InterPore2026

Surrogate Modeling of Particle Retention in Porous Media Enabled by a Massive Pore-Scale Simulation Dataset

20 May 2026, 11:35

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Dr Saeid Sadeghnejad (Institute for Geosciences, Applied Geology, Friedrich-Schiller-University Jena, 07749 Jena, Germany)

Description

The retention of suspended particles in porous media plays a critical role in a wide range of subsurface processes, including filtration, contaminant transport in environmental applications, and formation damage in subsurface energy applications. As flow with suspended particles flow through porous media, they may deposit or clog flow pathways, changing local porosity, and ultimately impacting large-scale hydraulic behavior (permeability). Although pore-scale computational fluid dynamics (CFD) coupled with discrete element models (DEMs) can resolve these mechanisms, their high computational cost prevents extensive sensitivity analyses. Moreover, the absence of large pore-scale datasets suitable for surrogate modeling represents a major research gap.
To address this, we systematically extended the pore-scale model of Sadeghnejad et al. (2022) to generate a large-scale dataset for machine-learning surrogate development. Key physical and geometric parameters, including particle size, concentration, injection velocity, and pore-space morphology, were varied across wide ranges. For each realization, the Eulerian-Lagrangian workflow (including Navier-Stokes flow simulation, individual particle tracking modeling, dynamic voxel-based deposition, and porosity/permeability updating) was executed until steady post-retention conditions were achieved. Approximately 130,000 simulation points were run, consuming ~49,000 CPU-hour, which is one of the largest particle-retention datasets reported to date. Moreover, outliers of the dataset were removed by the Isolation Forest algorithm. Seven machine learning models (i.e., Adaptive Gradient Boost (AGB), Decision Tree (DT), Extremely Randomized Trees (XRT), Extreme Gradient Boost (XGB), Gradient Boost Machine (GBM), Multi-layer Perceptron (MLP), and Random Forest (RF)) were trained on 80% of the dataset with standard hyperparameter values to predict the final porosity and permeability of the domain after particle deposition.
Initial evaluations identified XGB and XRT as the most promising surrogate candidates. Both models were subsequently refined through Bayesian hyperparameter optimization to enhance predictive robustness and generalization. Model performance was assessed using five-fold cross-validation and the metrics Mean Squared Error (MSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). The optimized models achieved excellent predictive accuracy, with R² values exceeding 0.98 for porosity and 0.90 for permeability, respectively. In addition to their accuracy, these surrogates provide orders-of-magnitude faster inference than pore-scale simulations, underscoring their suitability for rapid assessment of particle-retention behavior. Comparative performance metrics and predictive outcomes are illustrated in the following figure.