Deep learning is now pervasive in our daily lives, and industry is investing heavily to digitalize workflows and implement AI solutions. Yet, its adoption in the workplace - particularly in exploration and extractive industries - remains limited, and there are few established pathways for successful integration.
In this lecture, I explore the question of the relevance of deep learning by presenting highlights from my group’s multidisciplinary research, which fuses deep learning (primarily computer vision) with geosciences. We begin at planetary scale, combining satellite imagery with AI to interpret surface processes. Our work on Mars hyperspectral data led to a novel hybrid neural network that classifies surface mineralogy using CRISM reflectance spectra, achieving state-of-the-art performance across 39 mineral classes. The model - built using a variational autoencoder and calibrated using the Expected Cost metric adapted from medical AI -delivers denoised spectral reconstructions, supports better interpretability, and generalizes robustly to unseen Martian terrains. These methods can be translated to Earth-based remote sensing for critical mineral exploration, bridging planetary science and terrestrial georesource applications.
The focus of most of our work though is in the subsurface, where we developed SRT-Ai (Seismic Reflection Terminations Attribute), a deep learning framework trained on 160,000 synthetic seismic data to detect seismic reflection terminations - features crucial for stratigraphic interpretation but historically mapped manually. SRT-Ai achieves 99.9% accuracy on synthetic data and 91% accuracy on real seismic imagery, enabling semi-automated reconnaissance of termination using probabilities and reducing interpretation uncertainty. This complements our ongoing work in core and log image analysis, contributing to the broader application of AI in reservoir characterization.
Finally, in an ongoing industry-funded project, we apply deep learning to forward stratigraphic modeling (FSM), using GANs to simulate realistic carbonate facies and exploring Physics-Informed Machine Learning (PIML) to incorporate geological constraints into predictive models. We test data assimilation strategies - including model-to-model and Kalman filter-based approaches - and investigate surrogate models to accelerate simulation. Preliminary results suggest these approaches can significantly improve the realism, precision, and scalability of FSM outputs.
So, is deep learning the next frontier in exploration geology?
The answer is: increasingly, yes. Deep learning is no longer a speculative tool - it is a maturing technology capable of addressing long-standing geological challenges across scales. While domain expertise remains essential, AI offers a transformative complement, accelerating discovery, reducing uncertainty, and unlocking insight from ever-expanding datasets.