Artificial intelligence explores the underground

Underground investigations are often time-consuming and costly. Yet without knowledge of the properties and characteristics of the layers located deep below the surface, many important questions cannot be answered: Can data for future explorations around the deep geological repository be predicted quickly and reliably? Is a particular underground site suitable for obtaining deep geothermal heat and power, or for extracting natural gas? Are the geological conditions at a depth of 1,500 metres suitable for storing carbon dioxide? To make it easier to answer these and other questions, Romana Boiger, from the Laboratory for Waste Management in the PSI Center for Nuclear Engineering and Sciences, is working to establish new tools from the area of artificial intelligence for geological investigations.

Boiger’s attention is focused on so-called artificial neural networks in particular. These consist of several layers of interconnected artificial neurons. These are, in the final analysis, mathematical formulas that process input data and deliver a result. What makes this special is that artificial neural networks are capable of learning. For example, an artificial neural network that is supposed to distinguish between apples and pears can be trained by presenting it with images of apples and pears and simultaneously providing the correct interpretation. After a certain number of training runs, the artificial neural network is then prepared to correctly classify even unfamiliar pictures of apples and pears.

In her research, Boiger, a mathematician with a focus on data science and machine learning, uses a special type of artificial neural networks called convolutional neural networks (CNNs). These are especially well suited to the identification and analysis of patterns and simple features in images.

Scientifically uncharted territory

One novel application of CNNs is the subject of the study Boiger and colleagues published in May 2024 in the Swiss Journal of Geosciences. It is the result of an interdisciplinary collaboration between scientists from PSI and experts in geology and engineering at Nagra. In a first step, they used CNNs to analyse images of drill cores taken from the Trüllikon borehole in Northern Switzerland. This was part of Nagra’s site investigation programme to identify a suitable site for a deep geological repository. The test interval was selected from 55 metres of drill core from a depth of between 770 and 939 metres. «We wanted to find out if it’s possible to accurately determine the lithological formations and above all the mineralogical composition of the rock – such as the proportions of calcite, clay, and silicates – solely on the basis of drill core images. » Studies already exist to investigate the lithology, determining properties that can be observed with the naked eye, without the help of a microscope. On the other hand, determining mineralogy in this way is scientifically uncharted territory. «No one had ever done it this way before.»

For her research, Boiger used artificial neural networks that had already been trained. They had previously learned to distinguish between images of vehicles, animals, people, and fruit – as well as geological formations and rocks – using images from the ImageNet database, a collection of more than 14 million images.

The CNN models thus already had a certain knowledge base when they were presented with the Trüllikon drill cores. The 10 cm thick drill cores from various geological units, known as formations, were systematically photographed after washing. The photographs were then cut into slices. Boiger and colleagues proceeded step by step: They expanded the pre-trained CNN by a few layers, which they then specifically trained to distinguish between lithological formations on the basis of the images. This resulted in a new, larger CNN model. It was then expanded again by a few layers – and finally trained to recognise the mineralogical composition.

Aiming for one-centimetre resolution

The trained artificial neural network that Boiger used fulfilled its task brilliantly and determined the lithological formations of the Trüllikon drill core with an accuracy of 96 percent. Its performance in determining the mineralogical composition solely from drill core images, i.e., the actual scientific novelty, was also excellent. The results of the neural network and the standard model were compared with those of independent laboratory measurements – and the quality was equally good in both cases. «Our method can help to make future geological explorations less expensive and at the same time more efficient, precise, and objective,» says Boiger, explaining the benefits. «This is true regardless of whether the objective is extended investigations on deep geological disposal, or assessing potential CO₂ storage sites, or doing preliminary studies for geothermal energy.»

The problem with previous methods: They not only cost a lot of time and money, but are also limited in terms of resolution. MultiMin, the standard method for determining mineralogy along drill cores, is one example. It combines three elements: measurements made during drilling, laboratory measurements, and statistical methods. Only the laboratory measurements provide exact values, but obtaining these is time-consuming and costly. As a result, it is not possible to simply measure every centimetre or every fifth centimetre of core in the laboratory. In practice, measurements are sometimes taken at intervals of ten or 20 centimetres, for example, and in other cases at intervals of several metres. The laboratory values that MultiMin relies on often have very large gaps.

Boiger wants to obtain a much more accurate picture of the subsurface using artificial neural networks. «We can train the network in such a way that in the end we achieve exactly the desired resolution, for example one centimetre. That means cutting the images of the drill core into one-centimetre-thin slices, as you’d slice a loaf of bread, and then determining the mineralogy or any other rock property for each slice.» It should easily be possible to achieve that kind of resolution with CNNs  – and without great expenditures of time and money, since coarser resolution values from laboratory measurements to determine the mineralogy, density, etc. are needed only during the training phase. «Once the model is sufficiently trained, we are almost completely independent of the laboratory values. The images alone are sufficient to determine the mineralogy»

Network with all-rounder potential

CNNs have the potential to complement or even, for some applications, to replace the standard MultiMin method, Boiger believes. To achieve this, she plans to train them further and feed them with even more data. The more data you can incorporate during training, the more precise the models become – significantly improving the reliability and significance of the results. In the future, CNNs could be used immediately if a new borehole is drilled in the same region. «Then the model would already be pre-trained with images from other drill cores and could show within seconds which and how much of which mineral is present – at almost every place in the drill core.» One additional direction that the team is currently exploring is the possibility to be able to calculate parameters such as permeability and diffusivity together with the lithology and mineralogy. It is quite possible that CNNs could develop into true geological all-rounders over time.