ML-based techniques for enriching databases

This deliverable presents the machine learning (ML) and deep learning (DL) methods and trained models developed in Task 2.1 of Work package 2 of the Horizon Europe project GO-Forward during the first 21 months of the project. The objective of Task 2.1 is to develop ML-based techniques for providing calibration data for sedimentary-, diagenetic- and fracture-forward-modeling approaches. The calibration data, such as facies, seismic horizons, porosity or mineral phases is derived from primary datasets, e.g. seismic data, well-log data, image data (thin section or SEM data). The methods are to be used in the subsequent forward modeling workflows employed in Work package 3. Three distinct data domains are addressed in this deliverable: seismic data, scanning electron microscope (SEM) image data, and well-log / core-plug data. Seismic data: Self-supervised and semi-supervised deep learning methods were applied for data conditioning (denoising and interpolation of vintage seismic data) and automated interpretation (fault detection and horizon classification). These methods directly support geothermal exploration by enabling efficient processing of large volumes of legacy 2D and 3D seismic data. SEM image data: A two-stage image segmentation pipeline combining Random Forest classification and U-Net deep learning was developed for automated mineral phase identification in sandstone thin sections. The pipeline distinguishes pore space, primary quartz, quartz overgrowth, and other mineral phases from paired BSE and CL images. This provides quantitative input for diagenetic forward modeling of clastic reservoirs. The pipeline supports pseudo-labeling and iterative self-training to maximize performance with limited labelled data. Well-log data: Supervised ML regression models were trained on a compiled dataset of nine wells from the Netherlands (NLOG database) to predict core plug porosity from well-log measurements. After systematic core-log depth alignment and feature engineering, five ML algorithms were compared. Support Vector Regression (SVR) achieved the best performance, demonstrating the feasibility of ML-based porosity prediction as calibration input for stratigraphic forward models. All trained models, usage documentation, and associated datasets are published as open-access resources on Zenodo (see section 1.3) under EUPL-1.2 (code) with appropriate data provenance documentation.