• Labeling—or segmentation—of rock core computed tomography (CT) allows for micro- to nanoscale quantitative analysis of rock properties using computer simulations
  • Machine learning (ML) methods for segmentation will provide substantial benefits over traditional manual segmentation:
    • Non-biased results
    • Reduction of time to result of 3-4 orders of magnitude (several hours to seconds)
    • Accuracy approaching lab-measured properties
  • Selecting the optimal ML method for CT segmentation is vital, but the choice is not straightforward

  • A high degree of precision was found when both supervised and unsupervised segmentation results were compared to ground truth (90% and 95-97% respectively).
  • Unsupervised segmentation required no prior information (labeled images).
  • While training the ML algorithms took several hours, results were produced in ~5-10 seconds and the training is reusable.
  • Preliminary results are encouraging, but further testing on additional datasets and with a diverse set of samples is needed.

  • Porosity for unsupervised and supervised learning approximated the measured value, but further refinement is needed
  • Greatest differences in segmentations can be found at the interface and within pore throats