Geothermal energy, while a reliable baseload low-carbon resource, only comprise a small fraction of global renewable capacity due to high upfront costs and resource risks. Drilling wells accounts for ∼60 % of capital investment costs, thus finishing wells on-time and within budget has always been a crucial challenge for operators and challenges like fault structures, severe lost circulation, and high temperatures inherent to geothermal systems make this difficult. Early detection is crucial in taking corrective actions before problems escalate and leveraging machine learning (ML) technologies offers the potential to identify patterns that precede hole-related non-productive time incidents, such as stuckpipes or borehole instability. This research investigates the capability of ML to predict drilling incidents in geothermal wells in New Zealand, even when drilling on total losses. The models were performed on 2-h windows rolling every hour to demonstrate forward prediction. It also demonstrates how systematic and automated feature engineering outperforms naïve and manual feature engineering on incident prediction on several machine learning algorithms. The study proves that there is a huge potential for utilising ML to create real-time incident detection systems to assist drilling personnel in making decisions during drilling, thereby reducing operational risks and enhancing overall drilling and cost performance.