Improving the Reliability of Geological Data with Convolutional Neural Networks

Research Highlights

Geological practices, both in industry and academia, rely heavily on observational data from a variety of scales. However, ambiguities in the petrographic description of rock facies can reduce the reliability of this data. In a recent study, we explored the potential for using convolutional neural networks (CNNs) to classify rock facies from digital images.

We evaluated the performance of top-performing CNNs using transfer learning on three datasets containing carbonate core images across seven classes from the modified Dunham Classification. These datasets ranged in size from 7000 to 104,000 samples. We found that the Inception-v3 architecture was best suited to this classification task, achieving an accuracy of 92% when trained on the larger dataset.

Our research also showed that the size of the dataset plays a significant role in the performance of the models. Those trained on smaller datasets tended to overfit, highlighting the importance of using sufficiently large datasets when applying deep learning in the geosciences.

This study has important implications for the use of deep learning in the classification of carbonate rocks and has the potential to be easily modified for the classification of cores from different formations and lithologies. By improving the reliability of geological data, this research has the potential to advance our understanding of the Earth’s history and processes.