Standards, Training & Reasoning for AI in subsurface characTerisAtion (STRATA-AI)

Objective

To evaluate and develop AI and machine learning methodologies for integrating subsurface datasets, with an emphasis primarily on deep geological boreholes, to enhance geological understanding and predictive modelling in support of Geological Disposal Facility (GDF) siting.

Description of the WP

Safety cases, engineering designs, programmes and work packages to support the siting of Geological Disposal Facilities (GDF) generate and require vast quantities of high-quality, verified and validated subsurface data and information. To achieve this, Waste Management Organizations (WMOs) subsurface characterisation programmes from early characterisation through to concept design stage are underpinned by a broad range of data acquisition campaigns. The data collected in these campaigns are highly varied, with variable sampling, resolution, parameterisation and quality amongst a range of factors. Subsurface data may fall into quantitative measurements (e.g. a point density measurement, core mineralogical analysis), relative measurement (e.g. Seismic downhole VSP reflection data) or qualitative (e.g. a core photograph, outcrops). The requirement to integrate disparate and multimodal data types is a common and long-standing challenge in geosciences. This workflow often relies on qualitative or model-based analysis and interpretations presented in long form reporting to produce models and conclusions.  
AI and machine learning methods offer promising tools for integrating diverse and complex subsurface datasets. These models excel at convolving different data types, for example images, text and numbers, by converting them into data space and then learning relationships between them. For example, such techniques have been shown to link digitised core photographs with descriptive metadata, mineralogy or wireline log data, enabling the generation of one from the other. This approach supports both forward and reverse predictions, allowing for the synthesis or interpretation of missing or incomplete data based on learned patterns, together with an efficient check of the consistency of the different data response.
The databases aggregated in the siting process for a GDF are in many cases globally unique due to the extensive quality control of all incoming data and the predominant emphasis on sampling a spatially and stratigraphically discrete host rock unit. The strict control on 3D positioning of samples and the big data nature of the measurements lends these datasets ideally towards AI convolutional models.
The successful integration of subsurface data into a proof-of-concept AI-ready data framework can provide enhanced understanding of the behaviour, and ultimately the characterisation of GDF host rock units. Moreover, integrating selected representative data modalities and enabling mathematical operations across datasets allows to extract better insight and more value from the existing information/data by identifying underlying parameter correlations which are essential for model development, predictive analysis and digital workflows. To keep the scope realistic for the available budget and timeline, STRATA-AI will focus on a representative set of borehole-based test systems and selected data modalities, rather than attempting to cover all possible site-characterisation datasets. Additional modalities or multi-borehole extensions will be included only where suitable data and partner efforts allow.

Outcomes

The principal outcome of STRATA-AI will be an improved capability to organise, integrate and use heterogeneous subsurface characterisation data in a consistent, traceable and AI-ready way. This will enable borehole logs, images, text descriptions, laboratory measurements, hydraulic tests, geochemical data and, where available, downhole VSP seismic datasets to be jointly analysed within a coherent multimodal framework. 
Building on this capability, the WP will clarify the data-layer requirements needed for robust AI/ML analysis, including metadata, harmonisation, accessibility and quality control. STRATA-AI will also assess the applicability, added value and limitations of AI/ML models and workflows for site characterisation for deep geological disposal, from sensor-level data up to interfaces with geological, flow and transport modelling. Finally, the WP will deliver practical approaches for quantifying, reporting and communicating uncertainty in AI-assisted subsurface interpretation, enabling more transparent and defensible use of AI/ML outputs, accompanied by demonstration cases and by the release of pre-trained tools and reusable workflows, where feasible, for the EURAD community, including WMOs, TSOs and REs. 
These outcomes directly contribute to the EURAD SRA/Roadmap by generating scientific insight into AI-assisted subsurface-data integration, supporting innovation and optimisation of site-characterisation workflows, and producing reusable knowledge-management outputs for current and future EURAD applications.