InterPore2025

Advanced Pore Network Modelling (PNM) for Packed-Bed Reactors: A Neural Network Approach Using PR-CFD Data

19 May 2025, 17:10

15m

Oral Presentation (MS15) Machine Learning and Big Data in Porous Media MS15

Speaker

Cristina García Llamas (Eindhoven University of Technology)

Description

Packed-bed reactors are widely used due to their efficient heat and mass exchange. The hydrodynamics of these reactors is largely influenced by particle-fluid interactions. However, a unified theory representing the effect of particle shape on flow distribution and global parameters such as pressure drop is not well-described in the literature. Studying the hydrodynamics in a packed bed using Particle Resolved Computational Fluid Dynamics (PR-CFD) techniques represents an excellent alternative to experimentation. Nonetheless, PR-CFD is computationally expensive, whereas less resolved methods like Pore Network Modeling (PNM) offer a viable alternative. The accuracy of PNM simulations depends largely on the definition of physically accurate calibration factors, which can be derived from highly resolved PR-CFD simulations. This research aims to train an Artificial Neural Network (ANN) using PR-CFD data to obtain physically accurate calibration coefficients for various particle packings and particle size distributions.
We focus on packed bed reactors where the spacing between the walls is sufficiently large to disregard the influence of confining walls. We adopt a unit-cell approach, where the packed bed configuration is derived by periodically repeating this unit-cell in all three dimensions. This method has proven effective in modeling transport processes under steady-state conditions.

• Particle Resolved Computational fluid Dynamics (PR-CFD)
  The Navier-Stokes and continuity equations are solved under the assumption of incompressibility conditions. The computational domain is discretized with a conformal mesh composed of hexahedral elements. Periodic boundary conditions are applied across all domain boundaries.

• Pore Network Modelling (PNM)
  The PNM model utilizes the 3D structure information of the packed bed to generate a network of pores and throats. The hydrodynamic equations are solved by computing the mechanical energy and mass balance equations for each pore-throat-pore element. The overall pressure drop is model using a variation of the Hagen-Poiseulle equation that accounts for frictional losses. Due to the simplification of the particle packing into a pore network, the frictional forces associated with the particle shape are not properly described. Therefore, a resistance term should be incorporated to account for the local pore geometry. To obtain this resistance term, a calibration procedure is performed using an Artificial Neural Network (ANN), which is explained below.

• Calibration of the PNM model using an Artificial Neural Network (ANN)
  PR-CFD simulations are conducted using a wide range of particle packings. The ANN is trained using this particle-resolved Computational Fluid Dynamics (CFD) data, where each throat within a pore network serves as an individual data point for model training. The flow resistances associated to each throat are determined across a broad spectrum of local pore geometries. These flow resistances can subsequently be integrated into the PNM model as inputs to accommodate the throat-specific flow resistance associated with the local geometry.

This methodology enables physically accurate PNM simulations of packed beds for a wide range of particle shapes and polydispersities at a relatively low computational cost, which is critical for an optimal design of packed bed reactors.