NeuralDEM -- Real-time Simulation of Industrial Particulate Flows

TL;DR

NeuralDEM addresses the prohibitive computational cost of traditional DEM and CFD-DEM simulations by learning end-to-end deep surrogates that treat DEM dynamics as a continuous field while simultaneously modeling macroscopic processes with auxiliary fields. The core innovations are a physics representation that operates on field quantities and a multi-branch transformer framework that couples main-physics and macro-quantities, enabling real-time rollouts for large-scale particulate systems. Demonstrated on hopper drainage and CFD-DEM fluidized bed reactors, NeuralDEM achieves faithful long-term behavior, accurate macroscopic metrics, and generalization to unseen material properties and geometries, with substantial speedups on modern hardware. This approach eliminates the need for fine-grained microscopic parameter calibration and opens pathways for rapid design cycles and real-time engineering decision-making in industrial particulate flows.

Abstract

Advancements in computing power have made it possible to numerically simulate large-scale fluid-mechanical and/or particulate systems, many of which are integral to core industrial processes. Among the different numerical methods available, the discrete element method (DEM) provides one of the most accurate representations of a wide range of physical systems involving granular and discontinuous materials. Consequently, DEM has become a widely accepted approach for tackling engineering problems connected to granular flows and powder mechanics. Additionally, DEM can be integrated with grid-based computational fluid dynamics (CFD) methods, enabling the simulation of chemical processes taking place, e.g., in fluidized beds. However, DEM is computationally intensive because of the intrinsic multiscale nature of particulate systems, restricting simulation duration or number of particles. Towards this end, NeuralDEM presents an end-to-end approach to replace slow numerical DEM routines with fast, adaptable deep learning surrogates. NeuralDEM is capable of picturing long-term transport processes across different regimes using macroscopic observables without any reference to microscopic model parameters. First, NeuralDEM treats the Lagrangian discretization of DEM as an underlying continuous field, while simultaneously modeling macroscopic behavior directly as additional auxiliary fields. Second, NeuralDEM introduces multi-branch neural operators scalable to real-time modeling of industrially-sized scenarios - from slow and pseudo-steady to fast and transient. Such scenarios have previously posed insurmountable challenges for deep learning models. Notably, NeuralDEM faithfully models coupled CFD-DEM fluidized bed reactors of 160k CFD cells and 500k DEM particles for trajectories of 28s. NeuralDEM will open many new doors to advanced engineering and much faster process cycles.

Figures (18)

• Figure 1: NeuralDEM presents an end-to-end approach to replace discrete element method (DEM) routines and coupled multiphysics simulations with deep learning surrogates. Top: Hopper simulations. NeuralDEM treats inputs and outputs as continuous fields, while modeling macroscopic behavior directly as additional auxiliary fields. Bottom: Fluidized bed reactors. NeuralDEM is built to model complex multiphysics simulations, i.e., scenarios which necessitate the interaction of DEM and computational fluid dynamics (CFD). For fluidized bed reactors, air enters the domain from the bottom plane (CFD problem) and pushes the particles up (DEM problem). Both parts and the interaction thereof is modeled via the new multi-branch neural operator approach of NeuralDEM.
• Figure 2: Discrete element method. The force on a particle consists of particle-particle contacts ${\boldsymbol{F}}_i^{\text{(pc)}}$, the external force ${\boldsymbol{F}}_i^{\text{(ext)}}$, and the interaction with a surrounding fluid phase ${\boldsymbol{F}}_i^{\text{(pf)}}$.
• Figure 3: Neural operator learning. Neural operators aim to learn a mapping between function spaces, enabling outputs that remain consistent across varying input sampling resolutions. The neural operator $\hat{\mathcal{G}}$ approximates the ground truth operator $\mathcal{G}$ with three maps, composing encoder $\mathcal{E}$, approximator $\mathcal{A}$, and decoder $\mathcal{D}$. The approximation of $\hat{\mathcal{G}}$ is ideally independent from the number of sampled input points, and approximates the output function for an arbitrary number of points.
• Figure 4: In our physics representation we model the Lagrangian discretization of DEM as an assumed underlying continuous field. The approximator maps the encoded representation to one that can be decoded at any specified spatial location $j'$. The multi-branch neural operator is a family of deep learning architectures that processes multi-physics quantities and can distinguish between primary quantities, used to model the core physics in the main-branches, and secondary quantities which are used to predict additional desired quantities in the off-branches, both modeled as fields. The quantities come, e.g., from DEM simulations with coupled particles and fluid, which the architecture handles using specialized encoders and decoders. All modules processing the primary quantities influence each other. In contrast, those that process secondary quantities are independent, and use the tokens from the primary branch as additional information but cannot affect them.
• Figure 5: Schematic architecture of a multi-branch transformer block. DiT li22dit modulation is applied to each attention and MLP block but is omitted for visual clarity.