Particle Trajectory Prediction in Discrete Element Simulations using a Graph-Based Interaction-Aware Model

Abhishek Setty, Lukas Morand, Poojitha Ramachandra, Claas Bierwisch · 2025

Read Particle Trajectory Prediction in Discrete Element Simulations using a Graph-Based Interaction-Aware Model by Abhishek Setty, Lukas Morand, Poojitha Ramachandra, Claas Bierwisch (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-00215.txt.

Introduction

Abhishek Setty and co-workers foretold tiny bit motion in powder mock runs with a traffic-born graph model in a 2025 paper. Discrete element mock runs track every grain in mixers and hoppers. They are slow, so a learned surrogate that forecasts trajectories could speed design loops enormously. The team adapted a contact-aware graph network built for forecasting vehicles, whose nodes and edges suit tiny bits and contacts naturally.

Two test flows anchor validation: a unit cell with periodic boundaries under a wavy velocity profile, and a shear cell with sliding boundaries. New data transformations teach the model each boundary style, and two training schemes. One centered on positions, one on velocities, are matched head to head. Traffic intuition transfers to grains better than expected.

Catalog record

CREATORAbhishek Setty, Lukas Morand, Poojitha Ramachandra, Claas Bierwisch
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-00215

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Particle Trajectory Prediction in Discrete Element Simulations using a Graph-Based Interaction-Aware Model Abhishek Setty, Lukas Morand, Poojitha Ramachandra, Claas Bierwisch Abstract This study explores the applicability of a graph-based interaction-aware trajectory prediction model, originally developed for the transportation domain, to forecast particle trajectories in three-dimensional discrete element simulations. The model and our enhancements are validated at two typical particle simulation use cases: (i) particle flow in a representative unit cell with periodic boundary conditions (PBCs) in combination with sinusoidal velocity profile and (ii) shear flow in a representative unit cell with Lees - Edwards boundary conditions (LEBCs). For the models to learn the particle behavior subjected to these boundary conditions requires additional data transformation and feature engineering, which we introduce. Furthermore, we introduce and compare two novel training procedures for the adapted prediction model, which we call position - centric training (PCT) and velocity - centric training (VCT). The results show that the models developed for the transportation domain can be adapted to learn the behavior of particles in discrete element simulations. 1 Fraunhofer Institute for Mechanics of Materials IWM, Freiburg, Germany 2 Forschungszentrum Jülich, Institute of Quantum Control (PGI-8), D-52425 Jülich, Germany \dagger [email protected] Keywords: Trajectory Prediction, Particle Simulation, Discrete Element Method, Periodic Boundary Conditions, Lees - Edwards Boundary Conditions. 1 Introduction 1.1 Motivation The study of granular materials and complex fluid flow, which involves particle interactions governed by Newton’s equations of motion, holds significant importance for various industrial processes, such as food powder production, pharmaceutical manufacturing, and bulk handling [ 1 , 2 , 3 ] . To compute the motion of typically thousands to millions of particles in a system, the discrete element method (DEM) is the means of choice [ 4 ] . By accurately modeling the particle interactions using simple contact forces, DEM simulations can provide insights into the macroscopic properties of the material based on its microscopic interactions [ 5 , 6 ] . However, DEM simulations face significant constraints related to the time step size and the associated computational effort. The time step must be small enough to accurately capture the dynamics of particle interactions, particularly during collisions [ 7 ] . This requirement ensures numerical stability and accuracy but also increases the total number of time steps needed, thereby escalating the computational effort. These simulations are computationally intensive due to the necessity of calculating forces and updating positions for a large number of particles at each time step [ 8 ] . Additionally, storing the state of each particle and their interactions demand substantial memory, especially for large-scale simulations. To manage the computational load, simulations often need to be parallelized, which introduces further complexity in terms of programming and resource management, especially when using a graphics processing unit (GPU) [ 9 , 10 ] . The use of machine learning approaches to predict particle motion has the potential to remove the time-step constraint in DEM simulations, resulting in several significant benefits. Firstly, it allows for larger time steps, thereby reducing the total number of time steps required for a simulation. Also, there is a scope for a significant reduction in computational effort and time, making it possible to simulate larger systems or longer time periods [ 11 , 12 ] . Additionally, machine learning models are able to capture complex particle interactions in the domains such as gas - particle flow [ 13 , 14 ] , fluid flow for spherically shaped particles [ 15 ] , hydrodynamics [ 16 , 17 ] and also in detecting contacts of non-spherical particles [ 18 ] , improving the realism of the simulations. This can enable more detailed and precise predictions of material behavior under various conditions. Furthermore, the reduced computational load frees up resources, allowing for more extensive parameter studies and optimizations, ultimately accelerating research and development. In general, the works described above are part of a current trend in material science, namely the emergence of data-driven methods [ 19 ] . Since the last decade, data-driven models have been applied to various problems including representing materials relations [ 20 ] , designing materials and processes [ 21 ] , parameter identification [ 22 ] or experimentation analysis [ 23 ] . Recent trends are to bring together modern (semantic) data management techniques and software applications (e.g. machine learning applications) on platform environments, such as described in [ 24 ] . By providing data following the FAIR principles (findable, accessible, interoperable, and reusable), we can expect data-driven methods to reach a new level of effectiveness [ 25 ] . 1.2 Related Work One of the recent examples of accelerating DEM simulations using machine learning is the reduced-order modeling approach described in [ 26 ] . Therein, a ridge regression model trained on data from numerous DEM simulations is used to predict new system states. Another example is the use of pre-trained Convolutional Neural Networks (CNNs) to replace direct calculations of particle-particle collision and particle-boundary collision behavior and thereby accelerate the computational speed of a DEM simulation (\text{Conv-DEM}) [ 27 ] . This Conv-DEM model has been tested on granular flow in a rotating drum and a hopper. Instead of data-driven methods for predicting granular flow, an image-driven prediction was proposed for a wedge-shaped hopper DEM simulation [ 28 ] . However, this image-based method has the limitation of not being able to capture particle interactions and predict future trajectories. One of the early ideas to capture interactions among particles over time is to construct dynamic interaction graphs (aka DPI-Nets) [ 29 ] . This has proven to be working well for interactions of rigid bodies, deformable objects, and fluids. Alternatively, a continuous CNN has been developed and argued to be more accurate and faster than DPI-Nets for learning fluid flows [ 30 ] . Inspired by these results in fluid simulations by continuous convolutions, Walters et al. [ 31 ] tackled the problem of trajectory prediction of vehicles using continuous convolutions. In addition to this approach, graph networks have emerged as a promising technique for learning particle simulations, where each particle is represented as a node in a graph [ 32 ] . Similarly, graph-based trajectory prediction is gaining popularity in autonomous driving, which is of particular importance for the presented work. One of the popular choices in the transportation domain is the graph-based interaction-aware trajectory prediction (GRIP++) model introduced by Li et al. [ 33 ] . This method was ranked 1 in the ApolloScape trajectory competition in October 2019. GRIP++ was developed on top of other established methods such as convolutional social pooling [ 34 ] . Among the survey done on trajectory prediction methods for autonomous driving [ 35 ] , GRIP++ was found to be one of the preferred ones in graph convolutional network-based methods. Despite its success, to the authors’ knowledge, the GRIP++ model has not yet been applied to problems outside of the transportation domain. Only in a preliminary study, which formed the basis for the present work, the advantages of using GRIP++ in molecular dynamics were analyzed [ 36 ] . Therefore, together with our preliminary study [ 36 ] , this work is the first to assess the applicability of the GRIP++ model outside of

Read the complete text