Guest Series - 01 - AI in Geosciences by Sacha Roslin
I started my career in geology the old-fashioned way: in the field. Straight after university, I worked as a field engineer, running geophysical logging while drilling in boreholes. Later, I moved into the office, where I began using the very same data — wellbore logs — to perform geosteering — changing well trajectory in real time. That transition from acquiring data to interpreting it sparked something in me. I became increasingly interested in data analysis, especially the kind that goes beyond spreadsheets and moves into the realm of patterns, complexity, and prediction.
That curiosity eventually led me to machine learning, and now, to writing deep learning algorithms that process unstructured rock images and geological datasets. I work in this space not because it is fashionable, but because I genuinely love it. There is something incredibly satisfying about watching an algorithm learn, adapt, and reveal something meaningful from what seemed like chaos.
When I deliver the presentations, people often ask: Aren’t you afraid that AI might replace you? You are probably teaching your own competitor! My answer is always — no, I’m not afraid current development of AI algorithms and here is why.
Supervised vs. Unsupervised Learning
Most of the popular AI applications in geology today fall under supervised learning. This is the branch of machine learning where we train the algorithm using labelled data. You give it thousands of examples — "this is sandstone", "this is shale" — and the model learns to recognize them. It is fast, powerful, and incredibly efficient. Tasks like mineral identification, geological mapping, core description and labelling can now be performed by AI algorithms, often faster and more consistently than humans. I worked on mineral identification of copper minerals in cores, and I can assure that 98-99% accuracy is achievable. These days there are different software packages which successfully implemented deep learning techniques for geoscience applications. My opinion is that the question of replacing personnel to perform geological tasks which can be fulfilled by supervised machine learning algorithms is just a question of time and economic profitability.
3D video of rock samples (copper porphyry rocks) segmented by U-Net convolutional neural network (supervised learning). Blue - rock pores, green - pyrite, red - chalcopyrite.
But there is also another family of AI algorithms — the one which is less popular, less headline-grabbing, but deeply important in complex geology. This group of algorithms is known as unsupervised learning. Unsupervised algorithms do not rely on labelled data. Instead, they detect patterns, clusters, and structures in raw, unlabelled datasets. In geology, this can mean uncovering new facies groups, subtle textural trends, or hidden relationships in multi-dimensional data.
It is not always easy to validate these outcomes, which is why unsupervised learning is underused in commercial geology. But for those of us who understand the rocks, the formations, and the physical context — these algorithms are like explorers. They do not give you direct answers but help you to find the patterns and group the data for future interpretation or find the anomalies for resource exploration. Anomaly detection is especially powerful in geology where anomalies may represent ore zones, unexpected textures, or rare geochemical signatures. These are not always labelled in advance — we must let the data speak first. I personally believe the future lies in unsupervised learning. The ability to detect structure and patterns across millions of data points could allow us to reassess decades of geological data in just days.
But writing these algorithms — shaping them with insight and purpose and adapting them to specific geological environments — will remain the task of humans. Artificial intelligence, as it stands today, cannot replicate human creativity, critical thinking, or intuition — and in my view, it will not be able to for at least the next 5 to 10 years. Machines may assist, but they will never replace a geologist with unique expertise and vision.
As for now — and for the foreseeable future — the best way to stay competitive is to develop our coding skills. Understanding how these algorithms work is not just useful — it is essential if we want to drive innovation, not just watch it happen. Rapid development of AI algorithms is inevitable, and the recent evolution of generative models clearly demonstrates it. It is like a wave – and only we can decide if we will be removed by this wave and ride it.
https://www.linkedin.com/in/alexandra-roslin-403408240/
