Vision-language models for automated carbonate petrography and depositional environment interpretation

Carbonate petrographic analysis provides essential qualitative and semi-quantitative constraints on depositional environments and diagenetic evolution at microscale. However, conventional thin-section description and interpretation are labor-intensive, time-consuming, and susceptible to interpreter subjectivity, limiting reproducibility across studies. While recent deep learning methods have been applied to carbonate petrography, most are confined to narrow classification tasks and lack the holistic geologist-level capabilities required for comprehensive petrographic analysis. Advances in vision-language models (VLMs) offer new opportunities for joint image understanding and text generation, yet their application to carbonate petrography remains limited by vision encoders pretrained on natural images rather than geological textures. Here, we present a vision-language framework for automated carbonate petrographic description and depositional environment interpretation. The framework integrates an EVA vision encoder with a Llama2-based language model via a linear projection module to align visual and textual representations. A curated carbonate vision-language dataset (CarbonateVLD) comprising over 57,000 thin-section images, including augmented image-text pairs from diverse sources, was used for training to enhance generalizability across textures and facies. To translate generated descriptions into standard microfacies and facies-zone interpretations, a two-stage semantic reasoning transformer was developed to embed carbonate textural attributes in a high-dimensional representation space. Quantitative evaluation shows improvements of 36.7% in BERT-Similarity and 16.9% in SPICE score relative to baseline VLMs. Depositional environment interpretations achieve 93.3% accuracy when benchmarked against expert geologist assessments. By standardizing petrographic description and interpretation, this study demonstrates that vision-language modeling can reduce subjectivity in carbonate petrography while enabling efficient and reproducible large-scale carbonate petrographic description and depositional environment interpretations.