AI-Driven Optimisation of Drilling Parameters in High Temperature Geothermal Operations

Abstract

Geothermal drilling remains a dominant cost driver in high-enthalpy resource development because wells frequently penetrate hard, abrasive lithologies under high-temperature conditions that increase bit wear, drilling dysfunctions, and non-productive time. While machine-learning (ML) methods have demonstrated strong performance for rate of penetration (ROP) prediction and drilling diagnostics in petroleum operations, geothermal applications remain comparatively limited due to data scarcity, abrupt lithology transitions, and measurement-while-drilling (MWD) degradation at elevated temperatures. This study develops an end-to-end, geothermal-specific optimisation workflow that integrates supervised ML prediction (Random Forest, XGBoost, and Long Short-Term Memory networks), uncertainty quantification, and continuous-control reinforcement learning (Soft Actor–Critic, SAC) to recommend safe operational adjustments. Using adapted datasets representative of the Iceland Deep Drilling Project (IDDP) and the Cooper Basin Habanero enhanced geothermal system campaign, the LSTM achieved the best predictive performance (R² = 0.92), while feature-importance analysis identified weight-on-bit (WOB), rotary speed (RPM), and lithology as dominant drivers of ROP. Following technical validation of the drilling-time accounting, the SAC-optimised parameter trajectory reduced total drilling time from 100.0 hours to 95.9 hours, corresponding to a corrected improvement of 4.1% relative to baseline operations. Five-fold cross-validation produced a mean reduction of 4.1% (95% CI: [3.8%, 4.4%], SD = 0.15%). Uncertainty bands widened in high-temperature and highly fractured intervals, highlighting the need for uncertainty-aware decision support. Overall, the results demonstrate that integrated ML–RL optimisation can deliver measurable efficiency gains in geothermal drilling while providing transparency through uncertainty quantification and safety constraints suitable for real-world deployment. Iintegrated AI‑based optimisation frameworks can significantly improve geothermal drilling outcomes, offering faster drilling, reduced operational risk, and more cost‑effective well construction. These insights underscore the transformative potential of AI for enhancing geothermal project economics and enabling wider deployment of geothermal energy.