Scientists from Skoltech, Tomsk Polytechnical University, Gazprom Neft and Artificial Intelligence Research Institute (AIRI) developed a new intellectual method of creating detailed petroleum reservoirs permeability maps. This technique provides for highly accurate assessment of how easily oil can filter through the rock not only in the areas adjacent to wells, but in the inter-well space – exactly where the traditional methods cannot be easily applied. The method was tested using the data from the real oil field in West Siberia with the depth of pay zones of about 2400-2600 m, temperature of about 90 °C and effective net oil thickness up to 60 m.
Until recently, geologists and engineers had to manually combine different sources of information. Measurements by wells (logging and flow tests) give a rather accurate picture of permeability, but only in direct proximity to the bore. On the contrary, seismic surveys cover the entire reservoir, however, their attributes, including RMS-amplitudes, are only indirectly connected with filtration parameters. It results in the situation, when well measurements give only selective data, seismic – a general picture, and significant uncertainties remain between these two extremes.
The new method eliminates this gap by integrating all the available data into a single simulation. It is based on kernel regression – a mathematical mechanism weighing the contribution from every source depending on its reliability and distance from the well. Logging is believed to be the most reliable in direct proximity to the bore, the flow tests – within the drainage radius limits, and seismic performs as the global source filling the gaps where there are no direct observations.
The authors called the first stage of their methodology “pure combination”, because this is where they use only the wells data. With the help of the kernel, a specialized mathematical function, they assess, to which extent one can trust the measurements in every specific point on the map. Close to the well the weight of logging is big, but when you move to distance of several dozens of meters, flow tests start to dominate. The kernel parameters are selected automatically using the cross-check scheme: each well is excluded one after another, the map is built without this well and compare the forecast with real measurements. At this stage the map is received with determination coefficient circa R² ≈ 0.96 and typical artefacts are decreased significantly (bright halos around the wells emerging due to inconformity between different methods).
The received map becomes the basis for the next step – seismic data analysis. Numerous true-to-fact educational examples are required to train the neural network to recognize the links between 3D seismic cubes and permeability. There are not many in the standard environment; however, within the proposed approach points with high degree of trust located close to the wells form an expanded educational set. A small 3D fragment of seismic data is used for each point, so a 3D convolutional neural network (3D-CNN) is trained using the pairs “seismic cube – permeability”. This neural network is adjusted for work with RMS-attributes. It learns how to identify specific features of reservoir structures – vertical change of reflective properties, positional connection between sand rock and clays, acoustic contrasts – and all of them together correlate with permeability. After such training the neural network is capable of forecasting the parameters across the entire area of the field including the inter-well zones.
At the final stage all the three sources of data (logging, flow tests and forecasts by neural network) are combined again within the kernel regression framework. The kernel gets the third component in charge of the seismic contribution. Its weight is selected automatically so that the seismic data provides better level of detail where the well measurements are not enough, but does not distort real measurements. Eventually, the inter-well zones are no longer empty and start reflecting geologically reasonable reservoir structure, artificial borders between the areas with different well spacing density disappear. The quality grows noticeably: determination coefficient grows up to R² ≈ 0.972, and mean square root error decreases almost two times vs the option without seismic data.
This technology has successfully passed the test at one of the oil fields in West Siberia. The received maps accurately present the permeability values for the wells not used in the calculations and demonstrate a more detailed picture in inter-wells zones compared to traditional extrapolation methods.
