SNIC SUPR
Studies of granular surrogates
Dnr:

SNIC 2019/3-168

Type:

SNIC Medium Compute

Principal Investigator:

Martin Servin

Affiliation:

UmeƄ universitet

Start Date:

2019-04-01

End Date:

2020-04-01

Primary Classification:

10399: Other Physics Topics

Secondary Classification:

10207: Computer Vision and Robotics (Autonomous Systems)

Tertiary Classification:

10106: Probability Theory and Statistics

Webpage:

Allocation

Abstract

A 'granular surrogate' is a data-driven model that approximates the macroscopic properties of a real or simulated granular system with small computational effort. There is a great interest from the construction, forestry and mining industries to create digital twins of real systems, using granular surrogate models to monitor processes in real-time. In these applications, the granular surrogates are connected to the local IT infrastructure of the worksite, intercommunicating with connected vehicles and processing equipment to provide information about the granular material in a way that no real sensors can. This enables material tracking, fault detection, production planning and fleet management at a new level of precision. The aim of the project is to investigate methods for creating granular surrogates from detailed and computational intense simulations. The surrogate models need to be trained, off-line, on large data-sets from numerous numerical simulations that cover the relevant state-space. One goal is to develop a scheme for running large amounts (unfeasible for desktop computers) of granular flow simulations in parallel, needed for the development of surrogate models. An interface for batch simulation management on HPC infrastructure, targeting academic users of granular simulation, will be created in collaboration with the software provider Algoryx Simulation. A second goal is to test different methods for generating surrogate models, by running simulations, post-processing and model training. The third project goal is to investigate the parallel efficiency of individual granular simulations and the potential for running large-scale simulations that occupy one or several nodes each.