Coreless Geo: A Machine Learning and Retrieval-Augmented Generation Framework for Geomechanical Rock Property Prediction and Literature-Grounded Geological Interpretation

Coreless Geo presents an end-to-end system for geomechanical rock property prediction from well log data, making three concrete research contributions to the field. Contribution 1 — Correct Evaluation Protocol: This work demonstrates that the random 80/20 validation protocol prevalent in the well log prediction literature produces optimistically biased R² scores, and establishes Leave-One-Well-Out cross-validation as the correct standard for cross-well generalisation assessment. This has immediate implications for how future work in this area should be evaluated and reviewed. Contribution 2 — Literature-Grounded LLM Interpretation: This work introduces the first application of Retrieval-Augmented Generation to geomechanical well log prediction, grounding LLM-generated geological interpretations in a curated vector database of published geology literature. The fixed JSON output schema with mandatory citation fields prevents hallucination and enables expert peer review of AI-generated geological reports — a critical requirement for adoption in professional petroleum engineering contexts. Contribution 3 — Quantified Cross-Basin Domain Shift: This work provides the first explicit quantification of performance degradation when a North Sea-trained geomechanical model is applied to a geologically distinct basin, establishing cross-basin generalisation as an open research frontier. The cross-basin ΔR² metric is proposed as a standard reporting requirement for applied geomechanical ML work, particularly for models intended for deployment outside their training basin. The production Streamlit web application demonstrates that all of this functionality — prediction, SHAP explanation, and RAG-grounded geological reporting — is accessible to petroleum engineers without machine learning expertise, delivering complete results within seconds of uploading a well log CSV.  

[1] M. D. Zoback, Reservoir Geomechanics. Cambridge: Cambridge University Press, 2010. [2] L. Breiman, “Random Forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. [3] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., 2016, pp. 785–794. [4] C. Cortes and V. Vapnik, “Support-Vector Networks,” Machine Learning, vol. 20, pp. 273–297, 1995. [5] S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” in Advances in Neural Information Processing Systems, vol. 30, 2017, pp. 4765–4774. [6] Z. M. Ameen, H. Merkel, and A. Wimmers, “Estimation of Rock Strength Parameters Using Machine Learning: A Case Study from the North Sea,” Journal of Petroleum Science and Engineering, vol. 172, pp. 1285–1298, 2019. [7] C. Chang, M. D. Zoback, and A. Khaksar, “Empirical Relations Between Rock Strength and Physical Properties in Sedimentary Rocks,” Journal of Petroleum Science and Engineering, vol. 51, no. 3–4, pp. 223–237, 2006. [8] H. Wang, Y. Liu, Q. Chen, and Z. Zhang, “Machine Learning-Based Prediction of Well Logs Guided by Rock Physics,” MDPI Sensors, vol. 25, no. 3, pp. 1–18, 2025. [9] M. Ehsan, M. Gu, A. Akhtar, and S. Javid, “Pore Pressure Prediction Based on Conventional Well Logs Using Machine Learning,” Journal of Rock Mechanics and Geotechnical Engineering, vol. 17, no. 5, pp. 2890–2907, 2025. [10] T. Ma, R. Li, Y. Liu, and Z. Chen, “Progress and Development Direction of Intelligent Prediction Technology of Geomechanical Parameters,” Petroleum Science Bulletin, vol. 9, no. 3, pp. 451–467, 2024.