Versions Compared

Version Old Version 25 New Version Current
Changes made by

Timo Plath

Deepak Tunuguntla

Saved on

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

Earlier in November 2024, below food-for-thought information was shared with almost all of the WG1 members to help achieve our goal for 2025.

Benchmarking

Problem description

  • We will use a tumbling drum for our benchmark case as there should be experimental data from a number of groups along with it being simple enough that most codes can simulate. 

  • For the DEM Model we will use Hertz Mindlin as done in this paper with appropriate YM and time-step values, https://www.sciencedirect.com/science/article/pii/S0010465523004113 

    1. We will have a case with moderate particle numbers that we will use to calibrate against experiment. 

    2. We will do an additional case with a larger number of particles in a range of millions to benchmark performance

    3. We will do a shape case using cubes as they are simple enough that most codes will have some representation for it. We aim to calibrate against experiment and will do a larger case as a performance study as in the case with spheres.

What to measure?

The key metrics will be

  1. mixing index (same material, two colors)    Hongyang Cheng

  2. power (summation of the dashpot and friction dissipation energies)  torque arm Bruno Chareyre do you have a script ?  Code

  3. Torque Code

  4. profile outline .Code

To generate the above metrics, we aim to get post-process analysis scripts that people already have to compare between codes and to experiment. 

What we require

  1. Experiment and simulation data along with the configuration usedBase Case Anthony Thornton Drum.

  2. The code that you want to use and/or if you are willing to run cases. 

  3. If you can help with scripts for processing the results or have suggestions.

What we need to deliver

  • Officially we need to have at-least two open source codes with the implementations

    • Currently we have commitments for Yade and Mercury. 

  • Beyond that, any code that you want can be used as we aim to compare in a code agnostic manner.

What we want to achieve in 2025

  • Overview and comparison of different time-efficient computational strategies and methods (HPC vs Upscaling vs Surrogates)

    • Start with DEM glass beads

  • Comparison with real world ground truth

Hackathon plan

Presenting the WG1 Update

Identify available rotating drum datasets

Execution

  • Split into two groups HPC and Alternative Methods (AI/ML, Surrogate models)

    • G1 - HPC by nicolin

      • Assign tasks to complete the benchmark case of the rotating drum

    • G2 - Alternative methods by Deepak Tunuguntla Hongyang Cheng

      • Talk by Timo Plath on Reduced Order DPM

      • Setup a central github repo for all the relevant ML-based algorithms for granular materials

      • To be cont.

Additional references

Different calibration techniques

Yannick Descantes has a test case that we will use to check numerical stability in heap formation.

Timo Plath has a script in MercuryDPM for Rotating drum and Surrogate model calibration

Summary and action points

  • For this high performance comparison study, we will start with DEM using Silbert's glass spheres validated contact parameters for Linear Spring and Hertz-Mindlin

    • See Silbert’s paper and add a table of contact parameters here

    • Mono- and bi-disperse

  • For speed test

    • Turn on and off energy computations

    • Turn on and off data saving

      • How long do you want the run the simulation? x seconds

      • How often do we want to save data? y times

  • For surrogates

    • Generate our own datasets

Action points

  • Drum size, speed,

  • DEM parameters from Silbert’s paper

  • How do you generate the number of particles?, particle-wall friction,

  •  Hardware