Geophysical, hyperspectral, and remote-sensing surveys generate increasingly large and complex datasets that are difficult to interpret quickly. Turning those data into actionable information can be slow, computationally demanding, and difficult when survey geometries, geologic conditions, and target types vary.

Researchers at Pacific Northwest National Laboratory have developed a complementary suite of high-performance computing and artificial intelligence technologies for geophysical imaging, subsurface characterization, and mineral exploration. The suite combines parallel Geophysical Electromagnetic Modeling and Inversion of Natural and Induced (pGEMINI) sources, a scalable physics-based framework for 3D electromagnetic simulation and inversion, with four AI capabilities for rapid 3D electromagnetic inversion, near-real-time 1D inversion of airborne and drone electromagnetic data, mineral identification from hyperspectral imagery, and multimodal 3D inversion and anomaly characterization using ground-penetrating radar, and magnetic and electromagnetic data.

APPLICABILITY
The technologies can be deployed independently or combined within a broader characterization workflow to screen large datasets, guide field decisions, and develop high-resolution 3D models.

Potential applications include: 

• Critical Mineral and Material Exploration.
• Geothermal and Geological Exploration and Surveys.
• Environmental and Groundwater Surveys.
• Infrastructure and Engineering Site Characterization.
• Drone-Enabled Field Surveys.
• Other field programs that benefit from rapid interpretation or detailed 3D subsurface information.

AI models can provide rapid initial interpretations, anomaly screening, or updates during field operations. pGEMINI can provide detailed forward simulation and regularized 3D inversion for complex electromagnetic problems and can generate physics-based synthetic datasets for training and evaluating AI models. Together, these capabilities connect data acquisition, rapid AI-assisted analysis, detailed numerical modeling, and decision support without treating all geophysical problems as a single method or workflow.

For more information, please contact commercialization@pnnl.gov.

FEATURED TECHNOLOGY SUITE

pGEMINI: Scalable 3D Electromagnetic Modeling and Inversion
IPID 33887

pGEMINI is a parallel computational framework for 3D electromagnetic forward simulation and inversion in complex subsurface environments. It combines finite-element discretization on flexible unstructured meshes, scalable parallel solvers, sensitivity calculations, regularization, and gradient-based optimization within an integrated physics-based framework.

pGEMINI can represent irregular topography, complex geologic structures, spatially variable electrical properties, localized conductive or resistive targets, and complex source-receiver geometries. It supports active-source electromagnetic configurations, including controlled-source, ground, airborne, borehole, and drone electromagnetic—as well as natural-source magnetotelluric—data. The framework computes electromagnetic fields and model sensitivities and uses them to estimate 3D electrical-conductivity distributions from measured data.

In addition to direct geophysical interpretation, pGEMINI can generate realistic physics-based synthetic datasets for survey design, method evaluation, and development of AI-based inversion tools.

AI-Based Rapid 3D Electromagnetic Inversion
IPID 33337

This technology uses deep learning to estimate 3D subsurface electrical-conductivity structure from electromagnetic measurements. Training datasets pair diverse 3D conductivity models with synthetic electromagnetic responses generated using physics-based numerical simulation, including pGEMINI.

Once trained for a supported survey configuration, the AI model learns relationships between the measured electromagnetic response and the underlying 3D conductivity distribution and can generate a volumetric conductivity estimate in seconds. The resulting model can support rapid target screening, field assessment, or initialization of a subsequent physics-based inversion.

The technology is intended to complement, rather than universally replace, conventional inversion. Its performance depends on the acquisition geometry, data characteristics, training-model coverage, noise conditions, and similarity between field conditions and the training data.

Near-Real-Time 1D Inversion of Airborne and Drone Electromagnetic Data
IPID 33038

EM1DNet is an AI-based technology for rapid, sounding-by-sounding 1D inversion of frequency-domain airborne and drone electromagnetic data. It converts multifrequency electromagnetic measurements into layered-earth electrical-resistivity or conductivity profiles without running a conventional iterative inversion for every sounding.

Because each sounding is processed independently, the method scales efficiently to large surveys and can support interpretation during or shortly after acquisition. Field teams can use the results to assess broad electrical-property variations, refine coverage, and prioritize follow-up areas.

The technology provides 1D layered-earth estimates. Laterally constrained or fully 3D interpretation can be performed separately using pGEMINI or another appropriate physics-based workflow.

HyDeMiC: Hyperspectral Deep Learning-based Mineral Classifier
IPID 33348

HyDeMiC is a deep learning technology that identifies and classifies minerals using hyperspectral signatures. A convolutional neural network learns diagnostic spectral patterns and assigns mineral classes to hyperspectral observations, accelerating analysis of datasets that would otherwise require extensive and time-consuming manual interpretation.

HyDeMiC supports mineral mapping, exploration-target evaluation, and site characterization. Its classification results provide mineralogical information that can be evaluated alongside geophysical interpretations and other geologic evidence.

Near-Real-Time Multimodal 3D Subsurface Digital Twin for Inversion and Anomaly Characterization
IPID 33924

This AI-based digital twin supports rapid 3D characterization from GPR, magnetic, and electromagnetic data. It combines modality-specific deep learning for 3D inversion with anomaly detection, localization, segmentation, and classification.

For supported data types and configurations, the framework can generate 3D predictions in seconds to minutes. It can use additional measurements to update its subsurface representation and uncertainty estimates, helping field teams determine how new data change the interpretation and where further investigation is most valuable.

The capability supports anomaly localization, target characterization, survey adaptation, and prioritization of follow-up measurements. It is broader than the dedicated 3D electromagnetic inversion technology because it addresses multiple geophysical modalities and includes detection and classification.