⚡ electrical engineering

Combinatorial feature engineering and deterministic rock typing for seismic facies prediction in the mixed clastic-carbonate Asmari formation, SW Iran

This paper presents an integrated workflow combining deterministic petrophysical rock typing with a combinatorial feature engineering scheme that transforms raw seismic attributes into a discriminative feature space, significantly improving seismic facies prediction accuracy in the heterogeneous Asmari formation compared to conventional methods.

Original authors: Sam Azarpour, Igor Morozov, Iman Zahmatkesh, Mahmoud Poladzadeh

Published 2026-10-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sam Azarpour, Igor Morozov, Iman Zahmatkesh, Mahmoud Poladzadeh

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep beneath the earth, oil and gas are trapped in porous rocks that act like natural sponges, holding fluids within their tiny pores. To find these resources, geoscientists use sound waves sent from the surface to bounce off underground layers, creating a blurry, three-dimensional picture of the rock types below. However, in complex areas where different kinds of rock are mixed together, this picture often becomes confusing. The sound waves struggle to tell the difference between a rock that is full of oil and a rock that is solid and useless, even if they look very different under a microscope. This ambiguity makes it difficult to know exactly where to drill, leading to costly mistakes. The challenge is not just seeing the rock, but correctly identifying which specific types of rock will actually produce oil and which will not, especially when those rocks are layered in a chaotic mix of sand and limestone.

In the Mansuri oil field in southwestern Iran, a team of researchers tackled this problem in a formation known as the Asmari, which is a difficult mix of sandstones and carbonates. Instead of relying on standard computer methods that group rocks based on vague patterns, the team first used strict, physics-based rules to define exactly what they were looking for. They divided the rock into three clear categories: productive sand, porous limestone, and a "tight" background of rock that holds no oil. By grouping all the non-productive rocks into a single category, they removed the confusion that usually plagues these mixed systems. This created a clear, reliable target for their computer models to learn from, ensuring the machine was not trying to distinguish between rocks that sound the same but behave differently.

The researchers then faced the next hurdle: the raw data from the seismic waves contained too much noise and repetition to be useful on its own. Rather than simply picking the best few data points, they invented a new way to process the information. They used four different types of computer learning algorithms to look at the data in six different ways, creating a new set of twenty-four highly specialized features. Think of this as taking a blurry photograph and running it through several different filters that each highlight a specific detail, then combining those results into a single, crystal-clear image. This new, enriched set of data was then fed into a sophisticated neural network, a type of computer program designed to recognize patterns, to predict the rock types across the entire field.

The results of this new approach were striking. When tested against a well that the computer had never seen before, the new method correctly identified the rock types about 76 percent of the time. In contrast, the standard methods used in the industry only achieved accuracy rates of roughly 50 percent, essentially performing no better than a random guess in some cases. The old methods often failed to recognize specific rock types, either missing them entirely or confusing them with others. The new method, however, maintained a balanced performance, correctly identifying all three rock types with high reliability. Furthermore, the computer provided a confidence score for every prediction, allowing geologists to see exactly where the model was sure and where it was uncertain, adding a layer of trust to the final map.

The final maps produced by this new workflow showed a geologically sensible picture of the underground world. The predicted rock layers flowed smoothly and continuously, matching the known history of how the sand and limestone were deposited millions of years ago. The older methods produced fragmented, scattered results that did not make geological sense. By combining strict physical rules with advanced data processing, the researchers demonstrated that it is possible to see clearly through the noise of complex underground formations. This work offers a more reliable way to locate oil reserves, turning a blurry, uncertain image into a precise guide for exploration.

Technical Summary: Combinatorial Feature Engineering and Deterministic Rock Typing for Seismic Facies Prediction in the Asmari Formation

Problem Statement
Accurate reservoir rock typing in highly heterogeneous mixed clastic-carbonate systems, such as the Asmari Formation in SW Iran, presents a critical challenge in seismic reservoir characterization. Conventional workflows often rely on unsupervised statistical clustering to generate electrofacies, which frequently lack direct petrophysical meaning and may separate non-productive rocks that share similar fluid-flow potentials. Furthermore, raw seismic attributes often fail to resolve acoustic ambiguities inherent to these complex systems. Standard feature selection methods, such as linear correlation filters or stepwise regression, often prove inadequate for handling the redundancy and non-linear acoustic overlaps found in mixed lithologies, leading to high uncertainty in predicting productive "sweet spots" from heterogeneous backgrounds.

Methodology
The study proposes an integrated workflow that bridges deterministic petrophysics with ensemble-based feature engineering. The methodology proceeds through four primary stages:

Deterministic Litho-Reservoir Rock Typing:
Instead of unsupervised clustering, the authors define three specific reservoir classes based on strict petrophysical cut-offs derived from 18 wells in the Mansuri Oil Field:
- Class 1 (Pay Sand): Siliciclastic intervals with high porosity ( $\ge$ 12%) and sand volume ( $\ge$ 50%).
- Class 2 (Porous Carbonate): Productive limestone/dolomite intervals with effective porosity $\ge$ 8%.
- Class 3 (Tight Background): A unified class for all non-productive intervals (tight sands, tight carbonates, shales, anhydrites). This consolidation is crucial; it groups compositionally distinct but acoustically similar non-productive lithologies into a single background class to mitigate elastic non-uniqueness and prevent the model from forcing artificial distinctions between them.
Seismic Attribute Generation and Preconditioning:
The workflow utilizes 25 input attributes, including model-based acoustic impedance (AI), estimated porosity, and 23 additional seismic attributes (e.g., frequency, phase, amplitude, and similarity measures).
Combinatorial Feature Engineering (CFE):
To address the limitations of raw attributes, the authors introduce a combinatorial multi-classifier scheme. Rather than selecting a subset of raw attributes, they generate an enriched feature space:
- The three-class problem is decomposed into six binary splits (three one-versus-rest and three pairwise splits).
- Four distinct classifier algorithms (Regularized Logistic Regression, Support Vector Machines with RBF kernel, Shallow Feed-Forward Neural Networks, and Random Forests) are trained on each of the six splits using the original 25 attributes.
- The continuous discriminative outputs (scores or probabilities) from these 24 binary classifiers (4 algorithms $\times$ 6 splits) are retained as new "engineered" attributes.
- This results in a 24-dimensional engineered feature space, which is combined with the original attributes to form candidate sets.
Dimensionality Reduction and Classification:
Principal Component Analysis (PCA) is applied to reduce dimensionality while retaining 95% of the cumulative variance. The optimal feature set (the 24 engineered attributes reduced to 9 principal components) is fed into an Artificial Neural Network (CFE-ANN). The model's performance is evaluated against a conventional ANN (using raw features) and a commercial Probabilistic Neural Network (PNN) using a blind well (Well 096) and confidence-filtered facies sections.

Key Results
The evaluation on the blind well demonstrates the superiority of the proposed CFE-ANN workflow:

Accuracy: The CFE-ANN achieved an overall accuracy of 76.2%, significantly outperforming the conventional ANN (50.8%) and the PNN (49.2%).
Class Balance: Unlike the baseline models, which suffered from severe class-imbalance failures (e.g., the conventional ANN collapsed on Class 2 with a recall of 4.8%, while the PNN over-predicted Class 2), the CFE-ANN maintained balanced performance across all three classes (Recall: 84.2% for Class 1, 66.7% for Class 2, and 78.3% for Class 3).
Statistical Metrics: The CFE-ANN achieved a Cohen's kappa of 0.64 (substantial agreement) and a Matthews Correlation Coefficient (MCC) of 0.64, whereas the baseline models hovered near chance levels (kappa $\approx$ 0.25).
Geological Coherence: Spatial predictions along a seismic line revealed that the CFE-ANN produced laterally continuous, geologically plausible facies architectures consistent with the known interbedded sandstone-carbonate stratigraphy. In contrast, the PNN produced fragmented results, and the conventional ANN over-predicted sand bodies, suppressing carbonate intervals.
Confidence Filtering: The integration of classification confidence (based on the ratio of the maximum probability to the second-highest probability) allowed for the generation of filtered facies volumes, highlighting areas of high prediction reliability and exposing regions of ambiguity.

Significance and Claims
The paper claims that the primary contribution of this work is the successful integration of deterministic petrophysical constraints with a combinatorial feature engineering architecture to solve the "elastic ambiguity" problem in mixed reservoirs.

Target Definition: By consolidating non-productive lithologies into a unified background class based on rock physics rather than statistical clustering, the workflow ensures that target labels are seismically resolvable and physically meaningful.
Feature Transformation: The study demonstrates that transforming the seismic feature space via a multi-classifier ensemble (generating 24 derived attributes) is more effective than conventional attribute selection. This approach explicitly captures non-linear boundaries between rock types that raw attributes fail to resolve.
Reliability: The workflow provides a robust method for seismic facies prediction that overcomes class-imbalance failures common in complex reservoirs. The inclusion of confidence filtering offers a critical, often overlooked, measure of prediction reliability, allowing interpreters to distinguish between high-confidence geological trends and ambiguous zones.

The authors conclude that while the use of strict Boolean cut-offs for rock typing imposes crisp boundaries on continuous geological variations, the proposed framework significantly enhances prediction reliability in complex mixed clastic-carbonate systems compared to conventional machine learning baselines.

More like this