Professor and Head of AI for the Environment and Sustainability
Did you know that part of the solution for access to clean water and to promote the energy transition is right under your feet? The subsurface is a fantastic resource to store and extract freshwater, clean hydrogen fuel, store carbon emissions, or even extract heat from geothermal energy. But to do this requires a meticulous examination of the geological records hidden within rocks. Among these, carbonate rocks play a crucial role, as they are formed by the accumulation of skeletal grains from marine organisms. However, identifying these grains has traditionally been a task that demands specialized expertise and considerable time. This is where modern technology, specifically artificial intelligence (AI), steps in to revolutionize the field.
Carbonate sedimentology, the study of carbonate rocks, is inherently complex. The grains within these rocks come from a diverse array of organisms with various shell morphologies, which can be challenging to distinguish. The identification process is further complicated by the fact that these grains are often cut in random orientations across core sections and undergo diagenetic alterations over geological timescales. Traditionally, this intricate task has been performed by experienced carbonate sedimentologists, who rely on qualitative assessments prone to human bias and subjectivity. But data has shown that even experts disagree on the classification name for each rock!
In recent years, advances in computer vision and machine learning have shown great promise in automating geological analyses. Convolutional neural networks (CNNs), a type of deep learning algorithm, have been particularly effective in classifying lithofacies—rock types characterized by their physical and visual properties—from various geological images. Our initial study explored the potential of CNN architectures to classify carbonate core images according to the modified Dunham classification scheme (Fig. 1).
Figure 1: Modified Dunham Classification for carbonate rock. Image from Dawson et al, 2023.
We created three datasets of varying sizes, ranging from 7000 to 104,000 samples, containing images across seven classes from the modified Dunham classification. By training nine different CNNs from four architectural families on these datasets, we systematically evaluated their performance. Our results revealed that the Inception-v3 architecture achieved the highest overall accuracy of 92% when trained on the largest dataset. This underscores the importance of dataset size in training effective deep learning models. Models trained on smaller datasets showed a tendency to overfit, emphasizing the need for large, well-labeled datasets in geological applications.
These findings demonstrated the potential for AI to significantly enhance the consistency and reliability of carbonate classification, paving the way for more advanced applications in geoscience.
Building on the success of our initial study, we delved deeper into the capabilities of AI for geological analysis. The next logical step was to explore not just classification but also the identification and quantification of specific carbonate grains within high-resolution core images. This involved a shift from general facies classification to more detailed object detection tasks.
In this follow-up study, we compiled nearly 400 images from three Ocean Drilling Program (ODP) and Integrated Ocean Drilling Program (IODP) expeditions, manually labeling over 9000 individual carbonate components spanning 11 different classes. By applying a transfer learning approach, we evaluated the performance of two state-of-the-art object detection models (Fig. 2).
Our findings indicated that AI can produce results that closely mimic expert analyses, providing a consistent and reproducible method for carbonate grain identification. While the speed of detection is essential for real-time applications, the accuracy of our two-stage model makes it an invaluable tool for detailed geological studies. This study marked the first time a deep learning-based object detection task had been used to directly compare with human interpretations of carbonate biogenic components.
Figure 2: Example of object detection boxes showing the different types of skeletal components identified by our algorithm. From Dawson and John, 2024.
The integration of AI in geological research opens new avenues for understanding Earth's past. Automated identification and quantification of carbonate grains not only enhance the efficiency of core analysis but also reduce the potential for human error. This approach can significantly aid various applications, from carbon capture and storage to mineral resource exploration and paleoenvironmental reconstructions.
As data availability in geosciences increases, the deep-learning approach we present can be further refined. The frameworks developed in our studies demonstrate that AI, specifically CNNs, can be powerful tools in automating complex geological tasks. By merging the intricate knowledge of carbonate sedimentology with the precision of machine learning, we are poised to unravel the Earth's geological secrets with unprecedented accuracy and speed.
If you want to know more about this research, you can check the John Lab website (www.john-lab.org) or find the details of the study in two published papers:
Dawson, H. L., Dubrule, O., and John, C.M. 2023. ‘Impact of Dataset Size and Convolutional Neural Network Architecture on Transfer Learning for Carbonate Rock Classification’. Computers & Geosciences 171 (February):105284. https://doi.org/10.1016/j.cageo.2022.105284.
Dawson, H.L., and John, C.M. 2024. ‘Object Detection Algorithms to Identify Skeletal Components in Carbonate Cores’. Marine and Petroleum Geology 167 (September):106965. https://doi.org/10.1016/j.marpetgeo.2024.106965.