Digital Rock Sample Retrieval Using Variational Autoencoder and Clustering

Abstract The analysis of digital rock samples is paramount in the oil industry. In this context, computed tomography (CT) enables the volumetric, non-destructive representation of rock microstructures with high-quality detail, allowing experts to assess stratigraphic, sedimentological, and petrophysical properties of exploration fields. Efficient comparison of new samples with previously analyzed ones can significantly reduce the time cost of interpretation, yet datasets are often poorly organized and lack descriptive labels. In this way, content-based image retrieval (CBIR) is an effective approach to search and compare images based on similarity measures. Therefore, this work proposes an automatic CBIR method for CT images of rock samples, using an unsupervised deep learning strategy based on autoencoder architectures to extract meaningful feature representations. A clustering step was used, employing the DBSCAN algorithm to group similar samples and refine the representation space. Two clustering strategies were tested, with the most effective involving the creation of specialized models trained within defined clusters, rather than a single general-purpose model. The proposed approach achieved promising results, reaching a mAP score of 92.8% and an F1-Score of 93.4%. These results indicate that clustering reduced intra-group variability, enabling models to capture more subtle visual differences between samples and lowering retrieval errors.