2019 April 8

Location: Coors Tech

Attending:

Iga Jihyun Andy Antoine Bin Julian Bane Thomas

Key takeaways

  • Tensorflow officially suppports Keras
  • Tensorflow has new probablistic and open data distribuition tools
  • EDAA is a more formal group for us to potentially present/participate in
  • We now have nice, clean, open, RILD well log data from Kansas
  • Andy has impressive GAN puns.

Problems

  • Is 4000 wells enough to properly train a GAN?
  • Is Tensorflow’s official embedding of Keras better than Keras?
  • What are the unites of the well data?

Comments

  • Slides for TF updates are on Slack
  • Slides for EDAA are on Slack
  • Well log data for GAN (open data cleaned by Andy)
    • 4000 wells
    • 400 - 4000 ft
    • 1 ft spacing in depth
    • x-y coordinates of wells not in file, but available
    • Data is RILD (MT-like)
    • RILD ranges over several orders of magnitude
    • We should use log(RILD) as inputs
    • All well logs are from Kansas
    • “Let’s build the best GAN we can.” - Andy
    • “All we need is a good G.” - Andy
    • Downsampled by linear averaging every 60 ft.
    • 3600 samples per well (full resolution)
    • 60 samples per well (downsampled)
    • 100 MB file (full resolution link pending)
    • 2 MB file (downsampled on Slack)
  • Meeting didn’t go as planned timeline, too much in the meeting

Topics of discussion for next week

  • Breaking into three groups for pair programming of GAN
  • Two groups for network structure: Andy, Bin, Julian, Iga, Jihyun
  • One group for data viz: Thomas + Bane

Comments