Real-Time Lithology Prediction from Drilling Parameters Using a Multi-Layer Perceptron Neural Network

Abstract

Lithology identification is a critical process in drilling and reservoir management. However, traditional methods such as coring, cuttings analysis, and manual interpretation are costly, subjective, and difficult to scale. While logging-while-drilling (LWD) offers real-time insights, sensor offsets often introduce delays in formation evaluation, particularly in dynamic stratigraphic settings. To address these limitations, this study proposes a data-driven lithology prediction approach based on drilling parameters; Weight on Bit (WOB), Torque, Rate of Penetration (ROP), and Rotary Speed (RPM); to enable faster and more objective formation characterization. A physics-informed sequential deep learning workflow was developed using the Equinor Volve dataset, comprising five lithology classes: sandstone, marl, claystone, dolomite, and limestone. After data cleaning, Mechanical Specific Energy (MSE) was computed as a physics-based feature, while depth continuity was captured using a five-step lag sequence. To mitigate overfitting associated with the limited and imbalanced dataset, stochastic Gaussian noise augmentation expanded the original 270 samples to over 2,400. A deep Multi-Layer Perceptron (100–50–25 neurons with Softmax output) was trained using stratified five-fold cross-validation. The proposed model achieved an overall classification accuracy of approximately 98%, with an average cross-validation accuracy of 99.34%. Sandstone was classified with near-perfect accuracy, while marl and claystone also exhibited strong predictive performance. Misclassifications were primarily observed in ultra-minority carbonate classes, suggesting a few-shot learning challenge rather than systematic confusion between reservoir and non-reservoir lithologies. The results demonstrate that integrating physics-informed features, sequential contextual information, and data augmentation significantly enhances lithology prediction from drilling parameters. These findings indicate that deep learning, when guided by physical principles and contextual constraints, can serve as a viable alternative to expensive coring and labor-intensive analysis. Although real-time deployment challenges remain, the proposed framework offers a scalable and cost-effective solution for formation evaluation in drilling operations.