Permeability Prediction of Porous Media using CNN with Physical Properties
summary
Convolutional network trained on ~350 synthetic bead‑packing CT images predicts permeability by combining image features with porosity and surface area via MLP, using a VGG16 encoder transfer learning backbone for segmentation that cuts labeling from months to days. Augmenting CNN outputs with those two physical inputs raises test performance (slope/error metric ~0.8 versus ~0.69 for image‑only or empirical correlations). Models are checked against Stokes/flow solver results for selected samples.
AI-written summary · about
chapters (16)
- 0:00Intro
- 3:14Motivation: Flow and Reactive Transport
- 5:55Physics-Informed Machine Learning
- 6:57Semantic Image Segmentation
- 9:48Physics of Interest: Fluid Flow in Porous Media
- 10:42Data Generation (Images, Pore Network, Properties)
- 11:22Porous Media Properties
- 11:4810 Physics-Informed ML for permeability prediction
- 13:21111 CNN architecture with physical information
- 14:11al Impact of physical data
- 15:56wl Comparison with a shallow CNN and numeric data input
- 16:37l Impact of the Size of MLP Output
- 17:47Impact of combining method of physical data
- 18:19l Impact of Binary Image Input
- 19:25Visualizing Activations
- 20:211. Near Future Works
related talks
24mAutomating Porosity Detection with Deep Learning
47mInterpreting Deep Neural Networks (DNNs)
30mTransfer learning for image recognition in healthcare...
55mDeep Learning in Medical Image Diagnostics
51mCognition and Computation
39mInsights About Places with Deep Learning Computer Vision
40mPutting Deep Hybrid Learning to Work
44mWorking with Fashion Models