En
期刊文章
文章精选
当期目录
过刊浏览
高被引
高下载
特邀文章
专刊合集
待刊文章
Volume 113
您当前的位置：首页 > 期刊文章 > 过刊浏览 > Volumes 108-119 (2025) > Volume 113
Hybrid surrogate–physics framework for rapid initialisation of granular packings in discrete element simulations (Open Access)
Fatih Uzun *
Department of Engineering Science, University of Oxford, Oxford, UK
10.1016/j.partic.2026.03.020
Volume 113, June 2026, Pages 102-112
Received 8 January 2026, Revised 26 February 2026, Accepted 9 March 2026, Available online 27 March 2026, Version of Record 2 April 2026.
E-mail: fatihuzun@me.com; fatih.uzun@eng.ox.ac.uk
Highlights

• Hybrid surrogate–physics framework accelerates DEM packing generation.

• Neural network predicts particle equilibrium, bypassing dynamic settling.

• Physics-based overlap correction ensures rigorous mechanical stability.

• The proposed pipeline significantly reduces required computational iterations.

• This hybrid method successfully reproduces coordination number statistics.


Abstract

The initialisation of mechanically stable granular packings is a computationally expensive prerequisite for many discrete element method (DEM) simulations, as conventional dynamic settling requires long transient phases dominated by repeated contact resolution and energy dissipation. This work presents a hybrid surrogate–physics framework for rapidly approximating static equilibrium packings while retaining physical admissibility. A particle-wise neural network surrogate is trained to predict approximate equilibrium particle coordinates directly from initial conditions and material properties, using reference solutions generated by an energy-based DEM formulation. Because particle-wise predictions do not explicitly enforce local contact topology, the surrogate output is post-processed through a mandatory overlap-correction stage followed by a short physics-based relaxation to recover mechanically stable configurations. A physics-aware composite loss function incorporating boundary and overlap penalties improves the geometric plausibility of the surrogate predictions. The framework is evaluated on small-scale representative granular packing scenarios as a methodological proof-of-concept for computationally expensive applications such as die filling. It reproduces physically consistent final states while offering a substantial reduction in the computational effort required to reach equilibrium compared to full dynamic settling. Validation using packing structure and coordination number statistics confirms that the hybrid approach recovers both global and local characteristics of mechanically stable granular assemblies.

Graphical abstract
Keywords
Discrete element method (DEM); Surrogate modelling; Hybrid simulation; Granular packing; Physics-aware machine learning; Multilayer perceptron (MLP)
文章导读