PIK-Informer: A Physics-Informed Informer-KAN Framework for Joint Reservoir Fluid Identification and Water Saturation Estimation in the Youshashan Oilfield

Summary

Reservoir-fluid identification and water-saturation (Sw) estimation from well logs remain difficult because adjacent fluid classes exhibit overlapping responses, logging/reservoir relationships are nonlinear and formation dependent, and discrete class predictions may conflict with continuous saturation estimates. Existing Transformer-based methods often use dense attention, conventional prediction heads, and weak cross-task physical coupling. This study proposes PIK-Informer, a physics-informed Informer with Kolmogorov-Arnold network (KAN) framework that formulates interpretation as an ordered discrete/continuous task. ProbSparse attention and center-point pooling extract stratigraphic context while emphasizing informative depth positions associated with thin beds, boundaries, and fluid transitions. Compact KAN heads model segmented and locally nonmonotonic petrophysical relationships, while adaptive Archie, class/Sw interval-consistency, and depth-continuity constraints improve physical and geological coherence. Using 786 wells divided into 550/118/118 training, validation, and test wells, PIK-Informer achieved 92.86 ± 0.12% Accuracy, 0.910 ± 0.002 Macro-F1, 0.862 ± 0.003 Boundary-F1, and 0.052 ± 0.001 Sw root mean squared error (RMSE). Relative to the physics-informed graph attention Transformer (PiGAT), Accuracy increased by 0.53 percentage points and Sw RMSE decreased by 1.68%. All trainable models were evaluated over five runs, and external blocks were tested without local fine-tuning. Missing-data, noise, depth-shift, uncertainty, and independent MRTOS and ARTOC evaluations confirmed stable performance under incomplete inputs and cross-block distribution shifts. With 2.92 million parameters and approximately 95,420 samples·s⁻¹ throughput, the framework supports near-real-time field-scale batch interpretation.