Machine learning-based real-time lithology prediction from drilling acoustic signals

Accurate lithology classification is essential for optimizing drilling operations and improving reservoir understanding. Conventional well logging methods often suffer from delays, limiting real-time decision-making. This study introduces drilling acoustic signals as a high-frequency alternative and presents a novel hybrid MLP-XGBoost framework for lithology prediction. Acoustic features, including dominant frequency, amplitude, spectral centroid, spectral bandwidth, and zero-crossing rate, were extracted from drilling sounds recorded at the Johan Sverdrup 16/2 6 well. Three machine learning models were trained: Multi-Layer Perceptron (MLP), XGBoost, and the hybrid MLP-XGBoost model. The hybrid model, unique in its design, leverages the MLP as an adaptive feature engineer to learn complex lithology-specific patterns, while XGBoost performs robust classification, even with noisy data. This two-stage approach achieved the most balanced performance, with 86 % accuracy on the training set and 85 % on the testing set, outperforming the standalone MLP (78 % and 76 %) and XGBoost (98 % and 82 %). Feature importance and SHAP analyses revealed that dominant frequency, amplitude, and spectral centroid were key contributors, demonstrating how the model integrates both dynamic and steady acoustic characteristics to distinguish lithologies. Validation across eight North Sea wells yielded accuracies from 81.8 % to 89.5 %, confirming the model's generalization potential. Overall, this study establishes, for the first time, a robust and scalable machine learning framework for real-time lithology classification using exclusively acoustic signals, offering a practical complement to traditional formation evaluation techniques and opening new opportunities for integration with operational drilling parameters. • Machine learning frameworks were used for real-time lithology prediction using acoustic drilling audio signals. • The key acoustic features were dominant frequency, amplitude, and spectral centroid. • Combining deep learning with gradient boosting improves classification performance over standalone models. • Field validation using multiple well core sample lithologies yielded good accuracy. • Hybrid acoustic model offers a viable alternative to conventional well logging techniques.