Uncertainties in joint analysis of geological and multi-source geophysical data: lessons from a blind interpretation exercise

Developing accurate 3D geological models of the subsurface is crucial, as they provide the foundations for multiple uses (e.g., resource exploration and exploitation, geohazard assessment, and environmental geoscience). The construction of these models is an intrinsically integrative task, which jointly takes into account all available data and information from multiple sources, i.e. structural geology, stratigraphy, petrophysics, geophysics. Despite the progress made in automating the integration, in particular with recent advances in artificial intelligence, human interpretation remains essential. Consequently, the performance and limitations of human geological interpretation need to be carefully assessed particularly when subsurface data are incomplete, sparse and imprecise. In this context, the French geological survey – BRGM – has set up a blind interpretation exercise that enables the geo-interpreters to test their ability to answer two main operational questions when jointly analyzing geological and multi-source geophysical datasets (seismic, gravimetric, electric/magneto-telluric): (q1) Is it possible to detect and characterize structural traps and potential migration pathways at several kilometers depth? (q2) Do the errors associated with each of the different datasets influence / affect / bias the geological interpretation? If so, how?

To this end, the following procedure was applied: (1) a simplified 3D geological model was constructed using a real exploration project dedicated to the characterization of helium reservoirs in a deep Permian sedimentary basin; (2) two cross-sections were extracted from the model with realistic petrophysical properties to constrain geophysical forward models, i.e. gravimetric, magneto-telluric, and seismic; (3) these geophysical "truths" were intentionally degraded to reflect measurement errors and realistic processing. During the 6-hour exercise, the degraded geophysical datasets along with geological data from one borehole and from the 1:1,000,000 scale geological map were provided to three teams of interpreters - each consisting of a geologist and a geophysicist, with the aim of interpreting the two cross-sections.

This communication summarizes the main lessons learned from this exercise by discussing the interaction between data resolution, quality and reliability, and cognitive biases. It points out the value of fostering recurrent exchanges with data producers during the geological interpretation process. Finally, we propose recommendations for improving the links between data-centric and human-centric inversion procedures.