Geologists at the Institute of Applied Geosciences at KIT, Germany, presented a novel machine learning regression approach to derive the porosity and permeability, based on the microscopic assessment of 30 µm thin rock slices. Their findings are reported in Artificial Intelligence in Geosciences.
“Porosity quantifies the volume available for fluids and gases in rocks, whereas permeability characterizes the potential of a porous rock to transmit these fluids,” explains lead and corresponding author Benjamin Busch. “Both properties are relevant for geoenergy production (e.g., geothermal, hydrocarbons) and storage scenarios (hydrogen, natural gas or CO2).”
As the distribution, contents, and textures of minerals recorded in these thin sections can classically be related to the physical rock properties, machine learning regression, capturing non-linear and multivariate relationships, was selected by the researchers as the central method to test if this hypothesis works
“Applied to a dataset containing data from 51 wells, covering four major reservoir lithologies in central Europe, collected over 25 years, prepared by at least 21 petrographers, the models show a strong predictive performance with R²=0.87 (porosity model) and R²=0.82 (permeability model),” shares Busch. “Given the large dynamic range of data and variable, even non-unified, data acquisition, the associated errors (RMSE=2.23% (porosity) and RMSE=0.64 (permeability, orders of magnitude)) are very acceptable for reservoir characterization.”
As different subsurface use cases have different economic margins to operate, cost-limiting factors are explored wherever they can. “Therefore, the cost- and time-intensive retrieval of core material from the subsurface is often reduced, limiting the access to undisturbed rock material for detailed laboratory analyses to confirm the storage and production potential (reservoir quality),” says Busch.
As differences in cement contents can clog available pore spaces and the shape of minerals grown in the pores affect reservoir quality, their microscopic analysis has classically been paramount in understanding the distribution of high reservoir quality intervals. “Ultimately, understanding the distribution of cements and compaction structures, in a framework of the development of pressures, temperatures, and chemical conditions over millions of years, should lead to an improved prediction framework for unknown areas,” adds Busch.
In the future, the presented approach may be extended to include the microscopic assessment of cutting fragments. These are mm-sized fragments of rock material produced as a by-product at every drill-site worldwide. Assessing the type and texture of individual minerals and assessing the optically visible porosity, may still enable prediction of key reservoir properties based on these materials, and may reduce costs in future drilling operations.
###
References
DOI
10.1016/j.aiig.2026.100202
Original Source URL
https://doi.org/10.1016/j.aiig.2026.100202
About Artificial Intelligence in Geosciences
Artificial Intelligence in Geosciences is an open access journal providing an interdisciplinary forum where ideas and solutions related to artificial intelligence and its applications in geosciences can be shared and discussed. To support this discussion, we encourage authors to open source their code, data, and the labels used in AI.