InterPore2026

Decoupling the Non-linear Influence of Pore Structure on CO₂ Saturation: An Explainable Data-Driven Approach based on Microfluidic Experiments

19 May 2026, 13:50

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

晗 葛 (浙江大学)

Description

Geological CO₂ sequestration efficiency relies on pore-scale structural parameters. However, the complex, non-linear coupling among these parameters is difficult to quantify using traditional experimental correlations alone. In this study, we apply an explainable machine learning (ML) framework to uncover the dominant governing factors of CO₂ saturation, utilizing a high-fidelity dataset derived from our systematic microfluidic displacement-imbibition experiments. The dataset encompasses a wide range of topological scenarios, where pore-size distribution, pore-throat ratio, and coordination number were independently varied under different capillary numbers. We developed a multi-modal deep learning model that integrates Convolutional Neural Networks (CNN) for extracting topological features from experimental images and Multi-Layer Perceptrons (MLP) for processing numerical structural parameters. This hybrid architecture maps the inputs to initial and residual CO₂ saturation, achieving high predictive accuracy (R² ≈ 0.95) and robust stability across cross-validation folds (standard deviation < 0.05). Crucially, to move beyond "black-box" prediction, we employed SHAP (SHapley Additive exPlanations) analysis to decouple the interactions between topological features. The analysis reveals that pore-size distribution characteristics and structural heterogeneity are the primary predictors, exhibiting a non-linear influence that standard linear regression fails to capture. Furthermore, the ML-derived feature importance aligns with the physical mechanism of capillarity-connectivity competition, confirming that the coordination number and pore-throat ratio jointly dictate the capillary-viscous transition. This work demonstrates that applying explainable AI to experimental datasets provides a robust pathway for identifying critical sequestration criteria in heterogeneous porous media.