Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validation

Recent advancements in deep learning have significantly improved solving complex computational physics problems. This paper presents Physics-Informed Neural Networks (PINNs) with a spatially-dependent diffusion function to model two-phase flow in porous media, explicitly addressing the Buckley-Lever...

Full description

Saved in:
Bibliographic Details
Main Authors: Daulet Kalesh, Timur Merembayev, Sagyn Omirbekov, Yerlan Amanbek
Format: Article
Language:English
Published: Elsevier 2025-06-01
Series:Results in Engineering
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2590123025015099
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850233490108317696
author Daulet Kalesh
Timur Merembayev
Sagyn Omirbekov
Yerlan Amanbek
author_facet Daulet Kalesh
Timur Merembayev
Sagyn Omirbekov
Yerlan Amanbek
author_sort Daulet Kalesh
collection DOAJ
description Recent advancements in deep learning have significantly improved solving complex computational physics problems. This paper presents Physics-Informed Neural Networks (PINNs) with a spatially-dependent diffusion function to model two-phase flow in porous media, explicitly addressing the Buckley-Leverett problem. The study analyzes the PINNs using different approaches, including diffusion term and modification flux function to convex hull, employing both Multi-Layer Perceptron (MLP) and Attention-based neural network architectures. In addition, we investigate the application of the PINNs for laboratory experimental data by assessing the performance of PINNs in capturing the saturation front dynamics. Our results indicate that while the attention-based model achieves slightly higher accuracy, it is significantly time-consuming. In contrast, the MLP is the more efficient in terms of training time with a speedup in the range of 7−13. A sensitivity analysis on the constant diffusion coefficient shows that PINNs can approximately capture the pattern of the saturation front. We have found that adapting the spatial-dependent diffusion function provided better accuracy with the experimental data compared to using a constant diffusion coefficient.
format Article
id doaj-art-a080f987f1754df2931a40dc5563c31e
institution OA Journals
issn 2590-1230
language English
publishDate 2025-06-01
publisher Elsevier
record_format Article
series Results in Engineering
spelling doaj-art-a080f987f1754df2931a40dc5563c31e2025-08-20T02:02:55ZengElsevierResults in Engineering2590-12302025-06-012610543910.1016/j.rineng.2025.105439Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validationDaulet Kalesh0Timur Merembayev1Sagyn Omirbekov2Yerlan Amanbek3Department of Mathematics, Nazarbayev University, Kabanbay batyr 53, Astana, 010000, KazakhstanDepartment of Mathematics, Nazarbayev University, Kabanbay batyr 53, Astana, 010000, KazakhstanNational Laboratory Astana, Nazarbayev University, Kabanbay batyr 53, Astana, 010000, KazakhstanDepartment of Mathematics, Nazarbayev University, Kabanbay batyr 53, Astana, 010000, Kazakhstan; Corresponding author.Recent advancements in deep learning have significantly improved solving complex computational physics problems. This paper presents Physics-Informed Neural Networks (PINNs) with a spatially-dependent diffusion function to model two-phase flow in porous media, explicitly addressing the Buckley-Leverett problem. The study analyzes the PINNs using different approaches, including diffusion term and modification flux function to convex hull, employing both Multi-Layer Perceptron (MLP) and Attention-based neural network architectures. In addition, we investigate the application of the PINNs for laboratory experimental data by assessing the performance of PINNs in capturing the saturation front dynamics. Our results indicate that while the attention-based model achieves slightly higher accuracy, it is significantly time-consuming. In contrast, the MLP is the more efficient in terms of training time with a speedup in the range of 7−13. A sensitivity analysis on the constant diffusion coefficient shows that PINNs can approximately capture the pattern of the saturation front. We have found that adapting the spatial-dependent diffusion function provided better accuracy with the experimental data compared to using a constant diffusion coefficient.http://www.sciencedirect.com/science/article/pii/S2590123025015099PINNsTwo-phase flowBuckley-LeverettMachine learningMLPAttention-based
spellingShingle Daulet Kalesh
Timur Merembayev
Sagyn Omirbekov
Yerlan Amanbek
Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validation
Results in Engineering
PINNs
Two-phase flow
Buckley-Leverett
Machine learning
MLP
Attention-based
title Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validation
title_full Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validation
title_fullStr Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validation
title_full_unstemmed Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validation
title_short Application of physics-informed neural networks for two-phase flow model with variable diffusion and experimental validation
title_sort application of physics informed neural networks for two phase flow model with variable diffusion and experimental validation
topic PINNs
Two-phase flow
Buckley-Leverett
Machine learning
MLP
Attention-based
url http://www.sciencedirect.com/science/article/pii/S2590123025015099
work_keys_str_mv AT dauletkalesh applicationofphysicsinformedneuralnetworksfortwophaseflowmodelwithvariablediffusionandexperimentalvalidation
AT timurmerembayev applicationofphysicsinformedneuralnetworksfortwophaseflowmodelwithvariablediffusionandexperimentalvalidation
AT sagynomirbekov applicationofphysicsinformedneuralnetworksfortwophaseflowmodelwithvariablediffusionandexperimentalvalidation
AT yerlanamanbek applicationofphysicsinformedneuralnetworksfortwophaseflowmodelwithvariablediffusionandexperimentalvalidation