Evaluation of Machine Learning Models for Geothermal Downhole Temperature Prediction Using Hydrogeochemical Datasets

Deep beneath the Earth's surface, hidden within fractured rocks and ancient fault lines, lies a vast reservoir of heat. This geothermal energy offers a clean, steady source of power that could help nations move away from fossil fuels, but finding the best places to drill is a difficult and expensive gamble. In many regions, including the rift valleys of Ethiopia, scientists cannot simply look at the ground to see how hot it is miles down. Instead, they must rely on indirect clues, such as the chemical makeup of water that has circled through the hot rocks and returned to the surface. For decades, experts have used these chemical signatures to guess the temperature of the underground reservoirs, but these traditional guesses often carry a large margin of error. The challenge has been to find a way to turn these scattered chemical clues into a precise map of underground heat, a task that requires spotting patterns too complex for standard math to solve.

A team of researchers from Addis Ababa Science and Technology University has taken a new approach to this problem by using machine learning, a type of computer intelligence that learns from data rather than following a fixed set of rules. They gathered information from fourteen different geothermal fields across Ethiopia, compiling a dataset of nearly two hundred samples. For each sample, they recorded the temperature found deep underground and the specific chemical ingredients present in the water, such as the amounts of silica, potassium, sodium, and magnesium. The goal was to teach a computer to recognize the relationship between these chemical ingredients and the hidden heat, allowing it to predict temperatures in places where no direct measurements exist. The researchers tested five different computer models, ranging from simple linear equations to advanced tree-based systems that can handle complex, non-linear interactions between variables.

The results showed that the more sophisticated computer models were far better at the job than the simpler ones. When the researchers tested the models against data they had not shown the computer before, the most successful model, known as CatBoost, correctly predicted the underground temperature with a high degree of accuracy. It achieved a score indicating that it captured over ninety percent of the variation in the real temperatures, with an average error of about twenty-three degrees Celsius. Another strong performer, called Random Forest, was even better at minimizing the size of its mistakes on average, missing the true temperature by only about eleven degrees. In contrast, the simpler models, which rely on straight-line relationships, struggled significantly, producing much larger errors and failing to capture the full range of temperatures. The study found that the computer models learned that certain chemicals, particularly silica and potassium, were the strongest indicators of heat, while others, like magnesium, tended to decrease as the water got hotter.

Despite the success of these computer tools, the researchers noted that the models still faced challenges when predicting the very hottest temperatures. The predictions tended to be slightly less accurate in the extreme upper range, suggesting that the underground conditions in those specific zones might be more complicated than the chemical data alone could reveal. This could be due to hidden variations in the rock types or the way fluids move through fractures, factors that the chemical samples did not fully capture. The team concluded that while these machine learning models are powerful tools for reducing uncertainty in geothermal exploration, they work best when combined with a deep understanding of the local geology. By using these digital tools to interpret the chemical language of the Earth, scientists can make more informed decisions about where to drill, potentially saving time and money while unlocking a cleaner energy future.