Development of a GPU-based DEM solver for parameter optimization in the simulations of soil-sweep tool interactions

Nagy, Dániel and Pásthy, László and Tamás, Kornél (2024) Development of a GPU-based DEM solver for parameter optimization in the simulations of soil-sweep tool interactions. COMPUTERS AND ELECTRONICS IN AGRICULTURE, 227. No.-109482. ISSN 0168-1699

Preview	Text GPU_DEM.pdf - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (3MB) | Preview
Preview	Text (Graphical abstract) graphical_abstract.jpg - Additional Metadata Download (38kB) | Preview

Abstract

The Discrete Element Method (DEM) is a powerful technique for simulating granular materials in agricultural applications, yet it is notoriously computationally and memory-intensive. A typical DEM simulation of soil-sweep tool interactions involves tens of thousands of particles, each of which may potentially interact with many others. This results in hundreds of thousands of interactions being computed at every time step, while the need for numerical stability often requires very small time steps. Despite these challenges, DEM holds significant promise to allow the design of more efficient agricultural tools. This paper introduces an in-house-developed modular library based on CUDA C++ for GPUs, aimed at accelerating DEM simulations using a single GPU. The library is designed to facilitate efficient memory usage, employing a per-thread approach where each GPU thread computes one discrete-element particle. To accelerate particle–particle contact searches, we implemented a cell-linked-list algorithm. Our library utilizes the Hertz–Mindlin contact model, which has been widely adopted in agricultural DEM applications. Validation of the code was performed through comparisons with commercial software. Using our software, experimental measurements of a sweep tool moving through sandy soil were replicated with high accuracy by employing differential evolution for parameter calibration, achieving these results using 38 912 particles and running 2 700 instances within 16 h on a single GPU.

Actions (login required)

Edit Item