Machine learning (ML) is reshaping how geoscientists interpret the subsurface. Traditionally, seismic interpreters rely on manual correlation between well data and seismic attributes — a time-consuming, subjective process that often depends on individual experience. SubsurfaceAI changes that. With built-in, user-friendly ML capabilities, geoscientists can now apply advanced data-driven techniques directly within their interpretation workflow — without being machine learning experts.

This article explains how SubsurfaceAI’s ML workflow can predict well top properties, such as reservoir thickness, from seismic attributes. Using the Manville Formation in the Western Canadian Sedimentary Basin (WCSB) as an example, we’ll walk through a practical case study built on the Blackfoot dataset, a classic benchmark for seismic-well integration and attribute-based reservoir characterization.

1. Introduction: Bringing Machine Learning to the Seismic Interpreter

In many geoscience organizations, machine learning remains confined to data science teams. But the real value of ML comes when it is accessible to seismic interpreters, who understand the geology and the data’s geological context.

SubsurfaceAI bridges this gap. It embeds ML directly into the interpretation environment, making it easy for geophysicists and geologists to:

• Predict reservoir properties such as thickness, porosity, or net-to-gross directly from seismic attributes.
• Perform rapid attribute selection and model validation.
• Visualize prediction maps in 2D and 3D, alongside seismic and well data.
• Iterate quickly, testing multiple ML methods and attribute combinations.

The workflow is unified across different reservoir representation objects — whether it’s a 3D reservoir grid, a strata-grid, an interval, or a horizon. This consistency means that once you learn the workflow, it applies everywhere.

2. The Case Study: Manville Formation and the Blackfoot Dataset

To demonstrate the workflow, let’s use a dataset many interpreters are familiar with — the Blackfoot 3D seismic dataset. It covers part of the Manville Formation in Alberta, which includes fluvial sandstone reservoirs producing from channel complexes.

The goal of this case study is to predict reservoir thickness at the Manville top using seismic attributes — particularly spectral decomposition attributes, which capture frequency-dependent variations often related to lithology and thickness.

3. Overview of the SubsurfaceAI Machine Learning Workflow

The ML workflow for predicting well top properties follows nine systematic steps:

1. Interpret and prepare the target reservoir horizon.
2. Verify or import the target well top property (e.g., thickness).
3. Convert the well top property to a well log.
4. Extract seismic attributes to the horizon.
5. Define training data.
6. Upscale well log values to the horizon.
7. Build and train ML models.
8. Analyze model performance and interpret results.
9. Predict property maps and visualize results.

Let’s explore each step in detail.

4. Step 1: Horizon Interpretation and Preparation

The workflow begins with interpreting the target reservoir horizon — in this case, the Manville top — after a proper seismic-well tie.

Performing a reliable tie ensures that the horizon represents the true geological boundary in depth or time. In the Blackfoot example, the horizon is tracked across the seismic volume. Any gaps or holes left during horizon tracking can be filled using horizon interpolation tools in SubsurfaceAI.

The interpreted horizon becomes the foundation for attribute extraction, data upscaling, and model training.

5. Step 2: Ensuring Correct Well Top Properties

Next, we confirm that the well top property we want to predict — for example, reservoir thickness — is correctly calculated or imported into the SubsurfaceAI project.

Thickness may already exist as a well top property in your database, or it can be computed from formation tops (e.g., difference between top and base of reservoir). The key is to ensure the property is consistent across all wells.

6. Step 3: Converting Well Top Properties to Logs

Machine learning requires well-based data to be treated consistently — typically as logs rather than discrete top properties. SubsurfaceAI simplifies this conversion.

Using the command “Convert to Log” on the Well Top Property node, you can generate a Thickness Log under each well’s log folder. This effectively transforms the top property into a depth or time-based curve, which is critical for later steps like upscaling and cross-plotting.

The new log ensures that the thickness information can be spatially correlated with seismic attributes extracted at the target horizon.

7. Step 4: Extracting Seismic Attributes

Once the horizon and well data are ready, the next step is to extract seismic attributes around the target horizon.

SubsurfaceAI supports extraction of a wide range of attributes — amplitude, phase, frequency, and more advanced ones like spectral decomposition attributes (instantaneous frequency, amplitude at specific frequencies, etc.). These attributes capture subtle variations in seismic response linked to lithological or thickness changes.

In our example, multiple spectral components are extracted to the Manville horizon. The resulting attribute volumes serve as the input features for the machine learning model.

8. Step 5: Defining the Training Dataset

The command “Make Training Data” initiates the process of linking well data and seismic attributes. You’ll be prompted to select the wells to include in training.

This step defines the training samples — points where both the target property (e.g., thickness) and input attributes (from seismic) are known. Typically, these are extracted at the well locations and projected onto the interpreted horizon.

SubsurfaceAI automatically organizes the training data into a structured dataset ready for machine learning.

9. Step 6: Upscaling Well Logs to the Horizon

The next step is to align the well log data (thickness) with the horizon representation.

Using “Upscale Well Log” in the Horizon Training Data tree node, select the thickness log created earlier. SubsurfaceAI then upscales the well log values to the horizon’s sampling grid — producing a set of training points that match the seismic attribute resolution.

The results appear under the Upscaled Well Log folder. This harmonization ensures that well and seismic data are on the same spatial framework.

10. Step 7: Building Machine Learning Models

This is the core of the workflow — constructing and training the ML model.

By selecting the “Build Machine Learning Model” command on the Upscaled Well Log node, you enter the model builder interface.

Model Setup

• Training Wells: Choose which wells to include. By default, all from Step 5 are used, but you can exclude some for validation or scenario testing.
• Input Attributes: Select the seismic attributes to be used as predictors. This step may involve trial and error — testing which combinations yield the best performance.

Model Type

SubsurfaceAI provides multiple ML algorithms:

• Neural Networks: Including feedforward, cascading, and functional fitting types.
• Decision Trees: Supporting Random Forest and XGBoost options.

You can also define the training/test/validation split — typically 80% for training, 10% for testing, and 10% for validation by default.

After pressing “OK,” the model begins training, producing a machine learning model object that appears in the Machine Learning Model folder.

11. Step 8: Performance Analysis and Model Explainability

Training is only half the story — understanding model performance is crucial.

SubsurfaceAI includes a suite of explainable machine learning tools to assess and interpret your model:

• Feature Importance Ranking: Quantifies which seismic attributes contribute most to the prediction.
• Partial Dependence Plots (PDPs): Show how individual attributes influence the predicted property while holding others constant.
• SHAP (SHapley Additive exPlanations) Plots: Provide a game-theory-based understanding of how each feature affects individual predictions.

These visualization tools transform the ML model from a “black box” into an interpretable system — allowing geoscientists to validate the geological sense of the predictions.

If the results are not satisfactory, you can iterate by returning to Step 7 to adjust:

• Input attributes
• Model type (switch between neural network and decision tree)
• Hyperparameters or sample selection

Each iteration can be completed quickly, fostering an interactive modeling workflow.

12. Step 9: Prediction and Property Map Generation

Once the model performs well, you can apply it to the entire horizon using the “Predict Property Grid” command on the ML model node.

The model predicts the target property (reservoir thickness, in this case) across the full 2D horizon surface. The resulting Predicted Property Map appears as a new node in the project tree.

This surface can be visualized in both map view and 3D window, alongside seismic data and wells, to evaluate geological plausibility and identify reservoir trends.

13. Step 10: Quality Control and Visualization

Visualization and QC close the workflow loop.

By overlaying the predicted property map with seismic attributes, structural interpretation, and well data, interpreters can assess:

• Whether predicted thickness aligns with known geological trends.
• If predictions capture expected depositional patterns (e.g., channels, bars, lobes).
• How uncertainty varies spatially.

SubsurfaceAI’s integrated 3D visualization environment makes this inspection intuitive, letting you toggle between ML prediction surfaces, attribute maps, and structural frameworks.

14. Iteration and Experimentation

Machine learning in geoscience is inherently iterative. Data quality, attribute selection, and algorithm choice all affect outcomes.

SubsurfaceAI enables rapid experimentation:

• You can create multiple ML models using different attribute sets or algorithms.
• Predictions are immediately available for visualization and comparison.
• All results are organized hierarchically, maintaining a clear workflow history.

This flexibility encourages interpreters to explore “what if” scenarios — for instance, testing whether low-frequency spectral attributes improve thickness prediction, or comparing Random Forest vs. Neural Network results.

15. Why SubsurfaceAI’s Workflow Matters

The value of SubsurfaceAI’s ML workflow lies in its integration, transparency, and usability:

By embedding ML directly into interpretation software, SubsurfaceAI democratizes advanced analytics — giving domain experts control over model design and interpretation.

16. Geological and Business Impact

Predicting reservoir properties such as thickness from seismic attributes has tangible geological and economic benefits:

• Improved reservoir characterization — better understanding of thickness variations helps identify sweet spots.
• Enhanced well planning — ML-predicted maps guide infill drilling or horizontal well placement.
• Reduced interpretation bias — consistent, data-driven results support team collaboration.
• Accelerated workflows — from data preparation to prediction, the process can be completed in hours instead of days.

For companies managing multiple fields or large 3D datasets, these efficiencies translate directly into cost savings and faster decision-making.

17. Summary

The SubsurfaceAI machine learning workflow provides a powerful yet accessible toolset for geoscientists to predict well top properties from seismic attributes. Through the example of predicting reservoir thickness in the Manville Formation, we demonstrated a repeatable process:

1. Interpret the reservoir horizon.
2. Prepare and convert well top properties.
3. Extract seismic attributes.
4. Define and upscale training data.
5. Build and train ML models.
6. Analyze model performance.
7. Predict and visualize results.
8. Iterate for improvement.

By combining seismic, well, and ML capabilities in one environment, SubsurfaceAI enables interpreters to focus on geology — not coding — while still leveraging the power of modern machine learning.

The future of subsurface interpretation is intelligent, integrated, and explainable — and SubsurfaceAI is leading that transformation.