MIPT Researchers Develop a Fast Method for Predicting Rock Permeability
Oil, gas, and water move through tiny pores in rock formations. How readily a rock allows these fluids to pass determines how much of the resource can ultimately be recovered. To assess this, geologists typically take a core sample, build a three-dimensional model of its pore structure, and run calculations. The process, however, can take considerable time.
Researchers at the Moscow Institute of Physics and Technology (MIPT), working with colleagues from Iran, have developed a neural network algorithm that quickly calculates how readily a rock allows liquids and gases to flow in different directions. The method represents the rock’s internal structure – the arrangement of its pores and the connections between them – as a set of numerical parameters. The resulting description contains about 250 values, rather than requiring an enormous three-dimensional image.
A Compact Set of Data for a Complete Picture
Permeability is one of the key properties assessed during oil and gas exploration and field development. It determines how easily oil, gas, or water can move through an underground formation. A rock, however, may allow fluids to flow more readily in some directions than in others. In layered formations, or where pore channels are elongated along a particular axis, fluid flows easily along those channels but encounters greater resistance across them. This directional dependence is known as permeability anisotropy. To account for this effect, which requires complex calculations, the MIPT researchers have proposed using a fast neural network model.
The algorithm combines mathematical modeling with machine learning. First, the researchers calculate statistical characteristics of the rock’s internal structure, including the spatial distribution of its pores. These characteristics are then converted into a set of 250 parameters. The neural network predicts the nine components of the permeability tensor, which describes a material’s ability to transmit liquids or gases in different directions. It produces a prediction in a fraction of a second.
Rock Properties Under the Microscope
Russia already has substantial experience in using digital tools and AI to study rocks and predict their properties.
For example, Rosneft and Innopraktika have developed a software suite called RN-Tsifrovoy kern (RN-Digital Core) for digitally modeling core properties using computed tomography and electron microscopy. The software can calculate porosity, permeability, capillary pressure curves, and multiphase flow. An updated version can model three-phase flow involving oil, gas, and water, and calculate reservoir permeability directly.
Researchers are also advancing the field. Scientists at Kazan Federal University have proposed a method for predicting the gas permeability of carbonate reservoirs using X-ray microtomography and machine learning. This approach eliminates the need for the labor-intensive reconstruction of a complete three-dimensional pore network.
Meanwhile, researchers at Perm National Research Polytechnic University have developed a method for digitally determining the properties of the near-wellbore region – the rock surrounding a wellbore through which formation fluids flow into the well. They trained neural networks using data from 486 studies of producing wells. During validation, the best-performing model achieved an error rate of less than 7%.
Researchers from MIPT and Sichuan University have also presented a method called SWNN, which can generate three-dimensional models of porous materials from a single two-dimensional image. MIPT is effectively building an integrated research program: first reconstructing the rock’s internal structure, then predicting its properties.
Lower Costs, Greater Safety
In Russia, hard-to-recover reserves already account for more than 30% of production, and their share is expected to grow. This makes technologies that can support the economically viable development of complex fields increasingly important. They can help operators use resources more efficiently, reduce production costs, and improve environmental safety by enabling better assessments of how contaminants spread through groundwater.
The MIPT researchers have also shown that solving a complex physical problem does not always require feeding a neural network a complete three-dimensional dataset. Extracting the most informative features in advance reduces the computational load.
