Quantitative Evaluation and Screening of Complex Structural Traps in the Southern Junggar Basin Using a Support Vector Machine

Complex structural traps in the southern Junggar Basin require quantitative evaluation despite limited drilling-based labels. This study examined a support vector machine (SVM) workflow for 49 traps, including 16 labeled and 33 unlabeled traps. Eleven geological variables were considered, with grouped feature selection and model optimization performed within nested leave-one-out cross-validation (LOOCV). The final six-variable linear SVM achieved an accuracy of 0.5625 and a weighted F1-score of 0.5382 under strict nested LOOCV, matching the majority-class baseline accuracy (0.5625); nested leave-one-structural-belt-out validation yielded 0.4375, below the corresponding majority-class baseline (0.5625). None of the four Class II traps were correctly identified, while a source-rock-thickness-only SVM achieved higher accuracies under both validation schemes (0.6875 and 0.6250). Feature-selection stability was moderate (mean pairwise Jaccard similarity = 0.6732). The final model assigned 18 of 33 unlabeled traps to Class I, with 25 and 26 categorical agreements with fuzzy comprehensive evaluation and gray relational analysis, respectively. Seven illustrative traps were examined to interpret geological agreement and disagreement; their descriptive SVM scores do not constitute independently validated prospect rankings. The findings demonstrate the limitations of multivariable SVM screening with extremely small labeled datasets and the value of cautious geological cross-method interpretation.