Overview
This research investigates the application of machine learning (ML) in conjunction with various data features for predicting the permeability of porous media. The study focuses on synthetic representations of porous media, leveraging separately computed exact permeability values as ground truth. It explores the utility of structural, topological, and network measures as input variables for a standard ML algorithm.
Research Context
Flow in porous media is recognized as a complex phenomenon, often challenging to analyze using conventional analytical or numerical methods. The increasing availability of synthetic porous media representations and data from physical experiments creates an opportunity for research incorporating machine learning approaches. This context motivates the exploration of ML techniques to address the complexities associated with porous media permeability prediction.
Approach
The study's approach involved the extraction of a range of features from porous media data for use as input variables in a standard machine learning algorithm. These features were categorized into three types:
- Structural measures: These describe the geometric properties of the porous media.
- Topological measures: These characterize the connectivity within the porous media, with a specific focus on features derived from Topological Data Analysis (TDA).
- Network measures: These are obtained by simplifying the porous media into pore networks.
The research design included comparing results generated using different input variables. This comparison aimed to provide insight into the effectiveness of various measures for predicting permeability based on the underlying porous media structure. The study specifically investigated how TDA features could be integrated with ML for permeability prediction. The ground truth permeability was computed separately for the synthetic porous materials.
Findings
- The study demonstrated that structural, topological, and network measures can serve as input variables for a standard machine learning algorithm to predict the permeability of synthetic porous materials.
- Comparison of results obtained using different input variables indicated varying utility of these measures for predicting permeability based on porous media structure.
- Topological Data Analysis (TDA) was found to provide a useful set of features.
- These TDA features can be combined with machine learning techniques to yield meaningful results in permeability prediction.