Content Comparison

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

  • DEM-High Performance Computing (HPC), led by nicolin

  • DEM-UP (Upscaling), led by Hongyang Cheng

  • DEM-ML (Machine Learning), led by Deepak Tunuguntla

• 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 ON-DEM wiki so each member can upload papers that might be of interest for the review to avoid duplication and improve efficiency.