Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Introduction to Machine Learning techniques for Geoscience

Introduction to Machine Learning techniques for Geoscience

Jesper Dramsch

October 14, 2020
Tweet

More Decks by Jesper Dramsch

Other Decks in Technology

Transcript

  1. Machine Learning Techniques
    for Geoscience Applications
    Jesper Dramsch

    View full-size slide

  2. Where are we
    going today?
    ● My Background
    ● What’s Machine Learning?
    ● Geo to ML?!
    ● What’s important in ML?

    View full-size slide

  3. What is Machine Learning?
    And why should I care?

    View full-size slide

  4. if x > 5
    Classical rule-based and expert systems

    View full-size slide

  5. Computers
    figure it out
    based on the data

    View full-size slide

  6. Tossing Bot!
    (http://tossingbot.cs.princeton.edu/)

    View full-size slide

  7. Models that do not have external information
    Blackbox models

    View full-size slide

  8. What is AI, ML, and Deep Learning?

    View full-size slide

  9. Style Transfer for Multiple Tasks

    View full-size slide

  10. Everybody dance now!
    (https://arxiv.org/pdf/1808.07371.pdf)

    View full-size slide

  11. What is the connection
    Why do we care in geoscience?!

    View full-size slide

  12. Take seismic data

    View full-size slide

  13. Get some “interpretation”

    View full-size slide

  14. Or get some “inversion”

    View full-size slide

  15. Neural Networks
    The new hype that is older than all of us

    View full-size slide

  16. A Simple Neuron
    1
    1
    + 2

    View full-size slide

  17. A Small Neural Network

    View full-size slide

  18. A Deep Neural Network

    View full-size slide

  19. Different Combinations In Neural Networks

    View full-size slide

  20. Is that all?
    What else is there to make Artificial Intelligence?!

    View full-size slide

  21. Convolutional Neural Networks
    The machine learning that sees

    View full-size slide

  22. A neural network as it is used today

    View full-size slide

  23. The learned filters with different abstractions

    View full-size slide

  24. A Geoscience Application

    View full-size slide

  25. Matching of seismic data

    View full-size slide

  26. The differences between unmatched and matched

    View full-size slide

  27. Compared to a baseline method

    View full-size slide

  28. Take it to the next level

    View full-size slide

  29. Real Data (CC-BY A. Çuğun)
    30
    How do we generate data like this with a Neural Network?

    View full-size slide

  30. Real Data (CC-BY A. Çuğun)
    Network
    Data
    Legend
    31
    Generative Adversarial Networks

    View full-size slide

  31. Latent Space (CC-BY Hudson)
    Real Data (CC-BY A. Çuğun)
    Network
    Data
    Legend
    32
    Start from latent space

    View full-size slide

  32. Latent Space (CC-BY Hudson)
    Generator (CC-BY-SA T. Vaughn)
    Real Data (CC-BY A. Çuğun)
    Network
    Data
    Legend
    33
    Feed it to a Generator

    View full-size slide

  33. Latent Space (CC-BY Hudson)
    Generator (CC-BY-SA T. Vaughn)
    Fake Data (CC-BY Tony A.)
    Real Data (CC-BY A. Çuğun)
    Network
    Data
    Legend
    34
    That generates some Fake Data

    View full-size slide

  34. Latent Space (CC-BY Hudson)
    Generator (CC-BY-SA T. Vaughn)
    Fake Data (CC-BY Tony A.)
    Real Data (CC-BY A. Çuğun)
    Discriminator(CC-BY Brickset)
    Network
    Data
    Legend
    35
    Discriminator judges if sample is real

    View full-size slide

  35. Latent Space (CC-BY Hudson)
    Generator (CC-BY-SA T. Vaughn)
    Fake Data (CC-BY Tony A.)
    Real Data (CC-BY A. Çuğun)
    Loss (CC-BY S. MacEntee)
    Discriminator(CC-BY Brickset)
    Network
    Data
    Legend (CC-BY Dramsch)
    36
    Both networks learn regardless if D is correct

    View full-size slide

  36. What’s the application?

    View full-size slide

  37. Some seismic data

    View full-size slide

  38. An “inversion”
    Generator (CC-BY-SA T. Vaughn)
    generated by a neural network

    View full-size slide

  39. ● Generating Synthetic Data
    ● Earthquake localisation
    ● Drill Core Processing
    ● Clustering of Well Data
    ● Prediction of Volcanic activity
    ● Satellite Data
    What else is
    possible?

    View full-size slide

  40. Many Other Machine Learning Algorithms

    View full-size slide

  41. Seismogram Tagging
    (https://www.nature.com/articles/s41467-020-17591-w)

    View full-size slide

  42. Core Image Segmentation
    (https://joss.theoj.org/papers/10.21105/joss.01969)

    View full-size slide

  43. Well Log Fascies Prediction
    (https://doi.org/10.1190/tle35100906.1)

    View full-size slide

  44. Remote Sensing Data
    (https://doi.org/10.1117/1.JRS.11.036028)

    View full-size slide

  45. Some Tips to Get Started?

    View full-size slide

  46. Start Projects
    Application
    as important as
    Theory
    ● Participate in Hackathons
    ● Build Apps
    ● Start Small
    ● Compete on Kaggle
    ● Get familiar with the Math

    View full-size slide

  47. 70 Years of Machine Learning
    in Geoscience
    https://authors.elsevier.com/b/1bpiNEroEut0M

    View full-size slide

  48. Available Data
    For Projects and Experiments
    ● Open Data:
    https://wiki.seg.org/wiki/Open_d
    ata
    ● Full Reservoir with Wells:
    https://wiki.seg.org/wiki/The_No
    rth_Sea_Volve_Data_Village
    ● Seismic Interpretation:
    https://github.com/yalaudah/faci
    es_classification_benchmark
    ● Salt Identification in Seismic:
    http://bit.ly/kaggle-salt

    View full-size slide