WP4 - 3D soil property modelling

WP4 focuses on developing efficient and robust techniques for the analysis and interpolation of large, high-dimensional spatial datasets, such as 3D seismic and soil property data across extensive geographic areas. The adoption of these advanced machine learning approaches is driven by computational challenges of interpolating predicted properties along seismic lines in 3D – due to the extremely high number of data points and size of the 3D model grids – when using traditional interpolation methods (example shown in Figure; NGI (2025)). This work package is aimed at investigating the use of advanced machine learning models, such as Variational Autoencoders (VAEs) and Diffusion models to address the challenges posed by interpolation of complex, multimodal spatial data (2D/3D seismic data, CPTs and interpreted data) in a probabilistic framework.

Multimodal learning approaches and latent space representations (see Shi et al., 2019; Li et al., 2023) form the methodological backbone of this work. In particular, the project explores the application of Gaussian Processes (GPs) within the learned latent spaces to better capture and understand spatial correlations among different modalities.

WP4 also aims to establish rigorous methods for evaluating the reliability and generalizability of machine learning models. This includes systematic studies of where ML-based predictions (such as soil property estimations) perform well and identifying conditions under which they may fail, particularly under varying geological scenarios. Model interpretability techniques are employed to analyse the learned correlations and relationships, thereby providing fundamental geotechnical insights and advancing domain understanding.

References:

Shi, Yuge, N. Siddharth, Brooks Paige, and Philip H. S. Torr. Variational Mixture-of-Experts Autoencoders for Multi-Modal Deep Generative Models. arXiv:1911.03393. Preprint, arXiv, 8 November 2019. https://doi.org/10.48550/arXiv.1911.03393.

Li, Jintao, Xinming Wu, Yueming Ye, et al. Unsupervised Contrastive Learning for Seismic Facies Characterization. GEOPHYSICS 88, no. 1 (2023): WA81–89. https://doi.org/10.1190/geo2022-0148.1.

NGI (2025). Integrated Ground Model Report - Ground Model for the Ijmuiden Ver Gamma Wind Farm Site, Document No.: 20220204-05-R – Revision No. 2 (October, 2025).