FluidNet-Lite: Lightweight convolutional neural network for pore-scale modeling of multiphase flow in heterogeneous porous media

Mohammed Yaqoob, Mohammed Yusuf Ansari, Mohammed Ishaq, Unais Ashraf, Saideep Pavuluri, Arash Rabbani, Harris Sajjad Rabbani, Thomas D. Seers

Abstract: Highlights•Extended Dataset for Porous Media: Released 657 DNS samples from OpenFOAM, tailored for pore-scale modeling and breakthrough pattern analysis.•Lightweight Network Architecture: Developed FluidNet-Lite, a compact CNN that integrates viscosity ratio and contact angle for accurate flow modeling.•Physics-Aware Loss Function: Introduced GWAL, a loss function that enforces physics constraints for realistic fluid displacement predictions.•High Computational Efficiency: FluidNet-Lite is 16.67×<math><mo is="true">×</mo></math> faster and 1.68×<math><mo is="true">×</mo></math> more memory efficient than GANs, achieving IoU 0.92 and SSIM 0.89.

External IDs:doi:10.1016/j.advwatres.2025.104952