Deep Learning Geological Interpretation Using Geophysics

For Round 5 of the Queensland Collaborative Exploration Initiative (CEI), Caldera Analytics was engaged by Strategic Energy Resources (SER) to develop a deep learning model that performed geological interpretation of basement geology using geophysics. SER have two main projects in the Mt Isa region that are both located under significant amounts of younger cover, making geological interpretation a difficult task. A well trained deep learning model would give probabilistic interpretations along with quantifiable uncertainty, which would be a valuable addition to the limited decision making tools available to mineral explorers targeting economic deposits under cover. The machine learning model, which used gravity and magnetics as an input, was trained to predict six key geological groups, one of which was hydrothermal magnetite, a direct vector for magnetite iron-oxide-copper-gold (IOCG) deposits. A ResNet18 convolutional neural network model architecture was used to make predictions, with the MoCoV2 self-supervised learning technique used to warm up the model on the geophysics of Australia. This allowed the model to be fine-tuned on a small highly curated training dataset built from historical drilling basement intersections. The trained deep learning models when applied to the SER projects showed great promise as a valuable decision making tool for mineral explorers in areas under cover. The predictions of the six lithology groups align with the expected geophysical response of each group, and the ability to see the probability of each interpreted lithology allows explorers to be wary of uncertain areas that may need additional geophysical surveying to increase the confidence of the region. These predictions can be combined with the traditional decision making process for under cover exploration, further reducing risk and leading to better outcomes for mineral exploration.