Computers can learn geology – but must handle uncertainty

How rock types are logged in drill cores influences decisions in exploration and mining. At the same time, geological complexity means that different geologists can interpret rocks differently leading to inconsistencies. In his doctoral thesis, Filip Simán investigates how machine learning can be used to gain deeper geological understanding from drill core data and create more consistent drill core logs – but also why it’s not enough to leave the interpretations to models.

“Machine learning can provide a more consistent basis for geological interpretations, but the results must be transparent and explainable to the geologist,” says Filip Simán, PhD in Ore Geology at Luleå University of Technology.

Combining AI and geological expertise

In the thesis Teaching computers geology, he combines geological knowledge with machine learning models to classify rock types from drill core data. Different kinds of machine learning models were applied to data from Bolidens Rävliden North deposit.

The results show that the models can learn to identify patterns in the data, but their ability to generalise between different drill holes is limited.

“The models perform better when trained and tested within the same dataset, but their reliability decreases when applied to new geological contexts,” he says.

This means that AI-based tools cannot be used uncritically; they must be developed with consideration for how generalisable a project aims for the model to be.

Understanding and quantifying uncertainty

A key contribution of the thesis is the explicit treatment of uncertainty. When models are used in practice, geologists need to understand not only the prediction, but also how reliable it is.

Filip Simán utilises methods to quantify uncertainty in both geochemical calculations and machine learning predictions.

“It is not enough for a model to provide an answer. We also need to understand how uncertain that answer is,” he says.

He also shows that certain rock types, such as chemically altered ones, are harder for models to handle than simpler, non-altered classifications.

AI as decision support – not replacement

To improve usability, the thesis explores how model predictions can be explained. Using an explainable AI method called SHAP, the study identifies which variables influence model decisions. The method enables the AI model to explain which input variables are most influential. For example, which chemical elements contribute most to the classification of a rock type.

This reveals that the models often use geologically meaningful indicators, but they can also learn indirect relationships that may be ineffective for recognising specific rock types.

“The aim is not to replace the geologist, but to develop tools that can support and enhance decision-making. AI will not solve everything, but it can become a valuable partner for the geologist. AI can be used in geological analysis, but it requires an understanding of and control over uncertainties in order to be useful in practice,” says Filip Simán.

The doctoral thesis is the result of interdisciplinary collaboration between Ore geology and Machine learning at Luleå University of Technology, and Boliden.