Drag modelling for flows through assemblies of spherical particles with machine learning: A comparison of approaches

TL;DR

This paper compares machine learning approaches, including graph neural networks and genetic programming, to model drag forces in particle assemblies, aiming for interpretable and accurate models at higher Reynolds numbers.

Contribution

It extends previous GP-based drag modeling to higher Reynolds numbers by integrating GNNs and symbolic regression, enhancing interpretability and accuracy.

Findings

GNN effectively learns pairwise particle interactions.

Symbolic expressions approximate GNN results with slight accuracy loss.

GP can find simple, interpretable drag models.

Abstract

Drag forces on particles in random assemblies can be accurately estimated through particle-resolved direct numerical simulations (PR-DNS). Despite its limited applicability to relatively small assemblies, data obtained from PR-DNS has been the driving force for the development of drag closures for much more affordable simulation frameworks, such as Eulerian-Lagrangian point particle methods. Recently, more effort has been invested in the development of deterministic drag models that account for the effect of the structure of the particle assembly. Current successful deterministic models are mainly black-box neural networks which: 1) Assume pairwise superposition of the neighbours' effect on the drag, and 2) Are trained on PR-DNS data for a wide range of particle concentrations and flow regimes. To alleviate the black-box nature of neural networks, we use genetic programming (GP) to develop interpretable models. In our previous research, this has been proven successful in the Stokes regime. In the current contribution, we extend the application of GP to higher particle Reynolds number regimes. This is done by training a graph neural network (GNN) on the PR-DNS data to learn the pairwise interactions among the particles that constitute the drag variation. The significance of the input features of the GNN is assessed via a feature permutation approach. Then, the estimated pairwise interactions as extracted from the GNN are fed to a GP algorithm, which searches for symbolic expressions that fit the input data. A comparison between the trained GNN model and the resulting symbolic expressions is presented, to assess whether the symbolic expression can capture the underlying patterns learnt by the GNN. The comparison demonstrates the potential of GP in finding relatively simple symbolic models. At the same time, the accuracy of the symbolic models slightly fall behind the GNN.