Scaling up well log interpretation for groundwater salinity mapping in the age of Big Data

Abstract

The requirement to identify Underground Sources of Drinking Water under California Senate Bill 4 has the dimensions of a Big Data problem, especially as the public has access to vast amounts of relevant data in the form of petrophysical logs from a century of oil development. However, few data scientists have the domain knowledge and areal expertise to interpret these data. To promote conversation between hydrogeologists and data scientists, we created a simple, extensible computational model for well log interpretation that returns the groundwater salinity, lithology, and other geologic parameters that best explain the resistivity, porosity, SP, and GR measurements for a particular well. We are aware that there have been computational algorithms for well log analysis for almost as long as the field has existed. Our approach has been to design a model that can be put into the hands of experts in multiple fields. It is implemented using Tensorflow, which is an open-source software released by Google in November, 2015 that has been rapidly and widely adopted by machine learning researchers. The model takes the form of about 100 lines of Python code that contain standard empirical equations (such as Archie's equation) for deriving log data from well parameters. Tensorflow solves the inverse problem by automatic differentiation of the equations and gradient descent in the space of parameters. The analysis results are a plot of the full lithology and depth series for petrophysical parameters, which lets domain experts see, at a glance, if the model is behaving correctly. We have applied this model to wells from several California oil fields, including Lost Hills, Midway Sunset, and South Belridge. We are in the process of validating results by comparing them to core analyses and direct geochemical measurements.

AAPG Datapages/Search and Discovery Article #90266 © 2016 AAPG Pacific Section and Rocky Mountain Section Joint Meeting, Las Vegas, Nevada, October 2-5, 2016