InterPore2026

Investigating Machine Learning Models for Pore-Scale Multiphase Flow Using Lattice Boltzmann Simulations

19 May 2026, 14:35

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Chunyang Wang (Imperial College London)

Description

Machine learning is increasingly used to accelerate or replace pore-scale simulations of multiphase flow in porous media, yet a clear understanding of how different model classes perform—and fail—under realistic flow conditions remains limited. In particular, it is often unclear which modelling choices are most appropriate for capturing interfacial dynamics, geometry–flow coupling, and temporal evolution at the pore scale.

In this study, we present a systematic benchmark of representative machine learning architectures for pore-scale multiphase flow prediction using datasets generated from lattice Boltzmann method (LBM) simulations on micro-CT–derived pore geometries. The datasets span a range of flow conditions and multiphase configurations while retaining full access to velocity fields, phase distributions, and geometric information. This enables controlled evaluation of model behaviour under well-defined physical settings.

We compare several classes of ML models, including convolutional neural networks operating on voxelised domains, graph-based representations of pore networks, and autoregressive temporal models. Performance is assessed in terms of short-term prediction accuracy, stability under multi-step rollout, sensitivity to pore geometry, and generalisation across flow regimes. Beyond aggregate error metrics, we examine qualitative failure modes, such as loss of interfacial sharpness, accumulation of long-term drift, and reduced robustness near rapid interface rearrangements.

Based on insights from these benchmarks, we explore modest extensions to existing training strategies, including geometry-aware conditioning and rollout-consistent supervision, aimed at improving stability and interpretability rather than maximal accuracy. The results provide a practical reference for selecting and training ML models for pore-scale multiphase flow, and clarify the trade-offs involved when using data-driven surrogates alongside conventional LBM simulations.