2019 April 29¶

Location: Coors Tech¶

Attending:¶

Jihyun Andy Thomas

Key takeaways¶

Preparing a paper on training a generator on well log data
Possible participants: Andy, Thomas, Jihyun, Bin, Bane
Well-defined scope, open access publication

Problems¶

Data needs to be better cleaned
Identify other potential well data types besides resistivity
Check the units of the well data

Comments¶

Scope of paper
- Stick to CNN-based GAN
- Multi-dimensional training data
Motivation
- lack of geosci training data
- well logs are simple and high fidelity
  - simple geometry
  - ubiquitous
    - oil,water,mining
- generator
  - no petro-phys model/simulator
  - act as a prior for inversion
    - mention gradients?
  - links different types of logs
  - well log data recovery
  - unsupervised
Outline of paper
Intro
- Geology of Kansas
- GAN
- Lit review
  - Background on Kansas geology
Methods
- experiment/ data
- QC/data cleaning
- GAN
- Cross-validation
  - spatial stats
  - wavelet transform
  - Fourier transform
  - PCA
Results
- what the clean data looks like
- short
- no interp
- speed
  - training
  - generating
- statistics
  - cross-validation based on location
- visualizing well log plots
Discussion
- each result item
  - compare speed vs other methods
  - What did/didn’t the network learn
- speculation
- recommendations
- predictions/future work
Contributions of paper
- published curated dataset
- published trained network
- Discussion
Task breakdown
- Thomas: statistical analysis
- Jihyun: Gan literature review
- Andy: Data cleaning
  - may need help
  - Download zipped las files from KGS (link on mlgp website)
  - check units: depth and Ohm-m
  - Do our best to handle well geometry
  - ID Nans
  - Handle NaNs
  - ID most popular aliases
  - write functions in .py file
  - Function 1
    - takes: a well log, mneumonic
    - returns mask for nan, depth
  - Function 2
    - takes: data, mask, depths
    - returns: ‘clean’ data
  - Interpolate over no more than 10 ft
Split paper off from ML group?
- We’ll see what interests people have
Hani’s like a Pokemon, We gotta catch em.

Agenda for next week¶

Clean data
Set up communication channels for collaborating over summer
Go get a burger