Jean-Philippe Paiement

Chief Technology Officer, VRIFY

Talk Abstract

AI has not failed mineral exploration. Four assumptions have.

Drawing on 25 years of exploration experience and work with more than 200 mining companies, this talk takes apart the premises that quietly shape how the industry builds and buys AI for discovery: that the unlock is always new data, that the highest accuracy score wins, that the remaining deposits must be hiding somewhere new, and that a model's job is to hand you the answer. When the premise is wrong, no amount of data saves you. You solve the wrong problem faster.

Each fallacy is examined against real results. Data augmentation turns a handful of raw inputs into dozens of model-ready layers without a new survey. Random train/test splits leak across mineralized zones and inflate accuracy, while a 15:1 class imbalance lets a model reach 90% by learning to say no; holding out entire zones and reading minority-class F1, uncertainty and feature importance gives an honest number instead of a flattering one. Three generations of deposit-specific foundation models show validation accuracy and F1 rising on 10 of 11 deposit types as the shared training corpus grows, which points at a collective problem rather than an individual one. And 56% of new resource announcements between 2020 and 2024 came from existing projects, so the frontier is the ground already explored, not the seafloor. The four fallacies converge on a single mistake: building oracles when exploration needs instruments that produce reasoning. The output of a good exploration AI is not a target. It is an argument, and it should be measured on defensibility, auditability and iteration speed rather than on a score.

Speaker Bio

Jean-Philippe is a seasoned geostatistician with a strong background in structural, geological, and geochemical modeling and interpretation. His expertise extends to artificial intelligence, nonlinear interpolation, and simulation techniques. Throughout his career, Jean-Philippe has developed innovative workflows and approaches to minimize interpretational risks associated with geological data. With a wealth of experience in mineral resource estimation, he has successfully applied his skills to diverse geological environments worldwide, covering precious metals, base metals, and industrial minerals. In 2016, Jean-Philippe made a groundbreaking contribution to the field by pioneering the application of Machine Learning to mineral exploration. His achievements were recognized when he won the Integra GoldRush challenge, showcasing the potential of machine learning in mineral deposit targeting. Since then, he has remained at the forefront of advancing the application of machine learning to address geological and geophysical challenges, combining deep geological knowledge with supervised and deep learning techniques.