3D geological modelling with curlew and neural fields

Implicit structural geological modelling methods can integrate various geological constraints to rapidly constrain subsurface geometries, and are widely used for resource evaluation, geotechnical and hazard assessment, and reservoir characterisation. However, established approaches based on conventional interpolators (e.g., radial basis functions or co-kriging) often suffer from interpolation artefacts (“bubbles”) and can struggle to incorporate common constraints like stratigraphic relationships (inequalities) and geophysics data. 

In this contribution we present an update on progress developing curlew, an open-source python package for structural geological modelling using neural fields (https://github.com/samthiele/curlew/). This flexible modelling framework incorporates various local constraints (value, gradient, orientation and (in)equalities) and tailored global loss functions to ensure data-consistent and geologically realistic predictions. Progressive Random Fourier Feature encodings are adopted as a tool for improving the convergence and reliability of neural fields, and drop-out based approaches to uncertainty assessment are explored. We also present a newly developed method for deriving non-interpolated (analytical) geological prototype models and illustrate how these can be used as useful priors for hyperparameter optimization and the creation of subsequent data-driven (interpolated) models. 

Finally, the applicability of these approaches to real-world data is demonstrated through several case-studies, including the Altenberg-Teplice Caldera (Germany) and Stonepark-Pallas Green region (Ireland).