This paper introduces a deep learning workflow to predict phase distributions within complex geometries during two-phase capillary-dominated drainage.We utilize subsamples from Computerized Tomography(CT)images of roc...This paper introduces a deep learning workflow to predict phase distributions within complex geometries during two-phase capillary-dominated drainage.We utilize subsamples from Computerized Tomography(CT)images of rocks and incorporate pixel size,interfacial tension,contact angle,and pressure as inputs.First,an efficient morphology-based simulator creates a diverse dataset of phase distributions.Then,two commonly used convolutional and recurrent neural networks are explored and their deficiencies are highlighted,particularly in capturing phase connectivity.Subsequently,we develop a Higher-Dimensional Vision Transformer(HD-ViT)that drains pores solely based on their size,with phase connectivity enforced as a post-processing step.This enables inference for images of varying sizes,resolutions,and inlet-outlet setup.After training on a massive dataset of over 9.5 million instances,HD-ViT achieves excellent performance.We demonstrate the accuracy and speed advantage of the model on new and larger sandstone and carbonate images.We further evaluate HD-ViT against experimental fluid distribution images and the corresponding Lattice-Boltzmann simulations,producing similar outcomes in a matter of seconds.In the end,we train and validate a 3D version of the model.展开更多
基金supported by the International Cooperation Programme of Chengdu City(No.2020-GH02-00023-HZ)。
文摘This paper introduces a deep learning workflow to predict phase distributions within complex geometries during two-phase capillary-dominated drainage.We utilize subsamples from Computerized Tomography(CT)images of rocks and incorporate pixel size,interfacial tension,contact angle,and pressure as inputs.First,an efficient morphology-based simulator creates a diverse dataset of phase distributions.Then,two commonly used convolutional and recurrent neural networks are explored and their deficiencies are highlighted,particularly in capturing phase connectivity.Subsequently,we develop a Higher-Dimensional Vision Transformer(HD-ViT)that drains pores solely based on their size,with phase connectivity enforced as a post-processing step.This enables inference for images of varying sizes,resolutions,and inlet-outlet setup.After training on a massive dataset of over 9.5 million instances,HD-ViT achieves excellent performance.We demonstrate the accuracy and speed advantage of the model on new and larger sandstone and carbonate images.We further evaluate HD-ViT against experimental fluid distribution images and the corresponding Lattice-Boltzmann simulations,producing similar outcomes in a matter of seconds.In the end,we train and validate a 3D version of the model.
