Tulsa SPWLA
Monthly Luncheon Meeting
Thursday Sept 12 2019
Tulsa University
Helmerich Hall- Room 121
800 S Tucker Dr.
Tulsa, OK 74104
11:30 – 1:30 pm
Register and pay online or pay cash/check at the door.
RSVP: [email protected]
Cost - $25 for Professionals and FREE to students with student ID
Machine-Learning
Methods: Analysis of Rock Images and Beyond
Presented By: James J. Howard, DigiM Solutions
ABSTRACT: Big-Data,
Data-Driven Analytics, Artificial Intelligence and Machine Learning are all
terms found in many recent presentations within the petrophysics community.
Often these terms are used interchangeably, but they are distinct differences
amongst them. The common thread with these terms is that the application of
these various techniques and terminologies is based on the large volumes of
data that confront petrophysicists in the day-to-day operations of large,
highly instrumented fields. One example of Machine-Learning (ML) methods is to
provide a more accurate and robust approach to the processing of images,
including high-resolution images of pore systems acquired with micro-CT and SEM
techniques.
Each of
these imaging methods has a range of resolutions and sample volumes that cover
several orders of magnitude. In an ideal situation the images are characterized
by a range of intensity values that when evaluated as a histogram there is a
clear distinction between grains and pores. The reality is that image quality
is often compromised by instrument noise, overlapping phases in a given voxel
and other factors that generate an intensity histogram that is less discrete
and more difficult to process. A machine-learning based segmentation tool
provides a more robust solution to the phase separation challenge by including
a wide range of statistical measures of each voxel in the image.
Along
with the basic image intensity value for each voxel, the accumulated statistics
include information on nearest-neighbor and next-nearest-neighbor properties
derived from a series of filters and gradient measurements. Since the ML-based
segmentation includes nearest-neighbor information, it can be used to
distinguish phases with similar intensity but distinctly different surface
textures as observed by the user. Segmentation of pore space is validated
through visual inspection followed by comparison of static properties, e.g.
porosity and pore-size distribution, and finally dynamic properties such as
permeability, capillary pressure, relative permeability and upscaling. These
algorithms and workflows can be applied to other large dataset such as image
logs, whole core CT and core photos.
BIO: James
Howard is a Technical Advisor to DigiM Solution, LLC, an Image-Based Rock
Physics software company, which is giving him a late-career opportunity to
explore a long-held interest (since the 80s) in “digital rocks”. In an earlier
life he worked on a range of tools used for pore-scale characterization of
reservoir rocks, including the development of NMR interpretation methods in the
lab and the field, the application of MRI and microCT imaging to monitor flow
experiments, mineral mapping by SEM-EDS, and other curious topics. He has B.S
and Ph.D. degrees in geology and geophysics.