Intelligent prediction of formation water saturation by fusing geological pattern recognition and flair gas logging

• A novel geochemical feature set, including heavy hydrocarbons, is developed for robust water saturation prediction. • Two novel predictive models based on different mechanisms, CET and CKET, are proposed. • The CNN-based CET model is demonstrated to significantly outperform the KAN-based CKET in both performance and efficiency. • Learning geological patterns is a more effective AI strategy for well log interpretation than deriving symbolic formulas. Accurate prediction of formation water saturation (Sw) is fundamental to reservoir evaluation, yet remains limited by (i) reliance on a narrow set of conventional well logs and (ii) sub-optimal exploitation of geological pattern versus functional relationship priors in deep models. A depth-aligned formation feature dataset comprising 340 non-overlapping formations (80 sampled depth points each; 2384.3–3804.4 m) is constructed, expanding the conventional 5-log suite (GR, NPHI, RHOB, DT, SP) with 13 quantitatively corrected hydrocarbon components (C1–nC5, i/nC4–C8, benzene, toluene, C7H14). Two embedding replacement strategies for the Transformer are proposed: a CNN Embedding Transformer (CET) leveraging locality and weight-sharing inductive biases for spatial pattern learning, and a CKAN Embedding Transformer (CKET) exploiting Kolmogorov–Arnold functional decomposition via KAN convolutions. Across 18 configurations, the best CET model (c_em_tm) attains RMSE 15.13, MAE 8.75, R 2 0.729 and EVS 0.729 with 2.09 × 10 5 parameters and 1.33 s mean processing time per epoch; an accuracy-matched CKET model (ck_el_ts) yields RMSE 15.16, MAE 8.72, R 2 0.728, EVS 0.730 but requires 7.73 × 10 6 parameters (≈36.9 × larger) and 22.55 s per epoch (≈17 × slower). Thus, CET achieves comparable fidelity with a 97.3% parameter reduction and major computational efficiency gains, underscoring that an appropriately aligned inductive bias outweighs higher theoretical expressiveness. An ablation removing heavy/aromatic hydrocarbons (C6+) degrades EVS from 0.729 to 0.382 (CET) and from 0.730 to 0.101 (CKET), demonstrating that low-concentration “weak” geochemical signals are decisive for Sw discrimination and fluid typing. The results establish a lightweight, geologically coherent pathway for sustainable, real-time Sw evaluation and motivate future work in multi-scale feature fusion and interpretable attention analysis.