Data-Driven Closure Parametrizations with Metrics: Dispersive Transport

TL;DR

This paper develops a data-driven framework using CNNs and metrics to predict dispersive transport parameters in porous media, validated on pore geometries, with a focus on improving longitudinal dispersivity predictions.

Contribution

It introduces a novel combination of pore-scale simulations, volume-averaging, and neural networks to predict dispersivity directly from pore geometry and metrics.

Findings

CNN accurately predicts longitudinal dispersivity

Metrics help understand dispersivity parameter space

Transversal dispersivity prediction remains challenging

Abstract

This work presents a data-driven framework for multi-scale parametrization of velocity-dependent dispersive transport in porous media. Pore-scale flow and transport simulations are conducted on periodic pore geometries, and volume-averaging is used to isolate dispersive transport, producing parameters for the dispersive closure term at the Representative Elementary Volume (REV) scale. After validation on unit cells with symmetric and asymmetric geometries, a convolutional neural network (CNN) is trained to predict dispersivity directly from pore-geometry images. Descriptive metrics are also introduced to better understand the parameter space and are used to build a neural network that predicts dispersivity based solely on these metrics. While the models predict longitudinal dispersivity well, transversal dispersivity remains difficult to capture, likely requiring more advanced models to fully describe pore-scale transversal dynamics.