NeuralDEM: A Real-Time Deep Learning Surrogate for Industrial Granular Flow Simulation

Progress in computational capabilities has enabled the numerical modeling of extensive fluid-mechanical and particulate systems, which are essential to many fundamental industrial operations. Of the various numerical techniques, the discrete element method (DEM) offers one of the most precise depictions of diverse physical systems featuring granular and fragmented materials. As a result, DEM has gained broad acceptance for addressing engineering issues related to granular flows and powder dynamics. Moreover, DEM can be combined with grid-based computational fluid dynamics (CFD) approaches to model chemical reactions, such as those in fluidized beds. Nevertheless, DEM's high computational demands arise from the inherent multiscale characteristics of particulate systems, limiting simulation length or particle quantity. To overcome this, NeuralDEM offers a comprehensive framework that substitutes sluggish DEM computations with rapid, flexible deep learning proxies. NeuralDEM can simulate extended transport phenomena across various flow regimes using only macroscopic variables, independent of microscopic parameters. Initially, it reinterprets DEM's Lagrangian mesh as a foundational continuous field, while also directly representing macroscopic traits as supplementary fields. Additionally, NeuralDEM employs multi-branch neural operators that scale to real-time simulation of industrial-scale setups, spanning from gradual pseudo-steady states to rapid transients—challenges that have historically eluded deep learning solutions. Impressively, NeuralDEM accurately replicates coupled CFD-DEM fluidized bed reactors involving ۱۶۰,۰۰۰ CFD cells and ۵۰۰,۰۰۰ DEM particles over ۲۸-second paths. NeuralDEM promises to unlock novel avenues in engineering innovation and accelerate process workflows significantly.

کلیدواژه ها:

NeuralDEM ، Discrete Element Method (DEM) ، real-time simulation ، Industry ۴.۰ ، sustainable optimization