Rapidly inverting airborne electromagnetic data using Neural Networks

A collaborative research team from CGI, Invert Geophysics, and the University of British Columbia has unveiled a pioneering method for inverting airborne electromagnetic (AEM) data using machine learning. Traditionally, interpreting the electrical conductivity of the earth from aerial surveys has required massive computational power to solve complex physics equations. By training a deep neural network to recognize the relationship between survey data and subsurface conductivity, CGI and its partners have effectively bypassed the most "costly" part of the process.

To prove the concept, the team utilized H3DTD to generate 10,000 synthetic training models and then tested the network on a Skytem field dataset from the Kaweah sub-basin in California. The results were remarkable: once trained, the network produced realistic 2D conductivity models in just seconds on a standard laptop. This capability is a game-changer for groundwater management and mineral exploration, as it allows for near real-time interpretation of field data.

CGI provided critical value to this study by leveraging our proprietary expertise in 3D forward modeling and geostatistical tools like gstools to build the high-fidelity training datasets required for the network to generalize. This research serves as a successful proof of concept that paves the way for even more ambitious 3D machine learning inversions in the near future.

‍