Deep Neural Networks for Rate of Penetration Prediction: A Subsurface-Centric Approach Using Well Logs in the Volve Oil Field

Accurate Rate of Penetration prediction is pivotal for optimizing drilling efficiency, yet conventional models often neglect subsurface heterogeneity and rely heavily on operational parameters. This thesis introduces a subsurface-centric deep learning framework for ROP prediction based solely on well log data, addressing geological complexity through an iterative, three-phase methodology. In the reservoir interval (2600–3600 m), ten deep neural network architectures were evaluated, with the optimal model achieving high predictive accuracy (training MAE = 1.15, MSE = 2.99; validation MAE = 1.35, MSE = 3.92). To extend applicability across the full wellbore, including the non-reservoir section (146–2600 m), the same DNN model was applied, demonstrating a drop in performance (validation MAE = 4.96, MSE = 81.56) likely due to variations in lithological complexity outside the reservoir. To address this, unsupervised learning was employed to characterize lithological heterogeneity using k-means clustering, validated via silhouette scores, elbow method, t-SNE, and PCA. Additional features—including cluster labels, cluster probabilities, and a spatial-temporal metric capturing ROP-depth trends—were engineered and integrated into the DNN model. This refinement reduced training loss by 78.05% and validation loss by 76.61%, yielding improved performance (validation MAE = 3.14, MSE = 19.09) over the baseline model. The framework’s broader significance lies in its geology-first paradigm: trained on a single well’s logs, it enables ROP prediction across entire wellbores and nearby wells lacking historical data. This capability supports proactive bit selection, parameter optimization, and risk mitigation in complex formations, enhancing both efficiency and safety. Validated on the Volve Oil Field dataset, the work bridges geoscience and machine learning, offering a scalable, data-efficient strategy for global drilling operations.

Grantor dc:publisher.institution

University of Bradford

Author dc:creator

• Isam, Sherrif

Advisors dc:contributor.advisor

• Zhang, Qichun
• Qahwaji, Rami

• Rate of penetration (ROP)
• Deep neural networks (DNN)
• Well log imputation
• Lithological clustering
• Spatial-temporal feature engineering
• Well log interpretation
• Subsurface heterogeneity
• Drilling optimization
• Volve oil field

Statement dc:rights

• <a rel="license" href="http://creativecommons.org/licenses/by-nc-nd/3.0/"><img alt="Creative Commons License" style="border-width:0" src="http://i.creativecommons.org/l/by-nc-nd/3.0/88x31.png" /></a><br />The University of Bradford theses are licenced under a <a rel="license" href="http://creativecommons.org/licenses/by-nc-nd/3.0/">Creative Commons Licence</a>.

Language dc:language.iso

en

Repository record dc:identifier.uri

https://bradscholars.brad.ac.uk/handle/10454/20976.2

OAI identifier oai:identifier

oai:bradscholars.brad.ac.uk:10454/20976.2