Hybrid AI model nears perfect oil-presence prediction in synthetic reservoirs

A deep-learning framework that combined three ways of analyzing reservoir data achieved 98.74% accuracy when identifying oil-bearing records in a held-out synthetic test set, according to a study of 2,776 reservoir samples. The Synergistic Oil Recovery-Net integrated probabilistic learning, contextual self-attention, sequential sparse attention and a sample-specific fusion mechanism. On the untouched test set of 556 records, it correctly classified 549 cases, including 152 true positives and 397 true negatives, with four false positives and three false negatives. Its ROC-AUC was 0.9828 and precision-recall AUC was 0.9865, indicating strong discrimination within the dataset’s distribution. These results are promising as a technical demonstration, but they do not establish performance in an operating oil field because every record came from a single synthetic dataset.

The underlying public dataset, called the Synthetic Oil Reservoir Prediction Dataset, contained seven original predictors: rock type, porosity, permeability, trap type, seismic score, proximity to an oil field and estimated reservoir depth. Its binary outcome recorded whether oil was present or absent. The full dataset included 2,004 negative records, or 72.19%, and 772 positive records, or 27.81%. Before modeling, the researchers used a stratified split with random seed 42 to create a development set of 2,220 records and a held-out test set of 556, preserving class proportions in both subsets. The test records were not used for imputation, feature engineering, class balancing, feature selection, hyperparameter tuning, model selection or probability-threshold selection.

That separation was central to the study’s leakage-controlled design. Within each cross-validation fold, preprocessing parameters were estimated only from the training portion. Missing numerical values were replaced with the training-fold median, while missing categorical values were filled with the training-fold mode. Data-consistency checks examined duplicated records, non-finite values, negative permeability, porosity outside the dataset’s admissible range and invalid category labels. Numerical anomalies were screened with Isolation Forest, and flagged training observations were winsorized using the first and 99th percentiles estimated from that fold; flagged validation and test records were not automatically removed. The analysis was also repeated after excluding flagged training observations as a sensitivity analysis.

SMOTE-NC was then applied only to the training portion of each cross-validation fold, after imputation and feature generation but before final one-hot encoding. It created additional minority-class examples using nearest minority-class neighbors, with synthetic categorical values assigned from the modal category among those neighbors. Neither validation folds nor the held-out test set were oversampled. Consequently, all reported validation and test metrics were calculated using the original class distributions. The final model was fitted to the complete development set only after preprocessing, feature selection and hyperparameter choices had been made, and the held-out set was evaluated once as a final test.

The researchers also constructed target-independent engineered descriptors before fitting the model. These included a reservoir-quality index based on porosity and permeability, log-transformed permeability, depth-derived temperature and pressure proxies, and interaction terms involving rock type, porosity and permeability. Other indices combined geological, seismic and spatial variables into quantities named mobility ratio, sweep efficiency, biosurfactant efficiency, polymer efficiency, enhanced oil recovery potential and synergistic EOR effect. Despite those operational-sounding names, they are computational composites of the available synthetic predictors, not measurements of oil recovery, chemical concentration, viscosity, interfacial tension, injected-fluid performance or displacement efficiency. The authors explicitly warn that their feature rankings should not be interpreted as evidence of field-scale enhanced oil recovery performance.

Recursive feature elimination was used within each training fold to identify a compact subset of predictors. L2-regularized logistic regression served as the ranking estimator, and candidate subsets were compared using mean validation ROC-AUC. The final model received 13 selected predictors. To learn different aspects of each reservoir record, the network used three parallel branches. A probabilistic branch based on prior-data-fitted tabular classification produced posterior probabilities, predictive entropy and probability margin. A contextual branch treated each feature as a token, embedded it with trainable feature identities and processed those tokens through three Transformer blocks with four attention heads. A sequential branch used TabNet-style sparse masks across four decision steps, encouraging the model to select different features for different records.

The three branches were not simply averaged. Each representation was first projected into a common 64-dimensional latent space. A gating network then assigned each record its own three branch weights, allowing the model to emphasize uncertainty estimates, contextual relationships or sparse feature selection according to the input. Sparsity regularization encouraged selective feature use, while a diversity term discouraged the branches from learning identical representations. The fused output passed through dense layers and a sigmoid to produce an oil-presence probability. Hyperparameters were selected with 60 trials of a tree-structured Parzen Estimator using Hyperband pruning, optimizing mean validation ROC-AUC and using mean F1 as a secondary criterion. The selected configuration used AdamW, a batch size of 64, an initial learning rate of 3 × 10^-4 and early stopping after 12 epochs without validation-loss improvement.

Across five repeats of five-fold stratified cross-validation, the complete network produced 25 paired validation estimates. Mean accuracy was 98.60%, with a standard deviation of 0.27 percentage points and a 95% confidence interval of 98.49% to 98.71%. Mean ROC-AUC was 0.9809 and mean precision-recall AUC was 0.9831. The mean Brier score was 0.0138. At the pre-specified classification threshold of 0.50, held-out accuracy was 98.74%, with a bootstrap 95% confidence interval of 97.84% to 99.64%. Sensitivity was 98.06%, specificity was 99.00%, the F1 score was 97.75% and the Matthews correlation coefficient was 0.9688. The observed positive prevalence in the test set was 27.88%, while the model predicted a positive rate of 28.06%.

Calibration results also indicated that the predicted probabilities were broadly consistent with the observed test outcomes, although the evaluation sample was modest. The held-out Brier score was 0.0124, expected calibration error was 0.0160, the calibration slope was 1.08 and the intercept was −0.10. The reported ROC-AUC and precision-recall AUC bootstrap intervals extended to 1.0000, reflecting the small number of errors in the 556-record test set. Statistical comparisons used paired Wilcoxon signed-rank tests across the 25 cross-validation estimates, with Holm adjustment for seven baseline comparisons. The model exceeded k-nearest neighbours, support vector machines, multilayer perceptrons, an autoencoder-classifier combination, random forest, a deep-and-wide neural network and XGBoost in ROC-AUC. XGBoost was the strongest baseline, reaching 93.53% held-out accuracy and a ROC-AUC of 0.9497; its mean paired cross-validation ROC-AUC was 0.0320 below the proposed model. All seven comparisons remained statistically significant after Holm adjustment, with adjusted p values from 2.98 × 10^-5 to 4.17 × 10^-7.

Ablation experiments indicated that the complete configuration was not merely benefiting from the addition of several neural networks. The sequential-attention branch was the strongest individual branch, with a cross-validation ROC-AUC of 0.9431, while the contextual and sequential branches together reached 0.9638. Unweighted concatenation produced a ROC-AUC of 0.9689, lower than the full model’s 0.9809. Removing diversity regularization reduced the score to 0.9756, and removing sparsity regularization reduced it to 0.9741. The complete architecture also had the lowest Brier score among the tested variants. Attention analysis ranked five engineered proxies as the most influential: the final EOR score, synergistic EOR effect, sweep efficiency, mobility ratio and enhanced oil recovery potential. Together they accounted for 0.592 of normalized importance and were selected in all 25 folds, but their importance reflects predictive association within synthetic data rather than demonstrated physical or operational effects.

The central achievement is therefore a carefully validated multi-branch architecture for heterogeneous tabular reservoir descriptors, with sample-specific fusion showing stronger internal performance than the evaluated alternatives. The architecture also illustrates how uncertainty, contextual feature relationships and selective feature use can be combined without forcing them into a single representation. Yet the evidentiary boundary is equally important. The study used one synthetic dataset and did not include production histories, well-log sequences, pressure-transient measurements, fluid compositions, injection schedules or reservoir-management variables. Independent testing on field data from different lithologies, operating conditions, class prevalences and acquisition systems would be required to establish generalizability and practical value. Until then, the model’s near-perfect scores demonstrate consistency within a controlled synthetic benchmark, not proof that it can forecast oil presence or measure recovery efficiency in a working reservoir.

Subject of Research: Evaluation of a multi-branch deep-learning model for oil-presence classification in synthetic reservoir data

Article Title: Integrated analysis of geological and seismic parameters for oil presence prediction in reservoirs using the synergistic oil recovery-net model

Article References: Li, D. (2026). Integrated analysis of geological and seismic parameters for oil presence prediction in reservoirs using the synergistic oil recovery-net model. Frontiers in Earth Science, 14, Article 1900399. https://doi.org/10.3389/feart.2026.1900399

Image Credits: AI Generated

DOI: 10.3389/feart.2026.1900399

Keywords: oil-presence prediction, reservoir characterization, deep learning, attention mechanisms, petrophysics, seismic data, synthetic data, feature selection, model validation, Integrated, analysis, geological