Helmholtz AI Roadshow Research Field Energy

Nov 12, 2020; 9am-3.30pm
Event page: https://hifis-events.hzdr.de/event/33/

To connect to the consultant teams:

• Markus Götz et al @ KIT: consultant-helmholtz.ai@kit.edu
• Peter Steinbach et al @ HZDR: consultant-helmholtz.ai@hzdr.de

Connection Details

Roadshow Helmholtz AI for Energy
Time: Nov 12, 2020 09:00 AM Amsterdam, Berlin, Rome, Stockholm, Vienna

Join Zoom Meeting
https://us02web.zoom.us/j/83502005855?pwd=VHJRM3FCYVNjOTZ4aXVYeDRXVjRGQT09

Meeting ID: 835 0200 5855
Passcode: 249105
One tap mobile
+496938079883,83502005855#,0#,249105# Germany
+496950502596,83502005855#,0#,249105# Germany

Dial by your location
+49 69 3807 9883 Germany
+49 695 050 2596 Germany
+49 69 7104 9922 Germany
+49 30 5679 5800 Germany
+1 301 715 8592 US (Germantown)
+1 312 626 6799 US (Chicago)
+1 346 248 7799 US (Houston)
+1 646 558 8656 US (New York)
+1 669 900 9128 US (San Jose)
+1 253 215 8782 US (Tacoma)
Meeting ID: 835 0200 5855
Passcode: 249105
Find your local number: https://us02web.zoom.us/u/kRYXeBeXW

Agenda

Notes

Feel free to add notes here that surface during the discussion and are not contained in the slidedeck(s).

General Intro by Markus Götz

• Invite Link for the Helmholtz AI Projects “market place”: https://join.slack.com/t/helmholtzaiprojects/shared_invite/zt-i9cdle4m-VP655ozO92WmmyFXNSyy1Q
• DFG guidelines for good scientific practice: https://www.dfg.de/download/pdf/foerderung/rechtliche_rahmenbedingungen/gute_wissenschaftliche_praxis/kodex_gwp_en.pdf

Helmholtz AI YIG for Energy by Ralf Mikut and Benedikt Heidrich

• pyWATTS: https://github.com/KIT-IAI/pyWATTS
• Q&A:
  • FS: are the AI activities at KIT limited to time series and energy systems?
    • RM: no, time series are the focus of my group and to some extent for the YIG; the consultants have a more broad focus
    • RM: if vouchers fall in the domain of the YIG, the YIG is happy to collaborate/contribute
    • MG: consultants can help with more numerically inclined tasks too
  • RG: it might be helpful to collect all demands from the field of energy, presentation was KIT specific?
    • MG: project proposal to Helmholtz AI was reviewed externally, energy team was attributed to KIT
      • consultants are open for any question from field of energy
      • consultants team is very diverse
      • YIG focus is more specific
    • RM: please look for vouchers and bring your topic to the voucher system
      • for larger projects (1-2 years of work) -> please apply for a Helmholtz AI projects
  • RG: audience for vouchers? reach out to groups that do not apply ML yet?
    • MG: vouchers are support instruments (no obligation for advanced groups)
    • MG: to support science, maybe Helmholtz AI projects are the better tool
    • MG: all groups/institutes have been approached
    • RG:
      • will spread the word among colleagues
      • suggestion: maybe showcase the tools and stay more high level (explain the purpose and benefit from ML)
• Peter Steinbach (PS) introduces the research field Matter
  • Clarifies whom to talk to, orient yourself in line with your research field
  • Preliminary talks at HZDR possible if voucher is being considered

1on1 session with Fabian Schlegel et al

• Jan Schäfer
  • challenge: segment (overlapping) bubbles in 2D images
    • PS: proposed in earlier talks StarDist (https://github.com/mpicbg-csbd/stardist)
    • Charlie Debus (CD): improve labelling? nnUnet?
    • MG: try GANs to emulate simulation
    • MG: self-supervision?
      • https://arxiv.org/pdf/1805.01978.pdf (for classification/localisation variant available)
    • MG: size of data?
      • tomography: 1000 images in one stack summing to ~1GB
      • microscopes: ?
• Fabian Schlegel
  • Numerical simulations of multiphase flows
    • Build unified model for flow simulation from a large set of physical models
    • Predict model for certain multiphase problems
    • sweep parameters space and configs with snakemake and store in central DB
    • Strategy: collection of CFD grade -> data mine
    • Challenge: detection of overlapping bubbles
      • currently UNet variants (fully convolutional)
      • manually annotated dataset with inner and outer boundaries
      • planned: track particles in 3D volumes (from two 2D multiview)
      • build BubblyFlow: SQL database with experimental data
    • Voucher opportunities
      • Predicting occluded bubble parts
      • Combination of bubble segmentation datasets
      • Self-supervision/GANs for creating labels
  • Cases Repository
    • Compare baseline and simulation curves
    • Currently fuzzy-logic, MSE and Pearson coefficient
      • inplace of Pearson correlation coefficient, please consider the following as these are more robust with respect to non-linearity:
        • https://en.wikipedia.org/wiki/Kendall_rank_correlation_coefficient
        • https://en.wikipedia.org/wiki/Spearman's_rank_correlation_coefficient
      • overview of the limits/performance of several metrics: https://arxiv.org/abs/math/0209021
      • alternatively: https://de.wikipedia.org/wiki/Kullback-Leibler-Divergenz
    • Goal: get advise for where to investigate divergence between baseline and simulation

1on1 session with Richard Gloaguen et al

• Multi- and hyperspectral remote sensing
  • Mostly image-based classification/segmentation tasks
  • Resource characterization
  • Recently time-series approaches brought in
• Drill core scanning
  • Human bias with respect to operator
  • Supervised and unsupervised approaches
  • Tasks: upscaling of mineral maps
  • Challenges: feature selection and sensor/feature fusion
• Adapt code to run on A100s
• Hyperspectral point clouds 2D -> 3D
  • Classify points into ore categories
  • Reconstruct resource layers
• Upscaling for data fusion
  • Different scales of sensor resolution
  • Make them agreeable
• Data acquisition via UAVs
  • Plan flight paths
  • Informing surveyors while flying