loading page

Engineer Level Automated Interpretation of Geothermal Well Logs Using Convolutional Neural Networks
  • +1
  • Esuru Okoroafor,
  • Ahinoam Pollack,
  • John Murphy,
  • Roland Horne
Esuru Okoroafor
Stanford University

Corresponding Author:esurunma@gmail.com

Author Profile
Ahinoam Pollack
Walmart Labs
Author Profile
John Murphy
Ormat Technologies
Author Profile
Roland Horne
Stanford University
Author Profile


Geothermal well log analyses consist of utilizing pressure and temperature data measured along the wellbore to predict feed zones, reservoir temperature, and reservoir pressure. Interpreting geothermal production logs can be subjective and require great expertise to achieve repeatability. In situations where there may be several log data, interpreting the logs may be time consuming for quick decision-making processes. This work discusses the implementation of a multi-layered deep learning convolutional neural network to automatically diagnose sets of temperature and pressure well logs. The algorithm achieves results similar to those of a professional engineer. This algorithm enables the interpretation of many well logs in just a few seconds. Data input for this project is synthetic well data that mimics real data. 10,000 datasets were used. The data was split as follows: 9,800 and 200 for training and validation respectively. The algorithm used takes as input three “depth-series” logs of temperature, pressure, and temperature gradient, passes the data through a convolutional neural network including a flat layer and then a fully connected layer with five output variables which are the depths of the feed zones, the reservoir temperature, the reservoir pressure and the depth at which the reservoir pressure is known. The cost function for this model was the mean squared error. The optimizer algorithm used was Adam, and the learning rate had an exponential decay. The algorithm recorded the model state that had the lowest mean absolute validation error. The architecture was implemented in Keras with a TensorFlow backend. The best model found during the process of hyper-parameter tuning was used to predict the reservoir characteristics for the validation and testing data sets. The results show a good match between the predicted data and actual data with a training error of 0.9%, validation error of 2%, and test error of 7%. Future works will involve adding more real data to the training and validation set, increasing the number of feedzones that can be identified, and performing sequential analysis using interdisciplinary data.