InterPore2026

Data-Driven Prediction of Relative Permeability: Applications to CO₂ and Hydrogen Storage

22 May 2026, 09:50

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Mr Abdolali Mosallanezhad (PhD Student, Research Centre for Carbon Solutions (RCCS), School of Engineering and Physical Sciences, Heriot-Watt University, Edinburgh, UK)

Description

Relative permeability curves are one of the fundamental parameters in multiphase flow modelling, supporting applications that now extend into Carbon Capture and Storage and Underground Hydrogen Storage. These curves are traditionally obtained experimentally using sophisticated special core analysis instruments, resulting in a workflow that relies on a limited number of core plugs that cannot fully capture reservoir heterogeneity. As interest in subsurface storage increases, there is a clear shift towards data-driven approaches that can connect sparse, complex measurements with the continuous property fields required by reservoir simulators. Accordingly, this study aims to apply machine learning techniques to predict relative permeability curves for water (krw) and gas (krg) in sandstone cores during drainage experiments.

The work described here is built on a moderate-sized dataset of approximately fifteen hundred data points, each characterised by a set of features that includes temperature, pressure, porosity, absolute permeability, and key fluid properties such as gas and brine viscosities and their ratio. Along with normalised water saturation and irreducible water saturation, these variables offer a realistic testbed for modern data-driven petrophysical modelling in systems relevant to gas and brine. The complete analysis will include the modelling workflow, explore how predictions respond to other key inputs such as Interfacial Tension and wettability, and provide an initial investigation into how this framework can be extended to incorporate detailed rock and fluid characteristics and broader gas–brine systems, thereby enhancing the transferability and efficiency of relative permeability modelling for subsurface storage applications.

Across the literature, there is a trade-off between model flexibility and physical consistency. Conventional regressions and unconstrained neural networks fit the data but often violate key constraints, especially the fact that relative permeability lies between 0 and 1. Deep networks tend to overfit and break monotonic saturation trends, while tree ensembles like XGBoost and kernel methods like Gaussian Process Regression perform well, with GPR quantifying uncertainty. Building on these insights, we trained monotonic XGBoost models on CO₂–brine drainage experiments in sandstone, using the above features to predict four quantities at each point, namely irreducible water saturation, gas relative permeability at irreducible water saturation, and the normalised water and gas relative permeabilities.

Finally, the model is evaluated on a held-out test set that covers the full range of experimental conditions in temperature, pressure, permeability, and viscosity ratio, providing a direct assessment of its ability to interpolate within realistic conditions. Initial results for a CO2-brine system indicate that the monotonic XGBoost surrogate accurately reproduces the normalised water relative permeability, achieving an R² of 0.9829 and a mean squared error (MSE) of 0.001725, corresponding to a root mean squared error (RMSE) of approximately 0.0415 on a held-out test set. For the gas phase, the model achieves an R² of 0.9747, an MSE of 0.002670, and an RMSE of 0.0517. The close agreement with SCAL measurements (Figure 1) indicates that this method can serve as a reliable predictive tool when laboratory data are sparse or unavailable, therefore helping to reduce experimental workload and costs while still providing simulation-ready kr curves.