Versions Compared

Version Old Version 3 New Version 4
Changes made by

Deepak Tunuguntla

Deepak Tunuguntla

Saved on

Key

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

...

Brief Summary

Points discussed

  • Started with the presentation by Igor Bersinskii on utilisation of Machine Learning for calibration of DEM simulations

  • After Igor’s presentation, discussed what is required to achieve WG1’s objectives. Based on the discussion, we all agreed upon dividing WG1 into three sub-groups. These being

  • Besides defining the sub-groups, we also brainstormed and listed key topics for each sub-group. These are

    • Time-scales → DEM-HPC

      • A review of the different algorithms in DEM codes:

        • Broad-phase contact detection (Gross)

        • Narrow-phase contact detection (Fine/detailed)

          • Contact Fidelity  

        • Interaction History

        • Time-Integration 

      • HPC Implementation

        • Software : OpenMP, MPI, CUDA, Kokkos,GPUFort

        • Hardware : GPU,CPU,TPU

        • Numerical Precision

        • Adapt to next gen (eq: fine contact).  

        • Make recommendations accounting for the fast evolving hardware and software

      • Benchmark cases

        • Silo discharge (geometry from benchmark paper but decreased particle size to fit 1M particles → monosized spheres, size distribution with range, other shapes?),

        • Rotating drum (geometry from benchmark paper with decreased particle size to fit 400k particles) → monosized, polysized, shape

        • Get input from WG4 (academic cases) and WG5 (industrial cases).

    • Space scale → DEM-UP

      • Review capability of continuum/process-based models (e.g., MPM, SPH, FEM) for granular processes from quasi-static to rapid flow

      • Identify (common) spatial/temporal mapping schemes, e.g., averaging, coarse-graining, for these processes

      • Critically review existing coupling algorithms for DEM-UP and their limitations

      • Collaborate with WG2 and WG4

    • DEM-ML

      • Review available ML methodologies to build reduced-order, generative DEM models, e.g., learning spatial and temporal features (micro to macro) and microstructure

        • Classical ML or Deep Learning

          • PINNs, GNNs, Neural Operator Learning, e.g., UPTs

        • How is ML utilized currently in the community generally?

          • Calibration

          • Characterisation

          • …

      • Collect simulation data from well-established benchmark cases (e.g., triaxial in Geotech, ring shear in particle technology) in a format common to many DEM codes

        • Collaboration with WG4

      • Best practices for ML for Granular materials

        • Put together training materials for ML enthusiasts in granular matter community

      • ML in Uncertainity Quantification

        • Bayesian ML

  • Action points

    • Ask members of the WG to choose a sub-group to join.

    • Each member of the WG can suggest literature to achieve the objectives of the literature review for each sub-group.

    • Consider implementing a mechanism on the OnON-DEM wiki so each member can upload papers that might be of interest for the review to avoid duplication and improve efficiency.

Important Links