Résumé
The Barrois region is a karstified Limestone Plateau situated in the east of the Paris Basin. The regional karst appears underground in the form of a dense network of conduits of variable size (centimetric to pluri-metric scale), and at the surface in the form of sinking streams, dolines, springs, and various cave entrances. The karst network itself is polyphased, and was created over a complex history of karstification that involves and intertwines ghost-rock karstification, crypto-alteration, and epigenic alteration under cover. Most conduits there being inaccessible and difficult to detect, the only way to factor them explicitly in a broader flow and transport model would be to simulate them. Several attempts were made in that direction (Jaquet, 2004 ; Gouy et al., 2024). The latest one (Gouy, 2025) uses KarstNSim, an open-access code that stochastically simulates 3D karst networks constrained by geological and karstological knowledge and data. It captures most of the processes involved in shaping the Barrois karst network, but its handling of the simulation of conduit sizes lacks realism.In this work, we thus propose an enhanced conduit-sizing methodology that builds on a karst network skeleton. Conduit dimensions (e.g., equivalent radius and aspect ratio) are simulated at the skeleton’s discrete nodes using a 1D-curvilinear, branchwise Sequential Gaussian Simulation (SGS) scheme adapted from Frantz et al. (2021). Branches are part of the network between two junction nodes. Spatial correlation is represented through a multiscale variographic structure (intrabranch, interbranch, and global variograms), allowing the model to reproduce variability within branches, between branches, and at junctions. The method explicitly honors measured or interpreted values at selected nodes (springs, inlets, explored caves), including constraints related to conduit enlargement when traversing ghost-rock volumes. To relax the stationarity limitation of simple kriging with fixed neighborhoods and to better reproduce large-scale organization, we introduce an external drift so that the simulated radius is expressed as the sum of a deterministic trend and a zero-mean stationary residual field. The drift is formulated as a regression on normalized explanatory variables that reflect (i) the node vertical distance to the water table, and (ii) the upstream integrated curvilinear length, thereby capturing contrasts between vadose and phreatic settings and the general downstream increase in conduit size respectively. Regression parameters are calibrated using a weighted least-squares strategy on all available observations except ghost-rock-driven voids that would otherwise bias the drift. This workflow produces observation-consistent 3D conduit networks that display both a local spatial correlation from the variogram models and a long-range trend that is directly inferred from available observations. This approach allows the construction of a 3D network that can be exported directly in a suitable flow model such as MODFLOW-CFP.