VI Seminar #100: From Pixels to Minerals: Multimodal Machine Learning for Rock Cuttings Characterization

Presented by Yohanes Nuwara, Senior Data Scientist at Aker BP

Abstract

Characterizing subsurface geology from rock cuttings is an important phase of oil and gas exploration, but conventional mineralogical analysis requires sample preparation, laboratory measurements, and expert interpretation. The Released Wells Initiative (RWI), initiated in 2019, has generated a large digital dataset of approximately 700,000 cutting samples, combining laboratory measurements such as X-ray diffraction (XRD) and X-ray fluorescence (XRF) with white-light and ultraviolet photographs.

This work investigates how machine learning can derive additional information from these multimodal data. An XGBoost model was trained using data from 65 wells to predict XRD-derived mineralogy from XRF elemental compositions combined with image-derived color (L*a*b*) and texture features (gray-level co-occurrence matrix). Models combining XRF and image features achieved the strongest predictive performance, particularly for major mineral groups such as carbonates, silicates, clays, and sulfides, as well as individual minerals including calcite and quartz. The results demonstrate that integrating geochemical measurements with digital cuttings imagery can improve data-driven mineralogical characterization of subsurface formations.