Integrating geology-informed constraints into machine learning–based borehole interpretations for subsurface modelling: A case study from the Netherlands

Geological mapping and 3D subsurface modelling require consistent geological interpretations across large datasets with heterogeneous spatial coverage and information density. In the Netherlands, several subsurface models rely heavily on borehole lithological descriptions to map lithostratigraphic units and geological structures. Automated interpretation approaches based on machine learning (ML) are being developed to transfer expert geological interpretations to previously unseen boreholes, thereby increasing the number of interpreted boreholes that can be incorporated into subsurface models. However, existing neural network-based approaches for borehole interpretation often struggle to consistently respect the stratigraphic and spatial relationships derived from expert geological knowledge.  In practice, automated interpretations can produce stratigraphically inconsistent successions, with younger units incorrectly predicted to occur below older ones, or units appearing outside their known regional extent. This limitation stems from ML training objectives that prioritise local classification accuracy (e.g., categorical cross-entropy loss) over regional geological plausibility. 

To improve the geological plausibility of ML-generated interpretations, we introduce geology-informed loss functions that account for stratigraphic consistency and the spatial extent of lithostratigraphic units. The proposed loss functions are combined with a standard classification loss during model training on expert-interpreted boreholes and evaluated on previously unseen boreholes drawn from the same national dataset, comprising 7,500 boreholes in total. By varying the relative weight of each loss function during model training, we found that ML models trained with a combination of geology-informed loss functions and standard categorical cross-entropy substantially reduce geologically implausible stratigraphic transitions, increasing the proportion of stratigraphically consistent transitions from approximately 90% to up to 95%, and making fewer predictions of lithostratigraphic units outside their known regional extent.  These improvements in geological plausibility do not lead to a noticeable change in overall classification accuracy (≈ 75% across different loss-weight combinations). Incorporating geology-informed training objectives, therefore, provides a practical way to improve the plausibility and consistency of automated borehole interpretations used in large-scale subsurface modelling workflows.