AI-Driven Fusion of Geophysical and Geochemical Data Enhances Subsurface Characterization

Summary

A review article synthesizing the AI-based multimodal fusion of geophysical and geochemical data for subsurface characterization in mineral exploration, geothermal energy, CO₂ storage, and radioactive waste domains. The study is structured around six independent dimensions and critically addresses gaps in uncertainty propagation and validation.

Why it matters

Subsurface characterization underpins high-risk decision-making in energy, mining, and environmental safety. However, most current AI applications treat multimodal fusion as a technical convenience rather than a scientific problem requiring explicit physical consistency. The review identifies concrete failure modes often overlooked by practitioners—scale mismatch, modality dominance, and validation leakage. By distinguishing aleatoric, epistemic, methodological, and ontological uncertainties, the article calls for physics-informed and uncertainty-aware models. This approach offers a structured roadmap to make subsurface AI interpretable and decision-focused.

Highlights

• The literature is 'conceptually siloed.' Most studies fail to address scale mismatch, spatial dependence, or causal links between modalities when fusing geophysical and geochemical properties.
• Four distinct types of uncertainty (aleatoric, epistemic, methodological, and ontological) are formalized, with emphasis on how they propagate through multimodal workflows.
• New priorities highlighted include physics-informed learning with inter-modal physical consistency, constrained generative simulation, and foundation model representations for subsurface data.

Critical take

As a purely descriptive review, the article does not provide original experiments, benchmarks, or quantitative comparisons to validate its own six-dimensional classification. The proposed framework remains an untested regulatory lens. Calls for 'physics-informed learning' and 'constrained generative simulation' remain at an idealistic level due to the lack of methodological recipes or failure-mode analyses to concretize these proposals.