June 2019 – Technical Talk

BCGS Technical Talk – June 20, 2019

Speaker: Mike McMillan, Computational Geosciences

Title: Machine Learning in the Natural Resources Industry – examples and applications for mining, oil and gas and water exploration

Date/Time: Thursday, June 20, 2019 @ 4:30pm PST

Location: 4th Floor Conference Room, Room 451, 409 Granville St. (UK Building at Granville and Hastings), Vancouver

Abstract:

Machine learning and artificial intelligence (AI) are all the rage in 2019, and we take a look at recent applications of AI across a wide spectrum of problems in the natural resource industry. These vary from mineral and water prospectivity mapping to airborne induced-polarization detection, to borehole classification and seismic horizon picking. These new developments use deep convolutional neural networks to train the computer to detect subtle patterns across many geoscience data layers in order to help (and importantly not to replace) the geoscientist. Given a data-rich region with many overlapping geoscience data layers, the key is to come up with a well-defined question with examples of training labels. In the training labels we want both positive outcomes (ie. what you’re actually looking for) and negative outcomes (ie. what you’re definitely not looking for). These labels can be anything from gold-assays, rock types, fault types to salinity values in water. The type of label doesn’t really matter as long as it represents the thing or things you’re trying to find (or not trying to find). The areas without labels are the unknown regions that the machine learning algorithm will try to predict on based on signatures it learns from the training labels. In some settings this may be predicting gold grades in un-drilled areas, it may be predicting which electromagnetic decays have induced-polarization responses, or it may be predicting the likelihood of a water aquifer occurring in a remote region of the desert. The convenient aspect of these convolutional neural networks is that the algorithm architecture can be used to answer a multitude of questions, depending on the input training labels. This means that unlike geophysical inversion where a completely different code is required for magnetics, gravity or electromagnetics, we can generally use one AI framework, with some minor tweaks, for every problem. In this manner, we can throw in all available data sets and collectively use this information to help answer relevant questions that will help drive a data-driven cost-effective exploration program.

Leave a Reply