlinear relationships required for accurate upscaling of transmissibilities in heterogeneous and fractured porous media.
  • Nonlocal Nature: The approach explicitly handles the "nonlocal" aspect of upscaling, meaning that the upscaled parameter at a given coarse grid face depends on information from a wider local domain, not just immediate neighbors.
  • Efficiency Gain: By replacing expensive local simulations with fast neural network predictions, the method offers a significant computational speedup for coarse grid simulations.
  • Accuracy Preservation: Despite the speedup, the machine learning approach maintains high accuracy, often outperforming traditional upscaling methods, especially for nonlinear flow problems.
  • Adaptability: The method is shown to be applicable to both unsaturated (single-phase) flow and two-phase flow and transport problems.

Key Results

  • Accuracy: The proposed machine learning method achieves significantly lower relative L2 errors for pressure compared to classic upscaling. For Test 1 (nonlinear flow), errors were around 1-2% for the ML method versus ~15% for classic upscaling. For Test 2 (nonlinear flow and transport), errors were even lower (e.g., 0.920% for pressure, 3.957% for saturation).
  • Speedup: The online stage (neural network inference) is substantially faster. For Test 1, the coarse grid calculation with ML was approximately 35 times faster than fine grid calculations. For Test 2, the ML approach achieved a speedup of roughly 48 times compared to fine grid simulations.
  • Training Time: While the online prediction is fast, the offline training of the neural networks takes time, ranging from minutes to hours depending on the network complexity and GPU used (e.g., 25-28 minutes for Test 1 networks, 2-80 minutes for Test 2 networks).
  • Dataset Impact: The study suggests that the accuracy of the ML method depends on the quality and coverage of the training dataset, highlighting the importance of sufficient snapshots.

Limitations and Future Work

  • Dataset Construction Time: While the ML prediction is fast, the initial construction of the training dataset can be time-consuming, especially if it involves extensive fine-grid simulations.
  • Generalizability: The paper focuses on specific types of heterogeneous and fractured media. Further research could explore the method's applicability to a wider range of complex geological formations.
  • Fracture-Fracture Transmissibilities: For Test 1, the authors noted that fracture-fracture transmissibilities (Tff,NL) were nearly constant and thus set equal to their upscaled counterparts (Tff,UP), suggesting that dedicated ML models for all transmissibility types might not always be necessary. Future work could investigate this further.
  • Parameter Sensitivity: The paper mentions using threshold values (ε, εs) to handle small pressure differences and ensure positivity of transmissibilities. Further analysis of the sensitivity to these parameters could be beneficial.
  • Real-time Adaptation: While the method is fast for prediction, exploring more adaptive or online learning strategies during the simulation itself could be a future direction.

Summary updated

Abstract

In this work, we present a novel nonlocal nonlinear coarse grid approximation using a machine learning algorithm. We consider unsaturated and two-phase flow problems in heterogeneous and fractured porous media, where mathematical models are formulated as general multicontinuum models. We construct a fine grid approximation using the finite volume method and embedded discrete fracture model. Macroscopic models for these complex nonlinear systems require nonlocal multicontinua approaches, which are developed in earlier works. These rigorous techniques require complex local computations, which involve solving local problems in oversampled regions subject to constraints. The solutions of these local problems can be replaced by solving original problem on a coarse (oversampled) region for many input parameters (boundary and source terms) and computing effective properties derived by nonlinear nonlocal multicontinua approaches. The effective properties depend on many variables (oversampled region and the number of continua), thus their calculations require some type of machine learning techniques. In this paper, our contribution is two fold. First, we present macroscopic models and discuss how to effectively compute macroscopic parameters using deep learning algorithms. The proposed method can be regarded as local machine learning and complements our earlier approaches on global machine learning. We consider a coarse grid approximation using two upscaling techniques with single phase upscaled transmissibilities and nonlocal nonlinear upscaled transmissibilities using a machine learning algorithm. We present results for two model problems in heterogeneous and fractured porous media and show that the presented method is highly accurate and provides fast coarse grid calculations.
BibTeX
@article{vasilyeva2020learning,
  author = {Maria V. Vasilyeva and Wing Tat Leung and Eric T. Chung and Yalchin Efendiev and Mary F. Wheeler},
  title = {{Learning macroscopic parameters in nonlinear multiscale simulations using nonlocal multicontinua upscaling techniques}},
  journal = {Journal of Computational Physics},
  year = {2020},
  doi = {10.1016/j.jcp.2020.109323},
}
Figure