The state of geoscience today: Challenges & opportunities

The rise of ai in subsurface workflows

Subsurface characterization has long relied on expert interpretation of subsurface rock samples, well logs, and seismic data. As datasets have become larger, and domain experts increasingly need deeper insight into complex geological heterogeneity, machine learning (ML), especially deep learning, has become essential for capturing nonlinear relationships and extracting mean­ingful insights at scale. ML is now central to image data interpretation, log interpretation, seismic interpretation, and reservoir property estimation from seismic (Figure 1).

Why AI now?

Four converging factors enable the rapid adoption of AI in geoscience today:

• Massive growth in data volumes: Modern seismic surveys generate large-scale 3D subsurface volumes, enabling direct training of deep neural networks on volumetric data.
• Maturity of deep learning architectures: Transformers, diffusion models, and contrastive vision-language frameworks are now applied to seismic interpretation.
• Foundation models enabling cross-survey generaliza­tion: New models provide reusable seismic features across basins and data vintages, increasing robustness.
• Industrial demand: Operators seek faster, more accu­rate, and integrated workflows to improve well place­ment, reduce costs, and accelerate field development.

The energy industry generates vast amounts of subsurface data: Seismic surveys, well logs, core samples, rock data, and geological reports. Much of this information, however, is still underutilized. The main barriers are fragmented, discipline-specif­ic workflows, heterogeneous data formats, limited cross-disciplinary interoperability, and slow, manual interpretation processes.

Advances in AI now address these challeng­es by integrating seismic data, text, geological maps, and numerical logs within unified analytical frameworks.

AI transformations across geoscience disciplines

AI is delivering order-of-magnitude improvements across key subsurface workflows in the energy industry, fundamentally changing both the speed and quality of interpretation.

In seismic interpretation, modern AI-driven platforms such as EarthNET enable horizon and fault interpretation to be completed 20 – 200 times faster than traditional manual approaches. By automating pattern recognition and consistently applying learned geological features across large seismic volumes, these systems significantly reduce interpretation cycles while improving reproduc­ibility and reducing interpreter bias.

AI is also accelerating reservoir characterization by enabling the rapid and coherent integration of diverse data types, including core and core-plug measure­ments, cuttings data, well logs, and seismic volumes. This integrated approach supports the construction of consistent, multi-scale earth models that capture both fine-scale rock properties and larger-scale structural trends, leading to faster insight generation and more robust subsurface understanding.

Property prediction from well data

AI has now, for several years, been accelerating log editing and infill, as well as prediction of target properties based on available well logs. The ability to do this type of work efficiently and at scale has enabled basin-wide missed pay studies that have revealed overlooked exploration opportunities.

Ear thNET can now integrate core photos (Figure 1B), cuttings photos (Figure1C), thin sec­tions, logs (Figure1D), and geochemical data such as X-ray fluorescence (XRF) and X-ray diffrac­tion (XRD) for more reliable classification of facies and prediction of target properties in wells.

This multi-modal approach enables petrologists and petrophysicists to characterize samples and interpret well logs faster, with higher consistency, while using all the available data.

The interpretation of geological timelines and age of samples is now possible using computer-vi­sion-based fossil identification (Figure 1A) and biozone classification. Emerging, multi-modal models will soon merge fossil-derived timelines with well logs and seismic for chronological consistency, facilitating the creation of sequence-stratigraphic models based on all available data.

Property prediction from seismic

EarthNET users can leverage machine learning meth­ods that integrate well data with seismic information, enabling more accurate and scalable prediction of key subsurface properties. By jointly analyzing logs and seismic partial stacks, these approaches can infer elastic properties, lithofacies distributions, porosity, fluid satura­tion (Figure 1E), and other critical reservoir parameters across large spatial domains. This integrated learning reduces uncertainty while extending well-based insights into areas with sparse direct measurements. More broadly, machine learning represents a significant advancement in the way seismic data is mapped to underlying rock and fluid properties.

Traditional empirical and physics-based workflows are increasingly complemented by, and combined with, data-driven models that can capture complex, nonlinear relationships in the subsurface.

Structural seismic interpretation

The efficient AI-assisted workflows for fault detection (Figure 1F), horizon tracking, geobody delineation, and seismic facies classification facilitate the construction of the structural and stratigraphic models that host the property volumes created from the combination of well and seismic data. As a result, ML-based techniques are improving subsurface characterization and prediction across a wide range of applications, including hydro­carbon exploration and production, CO₂ storage and monitoring, groundwater assessment, and geother­mal resource development.

Leveraging foundation models for seismic interpretation

Foundation models are large, general purpose machine learning models trained on massive amounts of diverse data and designed to be adaptable to many downstream tasks (Figure 2). The emergence of such models reduced a significant bottleneck for the efficiency of AI-assisted seismic interpretation, i.e. the strong dependence of an­notated data.

Self-supervised learning is the underlying training paradigm where labels are automatically de­rived from the data itself, rather than manually annotated by humans.

Seismic foundation models have demonstrated superior per­formance across seismic denoising and seismic interpretation tasks compared to traditional deep learning methods. Such models can generalize across surveys without retraining. With “human in the loop” workflows, these models can reduce interpretation cycle times (up to 200 times), for tasks such as fault, horizon and geobody interpretation.

Physics-informed neural networks

Physics-informed neural net­works (PINNs) incorporate gov­erning physical laws directly into neural networks and can therefore leverage the know-ledge accumulated by research­ers and domain experts. We no longer have to choose between using a physics-based approach or a learning-based approach, but can now combine both using PINNs and get the best of both worlds. Recent work in Earth Science Analytics shows that by combining physics-based methods and learning-based methods in physics-informed foun­dation models we can achieve improved inversion robustness, better generalization across geologies and more physically consistent predictions.

Physics-informed AI bridges data driven ML with geophysical theory, transforming seismic inver­sion workflows.

Integrating disciplines in the era of foundation models and physics-informed neural nets

Within Earth Science Analytics, foundation models and physics-informed neural networks (PINNs) are already playing a central role in the integration of diverse sub­surface data. These approaches are currently being applied to jointly analyze well-log measure­ments and seismic data, enabling the construction of coherent, high resolution three-dimensional volumes that describe elastic pa­rameters and reservoir properties. By embedding physical constraints directly into the learning process, these models allow data driven insights to remain consistent with established geophysical and geological principles.

Looking ahead, our ambition is to move beyond pairwise data integration toward a deeper and more comprehensive unification of geoscience disciplines. This effort will focus on the development of multi-modal geoscience founda­tion models capable of ingest­ing, correlating, and reasoning across multiple heterogeneous datatypes simultaneously.

In addition to seismic volumes and well logs, these models will incorporate visual and petro­physical information derived from core images, cuttings images, and thin-section microscopy. Quantitative laboratory measurements, including core-plug data and X-ray fluores­cence (XRF) analyses, will further en­rich the learning process by providing direct constraints on rock properties and geochemical composition.

By enabling joint learning across these complementary data modalities, we aim to create integrated subsur­face representations that capture structural, elastic, petrophysical, and compositional variability in a unified framework. This multi-modal approach is expected to reduce uncertainty, improve interpretability, and ultimately support more robust reservoir char­acterization and decision-making in complex geological settings.

The path toward autonomous workflows

Agentic AI is a promising direction for pursuing our vision of multi-modal integration of geoscientific data (Figure 3). Agentic AI leverages autonomous LLM-driven systems capable of executing multistep geoscientific workflows and creat­ing self-improving multi-modal AI models. We believe that this par­adigm can open opportunities for better utilisation of all geoscientific datatypes, further automation of QC, and better propagation of un­certainty through multi-disciplinary integrated workflows.

We see a path to a future where agents can manage continuous model updates, run cross-disciplinary workflows, autonomously refine reservoir models, and support de­cision-makers with insights stemming from all available data. Advance­ments in artificial intelligence are transforming subsurface sciences, enabling the creation of high-fi­delity digital twins that function as dynamic, continuously updated representations of the Earth. These living models integrate seismic, well, production, geological, and petrophysical data into a unified framework that evolves as new information becomes available. By combining multi-modal learning, seismic foundation models, phys­ics-informed neural networks, and agentic AI, these systems support real-time scenario testing, uncer­tainty quantification, and automated decision-making. This convergence marks a fundamental shift away from fragmented, discipline-spe­cific workflows toward holistic, autonomous subsurface modelling environments. As these technologies mature, they empower geoscientists to derive deeper insights, accelerate interpretation cycles, and achieve more accurate and integrated Earth models – ushering in a new era where AI is central to how we understand, predict, and manage the subsurface.