Subsurface intelligence: mineral exploration, seismic interpretation, geothermal & CO2 storage siting

Finding ore bodies, hot reservoirs or safe CO2 stores means inferring what lies kilometres below ground from sparse, noisy and heterogeneous data. ML fuses seismic volumes, potential-field geophysics, geochemistry, well logs and digitised legacy reports to propose drill targets, interpret faults and horizons in 3D seismic, and characterise reservoirs. The benefit is fewer dry holes and faster screening, not replacement of the geoscientist: every target and interpretation is reviewed and drilled or injected only after expert and regulatory sign-off.

Problem & value

• Today: geologists interpret horizons and faults manually in seismic workstations, build prospectivity maps by weighted overlay in GIS, and dig through paper or PDF reports from decades of earlier campaigns. Discovery rates in mature mining districts are low, drilling is expensive and interpretation is slow and subjective.
• AI value: (1) faster, more consistent seismic interpretation (fault/horizon/salt/channel segmentation); (2) data-driven prospectivity mapping that combines many weak signals; (3) extraction of knowledge from legacy reports and logs; (4) quicker screening of many candidate geothermal or storage sites with quantified uncertainty.
• EU drivers: the Critical Raw Materials Act (Reg. (EU) 2024/1252) sets 2030 benchmarks of 10 % of annual EU consumption from domestic extraction, 40 % processing and 25 % recycling, with no more than 65 % of any strategic raw material from a single third country, and fast-tracks Strategic Projects [1]. The Net-Zero Industry Act (Reg. (EU) 2024/1735) requires an EU CO2 injection capacity of 50 Mt per year by 2030, placing obligations on oil and gas producers [2]. Geothermal heat and power are expanding in several member states.
• Evidence of use: KoBold Metals (US) runs AI-driven exploration programmes; Zanskar (US) applies ML to find hidden geothermal systems; Equinor and partners operate Northern Lights, the first cross-border commercial CO2 transport and storage service, off Norway [3]. No numeric discovery-rate improvements are claimed here; published vendor claims should be validated against your own back-testing.

Variants

• Mineral exploration (critical raw materials): prospectivity modelling for lithium, copper, nickel, cobalt, REE, tungsten; targets in Fennoscandia, Iberia, Balkans, Greenland. Greenfield and brownfield (near-mine) variants.
• Seismic interpretation & subsurface characterisation: 3D fault/horizon/salt/stratigraphy segmentation, seismic inversion to impedance and porosity, well-tie, reservoir characterisation for hydrocarbons, geothermal, CCS and hydrogen storage.
• Geothermal siting: combine heat-flow, gravity/magnetotellurics, fault maps and well data to rank hot-rock and hydrothermal prospects, including repurposing of oil and gas wells.
• CO2 / hydrogen storage siting & monitoring: saline-aquifer or depleted-field capacity screening, caprock integrity, 4D seismic time-lapse for plume monitoring (MMV).
• Public geological surveys: national surveys (BGR, GTK, BRGM, GEUS, TNO) building open prospectivity layers and digitising archives.
• Legacy-report mining: OCR plus LLM extraction over scanned assessment reports and drill logs.

How it works

1. data intake (SEG-Y, grids, assays, LAS logs, scanned reports)
2. QC, harmonisation, CRS alignment, depth/time conversion
3. feature engineering / foundation-model embeddings
4. task models: 3D U-Net/ViT segmentation | prospectivity classifier | inversion | LLM extraction
5. uncertainty (ensembles, Monte Carlo dropout, conformal intervals)
6. geoscientist review in interpretation / GIS software
7. target ranking
8. drilling, injection or survey decision (human, with regulatory permit)
9. new data (drill results, 4D seismic) feed back as labels

• Seismic segmentation: 3D convolutional or transformer models trained on labelled volumes; weak labels and synthetic data reduce interpreter effort. Foundation models pretrained on large unlabelled seismic archives are fine-tuned for faults, horizons and facies.
• Prospectivity: gradient-boosted trees, random forests or neural nets on stacked geophysical, geochemical and geological layers. The hard issue is extreme class imbalance (few known deposits) and spatial autocorrelation, so spatial cross-validation is essential.
• Physics-informed and hybrid models: ML inversions constrained by wave-equation or flow simulation (see SCI-005 surrogates) to keep outputs geologically plausible.
• Legacy knowledge: OCR/VLM parsing of scanned reports, LLM extraction of grades, depths and coordinates into a geodatabase, then RAG for geologists (see DOC-014).
• Human-in-the-loop: interpreters correct predicted faults and horizons; chief geologists approve drill targets; storage operators and regulators approve injection.

Data preparation

Subsurface data lives in workstation project databases (Petrel, Leapfrog), SEG-Y tapes and file shares, well-log archives (LAS/DLIS), assay spreadsheets, GIS projects and, above all, paper and scanned legacy reports and maps, often held by external contractors or authorities. Typical gaps: inconsistent coordinate reference systems and depth/time domains, missing well headers, unlabelled scans, duplicates between projects, proprietary formats, licence restrictions on third-party data and sparse labelled examples.

Preparation steps: inventory surveys, wells and reports with owners (PREP-001, PREP-004); scan and OCR legacy reports (PREP-006, DOC-014); restructure project folders and add metadata (PREP-009, PREP-010); clean assay tables (PREP-011); ingest into an OSDU-style catalogue (see DATA-011); extract labels from existing interpretations (PREP-020); define a glossary of lithology and facies terms (PREP-012). Tools: OSDU, segyio, lasio, GDAL, Tesseract or modern OCR/VLM parsers.

Expect 50-70 percent of effort in digitisation, harmonisation and quality control. GUIDE-003 gives a readiness check.

Implementation paths

Which path for whom

• Enterprise: B - OSDU-based cloud platforms in EU regions give a common data model and elastic compute for seismic volumes, without building a data centre.
• Developer: B - cloud APIs and managed GPU batch plus open geoscience libraries speed up prototyping; switch to C only for very large data volumes.
• Public: C - geological surveys and universities favour open tools and sovereign EU or German infrastructure for public-interest data and procurement rules.

Path A — Buy (off-the-shelf)

• Seismic and reservoir platforms with AI modules: SLB Petrel/DELFI, Halliburton Landmark DecisionSpace 365, Petrosys, Bluware (interpretation and deep-learning fault picking), Cegal Prizm (Norwegian, EU), Geoteric (UK, AI fault/horizon).
• Mineral exploration: Seequent Leapfrog (Bentley) and Oasis montaj for 3D modelling and geophysics; Datarock (core-image AI); Fleet Space (ExoSphere, satellite-enabled passive seismic); Minalyze (Sweden, EU) core scanning.
• Geothermal / CCS: Zanskar (US, geothermal developer), Eavor/Fervo (engineering developers, not software).
• Choose A when: you already own a workstation licence stack and want AI features inside it.

Path B — Build on cloud APIs

• Platforms: OSDU-based data platforms (OSDU Forum standard; Microsoft Energy Data Manager, AWS and Google Cloud implementations, SLB Delfi, Cegal) offer EU regions and a common data model [4].
• Models: managed ML on EU regions (Azure ML, Vertex AI, SageMaker) for training segmentation and prospectivity models; frontier LLMs (Claude, Gemini, GPT, Mistral as EU provider) with structured outputs for report extraction; Mistral OCR or Azure Document Intelligence for scans.
• Orchestration: Python with segyio/Devito, GeoPandas, Dask; MLflow; Databricks or Snowflake geospatial; ArcGIS or QGIS for review.
• Choose B when: you need custom prospectivity or extraction logic and can use EU-hosted cloud.

Path C — Self-host (open source)

• Seismic: open-source toolkits for deep-learning interpretation (e.g. open benchmark repos, OpendTect with ML plugins (dGB, NL), Madagascar, Devito and Fatiando a Terra; Open Subsurface Data Universe OSDU community builds.
• Geophysics inversion: SimPEG, pyGIMLi, GemPy (implicit geological modelling), PyVista.
• Prospectivity: scikit-learn, XGBoost/LightGBM, PyTorch; EIS Toolkit from the EU-funded EIS Horizon project for mineral prospectivity [5].
• Report mining: Docling or PaddleOCR with an open-weight LLM (Qwen, Mistral, Llama) on vLLM.
• Hardware: HW3 (one 48–141 GB GPU) for 3D volume training and inference at survey scale; smaller surveys fit HW2; large reprocessing needs HPC (HW4).

Path D — Custom training

Worthwhile when you own proprietary seismic or drill-hole archives: fine-tune a seismic foundation model on your volumes, or train a deposit-type-specific prospectivity model. Needs labelled interpretations (faults, horizons), hundreds of drill holes or known occurrences and a data engineering team; models do not transfer well across basins or deposit styles.

State of the art (2026)

• Seismic foundation models pretrained on large unlabelled volumes outperform task-specific U-Nets on fault and horizon benchmarks when labels are scarce; public benchmarks are still small and mostly synthetic or from few basins (e.g. F3 Netherlands, Parihaka NZ, SEAM).
• Generative AI: LLM copilots for geoscience data search, natural-language access to OSDU stores and report summarisation are in early production; diffusion models are used for seismic denoising and super-resolution.
• AI-led exploration companies (KoBold Metals, Zanskar) combine proprietary data acquisition with ML target generation; success is measured by discoveries, which take years to confirm.
• Limits and failure modes: extreme data scarcity for deposits; spatial leakage inflating validation scores; domain shift between basins and acquisition vintages; ML faults that are geologically implausible; hallucinated values in LLM extraction from poor scans; amplified bias toward already-explored areas (sampling bias). Predictions must carry uncertainty and be tested by blind drilling or hold-out regions.

Shared resources & combinations

• Shared stack: scientific-fm, simulation, ocr-docparse and vector-db are reused in SCI-005 (simulation surrogates), DOC-014 (document parsing for RAG) and SCI-011 (AI scientist agents).
• Combinations: satellite and geological-map layers from VIS-021; InSAR ground motion for mine and storage-site monitoring (VIS-061); geospatial analytics DATA-011; climate and physical-risk context SCI-010, PRED-036.
• GPU/HPC: the same HW3/HW4 nodes serve seismic training, surrogate simulation and LLM inference.

Effort, cost & KPIs

• Path A: weeks to enable modules; licence cost dominates.
• Path B: 3–9 months for a first prospectivity or interpretation pipeline plus data platform work; data cleaning is the largest cost driver.
• Path C/D: 6–18 months with geoscientists and ML engineers; HPC or GPU costs and data acquisition (new surveys, drilling) dwarf model costs.
• KPIs: interpreter hours per seismic volume; fault/horizon agreement with expert (IoU, Dice); hit rate of drilled targets versus baseline; cost per metre drilled per discovery; share of legacy reports digitised; storage-site screening time per candidate; calibration of uncertainty.

People & change

Involve the geoscience lead and the interpreters who will review every map and horizon from the start; they define acceptance criteria and supply labels. Management decides about data-sharing with partners and licensors. IT provides storage, GPU and workstation integration; the data owners (survey archive, contractors) must release data and clarify licences. Legal and compliance cover mining and storage permits, and the authorities and affected communities as stakeholders for CO2 storage and geothermal siting.

Data protection: mostly no personal data, so a DPIA is usually not needed; involve the data protection officer (DPO) only if landowner or staff data enter the system.

Works council: no in the typical case, because the system analyses geology, not employees; it becomes relevant if usage logs allow performance monitoring (section 87(1) no. 6 BetrVG), and the council should be informed about the planned AI-supported work processes (section 90(1) no. 3). Public bodies inform the staff council.

AI literacy: the AI Act Art. 4 asks for measures that support staff AI literacy; teach uncertainty maps, extrapolation limits and how to challenge a model. Pilot with one basin or prospect and compare against a blind interpreter. Concern: "the model replaces the interpreter" - interpreters remain accountable for drill and storage decisions.

Guides: GUIDE-009, GUIDE-008, GUIDE-003.

Risks, regulation & compliance

• EU AI Act: typically minimal risk; geoscience analytics is not an Annex III use case. It becomes high-risk only if an AI component is a safety component of critical infrastructure (e.g. automated control of a storage-site injection or a gas network); no Art. 5 prohibited practice applies. AI literacy (Art. 4) applies to staff.
• GDPR: mostly no personal data; landowner records, permit applications or community consultation data may contain personal data.
• Sector rules: CO2 Storage Directive 2009/31/EC (site characterisation, permits, monitoring) [6]; NZIA; CRMA permitting and Strategic Project criteria; national mining law, EIA Directive and Natura 2000 constraints; EU Data Act for industrial data sharing; licensed seismic data contractual restrictions on AI training.
• Technical and commercial risks: over-trust in black-box predictions, induced seismicity in geothermal stimulation, CO2 leakage; community and environmental licence-to-operate. Mitigation: documented validation, independent peer review, human sign-off on every drill or injection decision, traffic-light monitoring and public engagement.

Getting started

1. Inventory data holdings (seismic, wells, assays, reports), formats and licence limits; stand up an OSDU or geodatabase foundation.
2. Pick one well-defined task, for example fault segmentation on one survey or prospectivity for one deposit type in one district.
3. Digitise legacy reports with OCR plus LLM extraction and validate a sample against ground truth.
4. Build a baseline with spatial cross-validation and blind hold-out areas before using larger models.
5. Present ranked targets with uncertainty to geoscientists; log accepted and rejected suggestions and feed drill results back.

Sources

1. Regulation (EU) 2024/1252 (Critical Raw Materials Act) — https://eur-lex.europa.eu/eli/reg/2024/1252/oj
2. Regulation (EU) 2024/1735 (Net-Zero Industry Act) — https://eur-lex.europa.eu/eli/reg/2024/1735/oj
3. Northern Lights CCS project — https://norlights.com
4. OSDU Forum (Open Group) — https://osduforum.org
5. EIS Toolkit (Horizon Europe EIS project) — https://github.com/GispoCoding/eis_toolkit
6. Directive 2009/31/EC on geological storage of CO2 — https://eur-lex.europa.eu/eli/dir/2009/31/oj
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. KoBold Metals — https://www.koboldmetals.com
9. SimPEG open-source geophysics — https://simpeg.xyz