Summary Drill cuttings are small rock fragments generated by the action of the drill bit during rotary drilling operations. The rock type (lithology) identification and quantification from drill cuttings provides critical information for underground characterization. The common practice of mudlogging—to identify the lithology from cuttings—is time-consuming and subjective. The previous efforts in automating lithology characterization from drill cuttings focused on analyzing cuttings optical images with convolutional neural networks (CNNs). The success is limited as many samples are not visually distinct. HyLogger-3 is an automated drill cuttings and core profiling system, providing both optical images and hyperspectral data of samples. In this study, a multimodal machine learning (ML) approach is developed to automate lithology prediction from cuttings. A CNN classification model is first trained for extracting the image features from 2D optical images of cutting samples in red, green, and blue (RGB) channels. Mineral information is derived from HyLogger hyperspectral data by comparing measured reflectance spectra with pure mineral reference spectra and applying spectral unmixing methods to quantify the abundance of individual minerals within each sample. The extracted image features are in the form of a 1D list of float numbers and can be concatenated with another 1D list of mineral information to further build a multimodal ML model. An XGBoost-based regression approach is used in the multimodal ML modeling, where the percentage of different lithology types (mudlogs) in cuttings samples can be directly used to train the ML model, and the trained model can produce mudlogs on HyLogger-3 data from randomly selected blind samples. The blind test results indicate that the multimodal ML model delivers a practical solution for automated mudlog generation in both pure and mixed lithology samples, matching interpretations of geologists with discrepancies of 10.16% in root mean square error (RMSE) and 4.84% in mean absolute error (MAE).
Liang et al. (Sun,) studied this question.