Neural network architecture for solving geological classification problems within the framework of Explainable AI ideology

Introduction. Although artificial intelligence systems have been repeatedly applied to solve geological classifi ation problems, their widespread adoption is hindered by the opacity of traditional neural networks.
Aim. To develop a new neural network architecture designed to solve problems of geological classification (determination of stratigraphy, lithology, identification of reservoirs, etc.), which would be based on the principles of explainability and verifiability of the results obtained.
Materials and methods. The set goal is achieved due to the fact that in addition to the well logging curves recorded in the studied well, a “textbook” is also fed to the network input, which is a set of well logging curves and known interpretation results for several (preferably the nearest) wells of the field. The trained network identifies the presence of similarities in the behavior of curves in the vicinity of individual depth quanta in different wells, due to which, for each quantum of the studied well, it is possible to indicate examples of similar geological situations presented in the “textbook”, on the basis of which the desired class labels are set. The mechanism for identifying similarities is based on the use of neural networks with recurrent layers that map depth quanta onto a set of normalized vectors taking into account non-local features of the curve behavior.
Results. The proposed artificial intelligence system not only allows solving geological classification problems automatically, but also allows the user to see examples from the “textbook” that justify the neural network’s setting of certain labels. This significantly simplifies the task of checking (and, if necessary, adjusting) the results obtained before their approval, which ensures a high level of reliability in the user’s interaction with the system. In addition, based on the developed architecture, it is possible to additionally solve the problems of harmonizing archival data, as well as automated selection of wells in the environment with suitable geological conditions.
Conclusion. It is likely that the approach to creating neural network architectures developed in this work will help overcome existing obstacles to the widespread implementation of artificial intelligence systems in the oil and gas sector, since it most fully meets the requirements for transparency and safety of using such systems.

1. Wu X., Nyland E. Automated stratigraphic interpretation of well-log data. Geophysics. 1987, no. 52, pp. 1665–1676. https://doi.org/10.1190/1.1442283

2. Adadi A., Berrada M. Peeking inside the black-box: A survey on explainable artificial intelligence (XAI). IEEE Access. 2018, no. 6, pp. 52138–52160. https://doi.org/10.1109/ACCESS.2018.2870052

3. Confalonieri R., Coba L., Wagner B., Besold T.R. A historical perspective of explainable artificial intelligence. WIREs Data Mining and Knowledge Discovery. 2021, no. 11(1), e1391. https://doi.org/10.1002/widm.1391

4. Guidotti R., Monreale A., Ruggieri S., Turini F., Giannotti F., Pedreschi D. A survey of methods for explaining black box models. ACM computing surveys. 2019, no. 51(5), pp. 1–42. https://doi.org/10.1145/3236009

5. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence. 2019, no. 1, pp. 206–215. https://doi.org/10.1038/s42256-019-0048-x