Artificial Intelligence for Geothermometry – A Critical Review

The chemistry of a geothermal water sample encodes the temperature of the reservoir it came from, and reading that signal is one of the most cost-effective exploration tools geothermal has. Classical geothermometers do the reading through empirical equations that assume full fluid-rock equilibrium, an assumption routinely broken by mixing, boiling and re-equilibration on the way to surface. Geologic® prepared a critical review of the main attempts over the last two decades to replace those equations with artificial intelligence, an aspect previous reviews had not explored thoroughly.

Fourteen studies from 2001 to 2024 are dissected in more detail than previous works attempted, covering each model's architecture, hyperparameters, dataset size and pedigree, temperature range and validation discipline, across solute and gas-phase geothermometers, classification hybrids and unsupervised prospecting. The reviewed studies suggest machine learning can give more accurate reservoir temperatures than classical equations where enough quality data exist, with data sufficiency and generalisation the main limits. Models trained on tens of samples often transfer poorly between geologic settings, and the review closes with an agenda for the datasets, feature engineering, architectures and interdisciplinary collaboration the field needs next.