subsurfaceAI
A quick tour of ML in subsurfaceAI 2025

Machine Learning at the Core of subsurfaceAI 2025.1

subsurfaceAI’s 2025.1 release marks a watershed moment for digital subsurface workflows. Where geoscientists once stitched together disparate tools—sometimes literally drawing interpretations by hand—today they can call on a tightly integrated suite of machine‑learning (ML) models that accelerate every stage of exploration and production, from core photographs on the workbench to volumetric attributes in the reservoir model. Below, we unpack the seven cornerstone algorithms behind this release, show how each addresses a stubborn pain‑point in oil‑and‑gas interpretation, and explain why transparency and explicability matter just as much as predictive horsepower.

1.    Segment Anything Model (SAM) — Instant Core‑Photo Segmentation

Core photographs are treasure‑troves of textural and mineralogical information, but manual markup is painfully slow. subsurfaceAI integrates Meta’s Segment Anything Model (SAM) as the engine inside coreAI, automating the pixel‑perfect delineation of bedding, veins, vugs, and lithologic contacts. SAM’s prompt‑based interface lets a geologist click a single point on a feature—say, a stylolite—and instantly receive an accurate mask that snaps to boundaries even under non‑uniform lighting or broken core surfaces. Once segmented, downstream CNNs classify the masked regions and calculate properties such as grain‑size distributions or fracture apertures, feeding straight into petrophysical calibrations and SBED small‑scale heterogeneity modeling. The upshot: hours of manual tracing collapse into minutes, freeing core specialists to interpret rather than annotate.

2.    U‑Net Convolutional Neural Network — Faults & Geobodies at Seismic Scale

In seismic interpretation, the U‑Net architecture has become the gold standard for semantic segmentation, and subsurfaceAI’s Rapid Seismic Interpretation module leans heavily on a proprietary 3D U‑Net variant trained on thousands of labeled inlines, crosslines, and horizon‑slices. The encoder–decoder topology excels at capturing both global context (stratal dip trends) and fine‑scale discontinuities (throw‑related amplitude breaks). Interpreters can feed in full‑azimuth seismic volumes and—in a single forward pass—receive voxel‑level probability cubes for faults, channels, carbonate build‑ups, or salt bodies. Post‑processing tools convert these probability cubes into geocellular fault sticks, horizon patches, or geobody meshes, which can be exported to Petrel or SeisWare. By front‑loading machine segmentation, teams routinely shave weeks off prospect maturation schedules while improving consistency across interpreters and assets.

3.    Convolutional Neural Network (CNN) — Predicting Core‑Plug Properties

Laboratory measurements on core plugs (porosity, permeability, grain density, capillary pressure) remain the ground truth for reservoir characterization, but running every sample through the lab is costly and time‑intensive. subsurfaceAI couples high‑resolution core imagery with a purpose‑built CNN that learns the subtle links between visual texture and petrophysical response. After training on a subset of plugs with lab data, the model predicts properties for the remaining samples with surprising accuracy—often within the experimental repeatability of the lab itself. This bulk‑propagation of core measurements bridges the scale gap between sparse plug data and continuous log curves, enriching petro‑facies models and tightening saturation‑height functions without ballooning laboratory budgets.

4.    Fully Connected Neural Network — From Seismic Attributes to Continuous Logs

When seismic attributes (impedance, Vp/Vs, sweetness, spectral curvature) are mined for rock properties, relationships are rarely linear. subsurfaceAI therefore relies on a fully connected neural network (FCNN) that maps multidimensional attribute space directly to well‑log or reservoir‑property predictions. Because the architecture is “dense”—each neuron in one layer connects to every neuron in the next—it excels at capturing high‑order interactions among attributes (e.g., how instantaneous phase modulates amplitude‑versus‑offset behavior). Users typically train the FCNN on a handful of wells with both log and seismic data, then apply it across the entire volume to deliver 3D cubes of porosity, Sw, or even brittleness index. In mature fields, the same network can forecast production indicators such as flow‑unit transmissibility, informing perforation or chemical‑EOR strategies.

5.    XGBoost — Gradient‑Boosted Precision for Log Imputation & Facies Prediction

While deep nets dominate headlines, tree‑based ensemble methods still shine when data are tabular, noisy, and limited in volume. subsurfaceAI deploys XGBoost—an efficient gradient‑boosting library—for three recurring tasks:

  1. Log‑from‑Seismic Prediction: Where full‑resolution seismic is unavailable or quality varies, XGBoost handles missing attributes gracefully by learning from the most informative splits.
  2. Missing‑Log Reconstruction: In vintage wells, density or sonic curves are often absent. XGBoost trains on wells that do contain the missing curves plus common logs (GR, RHOB, NPHI) to fill gaps with probabilistic estimates.
  3. Facies Classification: By combining raw log values, derived ratios, and depth‑context features, the model assigns facies labels (e.g., fluvial channel sand vs. estuarine mud) that align with core descriptions and image logs.

Because gradient boosting sequentially corrects the errors of prior trees, it often outperforms standalone nets on smaller datasets—making it a practical workhorse in data‑poor, brown‑field scenarios.

6.    Random Forest — Robust Baseline & Uncertainty Workhorse

Complementing XGBoost is the Random Forest algorithm, another ensemble of decision trees but bagged rather than boosted. In subsurfaceAI’s IntegrationAI workflows, users employ Random Forests as a quick‑look baseline before committing to more compute‑hungry models. The algorithm’s built‑in out‑of‑bag error estimates provide an immediate sense of predictive confidence, and its resilience to overfitting makes it ideal for heterogeneous well logs gathered over decades. Common applications mirror those of XGBoost—log prediction, facies mapping, attribute‑to‑property transforms—but the forest’s probabilistic vote counts also translate naturally into uncertainty volumes, which reservoir engineers can propagate into Monte Carlo STOOIP or EUR calculations.

7.    Explainable Machine Learning — Trust Through Transparency

Predictive accuracy carries little weight if engineers cannot trust or understand the model. subsurfaceAI embeds a full suite of explainable ML tools that open the black box:

  • Partial Dependence (PD) & Individual Conditional Expectation (ICE) Plots: Visualize how changing a single seismic attribute (e.g., acoustic impedance) impacts the predicted porosity while holding others constant. ICE overlays reveal heterogeneity among wells.
  • Beeswarm Summary Plots (Shapley Values): Rank the attributes that most strongly drive each prediction. A single glance tells whether phase variance or envelope amplitude is steering the porosity cube.
  • Dependence Plots: Show how the influence of one attribute depends on another—crucial for spotting seismic‑quality artifacts or geologic crossover effects.
  • Feature‑Importance Bar Charts: Offer a high‑level audit trail for management presentations or regulatory submissions.

Behind the scenes, Shapley value calculations ensure the attributions are theoretically fair and additive. The net result: interpreters can defend model choices, reservoir engineers can quantify risk, and regulators gain a clear line of sight into AI‑assisted reserves evaluations.

8.    The Workflow Impact: From Siloed Tasks to a Cohesive ML Pipeline

What distinguishes subsurfaceAI is not merely the individual algorithms but the way they dovetail inside a unified pipeline:

  1. Data Ingestion & QC — Raw cores, logs, and seismic volumes enter a centralized repository that tags metadata and tracks provenance.
  2. Segmentation & Classification — SAM and U‑Net transform imagery into geologically meaningful objects (lithologic layers, faults, channels).
  3. Property Prediction — CNNs, FCNNs, XGBoost, and Random Forests convert those objects and their associated attributes into quantitative rock and fluid properties.
  4. Explainability Layer — Shapley‑based visuals validate each prediction step.
  5. Integration & Export — Results feed seamlessly into Petrel, SeisWare, or reservoir simulators, preserving uncertainty envelopes for downstream decision‑making.

Because every algorithm is containerized, users can swap models in or out without breaking the workflow or rewriting code. Training can even run on‑premise GPU clusters for data‑sensitive operators.

9.    Looking Ahead 

The 2025.1 release already propels subsurfaceAI well beyond conventional interpretation suites—yet our R & D roadmap is only gaining speed. We are:

  • Advancing transformer‑based models that learn directly from mixed 2D/3D data.
  • Developing physics‑guided networks that embed rock‑physics priors for more geologically consistent predictions.
  • Launching true two‑way data exchange with Petrel in Q2 2025, followed by the success of SeisWare project connector in 2025.1, eliminating the hand‑off friction that often stalls AI initiatives.

In short, subsurfaceAI’s machine‑learning stack turns decades of subsurface craftsmanship into repeatable, auditable, lightning‑fast workflows. Whether you are chasing frontier plays with sparse data or fine‑tuning mature assets for enhanced recovery, these algorithms deliver the predictive clarity—and the explainability—that modern E & P demands. The future of subsurface interpretation is here, and it is intelligently automated.

Looking ahead to version 2025.2 (July 2025), expect a deeper integration of large language models (LLMs) that will further revolutionize subsurface workflows—plus a host of user‑driven enhancements and a quantum leap in usability, including seamless integration with existing Petrel projects. Stay tuned!