A Deep Learning Geothermometer for Deep Reservoirs

Knowing the reservoir temperature shapes every early geothermal decision, yet measuring it means drilling. Classical chemical geothermometers infer it from spring chemistry but assume a fluid-mineral equilibrium that mixing and degassing routinely break; multicomponent methods require complex optimisation and mineralogical expertise. This 2025 Geologic® study introduces aiION®, a deep learning chemical geothermometer that predicts deep reservoir temperature from eight routinely measured inputs: the major ions, silica and pH of a water sample.

The model was trained on 674 rigorously curated water samples from Nevada's Great Basin, one of the largest geothermal provinces in the world, with target temperatures combining four inference methods: classical silica geothermometers, iGeoT multicomponent geothermometry, a 2,365-entry regional bottom-hole temperature database and a published machine learning geothermometer. Among four benchmarked algorithms, the deep neural network won decisively, explaining over 97% of temperature variance with a test-set mean absolute error near 2.6 °C. Validated on 42 new wells from geothermal fields worldwide, aiION® held an R² of 0.83 but a mean absolute error of 22.8 °C on that external set, and many validation samples also fell outside the ±5% charge-balance limit, so part of that error may come from input data quality.