Three-dimensional geological interpretation is a knowledge-intensive process that requires human experts to interpret subsurface data and mobilize domain knowledge. However, interpretations can be subjective and error-prone, as data under-sampling forces experts to make implicit choices, while most of current modeling algorithms capture only a fraction of expert knowledge and offer limited transparency into the reasoning behind an interpretation. A knowledge-driven modeling formalism was recently proposed to address these limitations. It makes the interpretation reasoning explicit as an iterative process in which experts sequentially observe parts of the input dataset, identify the geological structure type each observation belongs to, verify interpretation consistency, and revise their conclusions when needed. Building on this formalism, we propose a Graph Neural Network (GNN) approach with the objective of learning and automating this reasoning process. The interpretation reasoning is represented as a concept graph — a heterogeneous graph in which observations, interpretation situations, structural hypotheses, and detected anomalies are encoded as nodes linked through semantic relations. This graph captures the reasoning chain, from the selection of parts of the input data to the identification of the geological structures they correspond to. To generate a training dataset, we enumerate alternative reasoning sequences from the expert's reasoning sequence using a combinatorial engine based on Lehmer factoradic indexing. This yields a structured corpus that covers the variability of interpretation paths over the same geological section, which will serve as input to a GNN trained to learn which reasoning steps are geologically plausible and to autonomously select the next interpretation action at each step of the process.

Keywords: graph neural networks, geological structure interpretation, combinatorial sequence generation, iterative interpretation process, knowledge-driven formalism