AI vs. Human Decision-Making
Humans are unpredictable and make mistakes. When designing automated systems, the goal is to remove human error and create something reliable and predictable. But how predictable are AI-powered geosteering systems?
Source:
Rune Rolvsjord
Sergey Alyaev, Senior researcher at NORCE
— Geologists use their expertise to interpret geology and make steering decisions when drilling a well. But their choices are subjective and sometimes suboptimal. We are developing an AI that avoids human errors and makes better geosteering decisions, says Sergey Alyaev, a senior researcher at NORCE.
Together with colleagues from UiS and Stanford, researchers at the NORCE-led research centre DigiWells, have developed an AI-powered geosteering system to tackle the challenges with subjective interpretation and suboptimal choices.
The AI learns geological patterns from data. When trained, it can simultaneously track thousands of possible scenarios. But despite its ability to model many scenarios, it must commit to a single decision at every step, just like a human.
— Even when we use AI, slight variations in decisions occur, and lead to significantly different well paths. During this project, I learned that an AI can also be influenced by "luck". And we need to account for that when developing these systems, says Alyaev.
The Geosteering World Cup
The Rogii Geosteering World Cup is a large-scale virtual competition. It provides valuable insights into human geosteering performance in layered geological scenarios. It allowed the team to compare the AI with human experts.
— Unlike real-world drilling, where repeated testing is logistically impossible, the GWC offers a fair and realistic benchmark for AI decision-making under complex conditions. It can be a key step toward real-world validation, says Alyaev.
The results from GWC may reveal a realistic yet rarely seen picture of variability arising from geological uncertainty, limited real-time data, and rapid decision-making.
The figure shows the performance of Human participants and repeated attempts by the same AI agent. The results are evaluated based on the following: • Geological interpretation accuracy (x-axis) – how well participants interpreted subsurface layers and the well location. Note that the AI system had a range of interpretations for each run. • Percentage of the well within the target layer (y-axis) – how effectively they steered towards the optimal drilling zone.
GWC 2021 results (see Figure) showed a broad distribution of human performance.
The AI system, leveraging automated interpretation, probabilistic modeling, and advanced decision-making algorithms, delivered highly competitive results.
- The AI systems average performance was superior to that of human participants, and its best individual run outperformed all human competitors, explains Alyaev.
This study is published in the concluding paper of Ressi Bonti Muhammad's PhD. His work provides critical insights into the behavior of AI-powered geosteering systems and their decision-making variability in uncertain geological conditions.
Key Takeaways
So, what did the study teach us about using AI within geosteering?
- Modeling uncertainty is essential – Understanding geological uncertainty improves geosteering reliability.
- Modeling decisions is even more critical – AI systems must optimize decision-making strategies, not just interpretation.
- Experiments provide critical insights – Controlled testing environments allow assessment of AI robustness and consistency.
— These findings highlight that while AI-driven geosteering is highly effective, decision variability remains a key challenge. By refining uncertainty modeling and decision optimization, AI-based geosteering can achieve even greater reliability and outperform human experts in complex drilling scenarios, Alyaev sums up.
