Microstructure-informed probability-driven point-particle model for hydrodynamic forces and torques in particle-laden flows

Ahmadi, Arman; Wachs, Anthony

doi:10.1017/jfm.2020.453

Cited by 57 publications

(61 citation statements)

References 97 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Euler-Lagrange simulations, on the other hand, simulates individual particle trajectories using Newton's laws, thus relying on simpler closures and providing insight into the particle-fluid interaction and particle-particle collisions (Subramaniam 2013;Capecelatro & Desjardins 2015). Finally, particle-resolved direct numerical simulations (PR-DNS) are able to capture the flow field around each particle, providing data and relations that proved invaluable to inform statistical theories applicable at the meso-and macroscales (Tenneti & Subramaniam 2014;Luo et al 2016;Ozel et al 2017;Esteghamatian et al 2018;Seyed-Ahmadi & Wachs 2020;Tavanashad, Passalacqua & Subramaniam 2021). The high computational cost, however, limits the number of particles and flow regimes that can be simulated; thus, despite continuous growth in high-performance computing, this approach is still not viable to capture mesoscale structures (Sundaresan et al 2018).…”

Section: Introductionmentioning

confidence: 99%

Experimental analysis of particle clustering in moderately dense gas–solid flow

Fong

Coletti

2021

J. Fluid Mech.

View full text Add to dashboard Cite

In collisional gas–solid flows, dense particle clusters are often observed that greatly affect the transport properties of the mixture. The characterisation and prediction of this phenomenon are challenging due to limited optical access, the wide range of scales involved and the interplay of different mechanisms. Here, we consider a laboratory setup in which particles fall against upward-moving air in a square vertical duct: a classic configuration in riser reactors. The use of non-cohesive, monodispersed, spherical particles and the ability to independently vary the solid volume fraction ( $\varPhi _V = 0.1\,\% - 0.8\,\%$ ) and the bulk airflow Reynolds number ( $Re_{bulk} = 300 - 1200$ ) allows us to isolate key elements of the multiphase dynamics, providing the first laboratory observation of cluster-induced turbulence. Above a threshold $\varPhi _V$ , the system exhibits intense fluctuations of concentration and velocity, as measured by high-speed imaging via a backlighting technique which returns optically depth-averaged fields. The space–time autocorrelations reveal dense and persistent mesoscale structures falling faster than the surrounding particles and trailing long wakes. These are shown to be the statistical footprints of visually observed clusters, mostly found in the vicinity of the walls. They are identified via a percolation analysis, tracked in time, and characterised in terms of size, shape, location and velocity. Larger clusters are denser, longer-lived and have greater descent velocity. At the present particle Stokes number, the threshold $\varPhi _V \sim 0.5$ % (largely independent from $Re_{bulk}$ ) is consistent with the view that clusters appear when the typical interval between successive collisions is shorter than the particle response time.

show abstract

Section: Introductionmentioning

confidence: 99%

Experimental analysis of particle clustering in moderately dense gas–solid flow

Fong

Coletti

2021

J. Fluid Mech.

View full text Add to dashboard Cite

show abstract

“…This dilemma is referred to as the curse of dimensionality [214]. Two very promising frameworks that derive hydrodynamic forces acting on individual particles in a dense suspension have recently been proposed as the pairwise interaction extended point-particle (PIEP) [212] and the microstructure-informed probability-driven point-particle (MPP) model [18], respectively. Both models have in common that data from particle-resolved DNS of a flow through an array of fixed spheres is used to derive the effect of particle hiding and shading on the resulting fluid drag acting on individual particles in that same configuration (Fig.…”

Section: Machine Learningmentioning

confidence: 99%

“…The algorithm is optimized by using the drag force computed by the particle-resolved DNS as training data to quantify the pairwise interaction. This interaction can be mapped out in terms of force maps [212] or probabilities of finding a neighbor in a certain distance for a given recorded drag force [18]. Nonlinear least square methods, such as the Levenberg-Marquardt algorithm, are then used to optimize a set of open parameters that enter, e.g., a spherical harmonic expansion, to recover the hydrodynamic forces obtained from particle-resolved DNS with satisfactory agreement.…”

Section: Machine Learningmentioning

confidence: 99%

“…1): (i) the microscale of individual particle-particle interactions can be Fig. 1 Concept of upscaling in sediment transport simulations after Seyed-Ahmadi and Wachs [18]: the continuum model shows a turbidity current propagating along the x-direction (from Zhao et al [19]) represented by particle-resolved direct numerical simulations (DNS) that compute the motion of every grain by solving the Newton-Euler equations of motion and account for the fluid forces by resolving the flow field around it, (ii) the mesoscale that involves millions of particles is typically represented by particles smaller than the grid size of the fluid solver, so that grains are simplified as mass points and the relevant forces that govern their motion are computed by empirical correlations, and (iii) the macroscale can cover several kilometers, and the corresponding computational approach can no longer resolve the motion of individual grains. Instead, sediment transport is dealt with by continuum models of coupled transport equations.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Incorporating grain-scale processes in macroscopic sediment transport models

Vowinckel¹

2021

Acta Mech

View full text Add to dashboard Cite

Sediment transport simulations face the challenge of accounting for vastly different scales in space and time that cannot be tackled by a unifying approach. Instead, processes are subdivided into a microscale at the particle level, a mesoscale of a large finite number of particles, and a macroscale that computes the sediment motion by means of advection–diffusion equations. The different processes occurring at different scales are simulated using different computational approaches. However, modeling sediment transport at multiple scales with high fidelity requires proper closure arguments that interconnect the different processes. Ultimately, we will need efficient macroscale models that can readily be utilized for engineering practices covering, e.g., entire river reaches or even estuaries. In recent years, highly resolved simulations have become a valuable tool to provide these closure arguments for sediment transport models on the continuum scale. In this paper, we will review the most relevant approaches to simulate sediment transport at different scales and discuss the perspectives of four most promising modeling techniques that can help to improve sediment transport modeling. On the grain scale, these enhancements include the impact of mechanical properties of cohesion and biocohesion as well as the shape of non-spherical sediment grains on fluid–particle and particle–particle interactions. On larger scales, we review constitutive equations for the macroscopic rheological behavior of sediment beds that may decouple the relevant scales for fluid and sediment motion. Furthermore, we discuss machine learning strategies as an efficient means to derive scaling arguments across multiple scales.

show abstract

“…They ultimately complement their data-driven model with the original physicsdriven PIEP approach for performance improvements. In our own effort to develop a data-driven neighborhooddependent closure model, we proposed the microstructure-informed probability-driven point-particle (MPP) approach [2]. Using PR-DNS data of stationary arrays of particles, the MPP model first identifies consistent, non-random patterns of neighboring particle locations according to a selective data filtering strategy.…”

Section: Introductionmentioning

confidence: 99%

Physics-inspired architecture for neural network modeling of forces and torques in particle-laden flows

Seyed-Ahmadi,

Wachs

2021

Preprint

Self Cite

View full text Add to dashboard Cite

We present a physics-inspired neural network (PINN) model for direct prediction of hydrodynamic forces and torques experienced by individual particles in stationary beds of randomly distributed spheres. In line with our findings, it has recently been demonstrated that conventional fully connected neural networks (FCNN) are incapable of making accurate predictions of force variations in a static bed of spheres [1]. The problem arises due to the large number of input variables (i.e., the locations of individual neighboring particles) leading to an overwhelmingly large number of training parameters in a fully connected architecture. Given the typically limited size of training datasets that can be generated by particle-resolved simulations, the NN becomes prone to missing true patterns in the data, ultimately leading to overfitting. Inspired by our observations in developing the microstructure-informed probability-driven point-particle (MPP) model [2], we incorporate two main features in the architecture of the present PINN model: 1) superposition of pairwise hydrodynamic interactions between particles, and 2) sharing training parameters between NN blocks that model neighbor influences. These strategies helps to substantially reduce the number of free parameters and thereby control the model complexity without compromising accuracy. We demonstrate that direct force and torque prediction using NNs is indeed possible, provided that the model structure corresponds to the underlying physics of the problem. For a Reynolds number range of 2 Re 150 and solid volume fractions of 0.1 φ 0.4, the PINN's predictions prove to be as accurate as those of other microstructureinformed models.

show abstract

Microstructure-informed probability-driven point-particle model for hydrodynamic forces and torques in particle-laden flows

Abstract: Abstract

Cited by 57 publications

References 97 publications

Experimental analysis of particle clustering in moderately dense gas–solid flow

Experimental analysis of particle clustering in moderately dense gas–solid flow

Incorporating grain-scale processes in macroscopic sediment transport models

Physics-inspired architecture for neural network modeling of forces and torques in particle-laden flows

Contact Info

Product

Resources

About