/ML_for_physicist

Repository for Machine Learning for Physicist 2020 summer training program

Primary LanguageJupyter Notebook

ML for physicist banner

Machine Learning for Physicist 2020 Summer Course

Instructor: Prof. Florian Marquardt
Institution: Max Planck Institute for the Science of Light, Erlangen, Germany

Links

My Mini Project

Access to my notebook here

neural network eor screening

At the end of this summer training program, I made a mini-project. I name this mini-project: Neural Network Application for Enhanced Oil Recovery (EOR) Screening. EOR screening is part of important practices in the oil and gas industry. Main reference of this project is EOR screening metric proposed by Taber et al (1997) in their SPE-35385-PA paper.

The neural graph above is created in NN SVG website. Thanks to @alexlenail.

Some key facts of this project are:

  • I generated datasets using random technique, 245 dummy field observations, each containing 6 reservoir fluid and rock parameters. These datasets are the my train datasets.
  • These 6 parameters are the input features for the NN.
  • The datasets are labeled with each of 7 EOR methods.
  • I generated as well, another 50 dummy field observations, without label. These datasets are the my test datasets, for prediction.
  • In the first trial, without hyperparameter tuning, 1 hidden layer with 100 hidden neurons is used.

Find the train CSV data here and the test data here

What can still be improved?

  • To increase accuracy of prediction, the other 3 categorical parameters that were not used before, could be included. The thing needed is to encode categorical parameters.
  • Real datasets can be obtained. This paper by Lake & Walsh (UT Austin, 2008) listed all publications about EOR, and this XLSX file by World Economic Outlook (2018) listed as much as 233 EOR fields.
  • This real dataset will be released soon.