Fast evaluation of pressure and saturation predictions with a deep learning surrogate flow model

Maldonado-Cruz, Eduardo; Pyrcz, Michael J.

doi:10.1016/j.petrol.2022.110244

Cited by 11 publications

(2 citation statements)

References 44 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, the spatial correlation between pixels within the image signal is ignored when using MSE alone as a fidelity measure. To more accurately measure the fidelity of the surrogate model, we introduced SSIM [15] with reference to previous studies [5,16,17]. The SSIM proposed by Wang et al [15,18] forms a local SSIM by comparing local image blocks at the same location of two image signals.…”

Section: Variable Loss Functionmentioning

confidence: 99%

Predicting CO2 storage pressure and saturation based on deep learning surrogate model

Jianchun¹,

Xu²,

Fu³

2023

Preprint

View full text Add to dashboard Cite

CO2 sequestration which is one of the important means to avoid global warming today provides an economically viable technical means to reduce greenhouse gas emissions. To maximize the sequestration efficiency while evaluating the effect of formation uncertainty, a numerical simulator of multiphase flow is required to simulate highdimensional nonlinear multiphase flow within a nonhomogeneous porous medium. Due to the inherent inhomogeneity of the formation of porous media and the nonlinear coupling of multiple complex physical processes, a significant amount of repetitive numerical simulation processes impose considerable computational costs and require prolonged computational time to obtain simulation results.In this paper, we propose an efficient and fast flow surrogate modeling process for deep learning, proposing that the extended hyperparameter optimization process will incorporate the neural network architecture and the loss function as relevant parameters into the optimization process. Subsequently, we conducted experiments based on the workflow proposed in this study for the case of CO 2 storage in a homogeneous deep saltwater layer and achieved accurate predictions at 120-time steps with mean MSE of 5E-5 and 2E-5 for gas saturation and pressure, and MSSIMs of 0.9989 and 0.9998, respectively, under different production parameters and well placement settings.

show abstract

Section: Variable Loss Functionmentioning

confidence: 99%

Predicting CO2 storage pressure and saturation based on deep learning surrogate model

Jianchun¹,

Xu²,

Fu³

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…A number of studies have applied image-based approaches and snapshots of simulation data over a spatially discretized input domain for surrogate modeling of subsurface flow and transport problems. Most of these works leverage convolutional neural networks (CNNs) to learn the nonlinear mappings from the input properties (e.g., permeability) to the output states (pressure and saturation) on regular Cartesian meshes (Mo et al 2019;Tang et al 2020;Wang and Lin 2020;Wen et al 2021;Zhang et al 2021;Jiang et al 2021;Yan et al 2022;Maldonado-Cruz and Pyrcz 2022). While CNNs are powerful in approximating PDE solutions, they are restricted to a specific discretization of the physical domain in which they are trained.…”

Section: Introductionmentioning

confidence: 99%