Oil extraction requires understanding of underground pressure and oil-water distribution at depths of several kilometers. However, due to rock heterogeneity and multiphase fluid interference, accurate prediction is extremely challenging. Recently, a research team from the Industrial Control Network and System Department of the Shenyang Institute of Automation (SIA) of the Chinese Academy of Sciences, proposed a novel method called PI-DeepOKAN, which deeply integrates physical laws with artificial intelligence, providing an efficient new tool for reservoir dynamic simulation and intelligent decision-making.

The results has been published in Expert Systems with Applications, under the title PI-DeepOKAN: Physics-Informed Deep Operator Kolmogorov-Arnold Network with Output-Head Reweighting for Multiphase Flow Prediction. The first author is doctoral student ZHANG Xiaodi, and the corresponding authors are Associate Professor CHENG Haibo and Professor LI Dong.

Traditional purely data-driven models rely heavily on large amounts of labeled samples and struggle to guarantee physical consistency and generalization capability, while conventional numerical simulation methods are computationally time-consuming and cannot meet the engineering demands for real-time prediction and rapid optimization.

The research team constructed a dual-branch deep operator architecture with output-head reweighting to achieve a unified representation of static geological parameters and dynamic production conditions. Recognizing the distinct evolution patterns of the pressure field and the saturation field, the researchers introduced an attention mechanism that adaptively adjusts the fusion weights of geological parameters and well-control information, overcoming the limitation of fixed multi-source information fusion found in conventional deep operator networks.

Furthermore, the study incorporates the learnable nonlinear mapping capability of theKolmogorov–Arnold Network (KAN), significantly enhancing the model’s ability to characterize complex multiphase flow behavior in highly heterogeneous reservoirs under limited sample conditions. The team also constructed a full-chain physics-informed loss function that encompasses multiphase flow governing equations, initial conditions, internal and external boundary conditions, and fault interface constraints, ensuring that the model predictions strictly conform to reservoir flow mechanisms.

Experimental results demonstrate that PI-DeepOKAN can accurately capture pressure propagation and waterflood front evolution in complex reservoirs. While maintaining physical consistency, it dramatically improves prediction efficiency, serving as a highly efficient surrogate model for rapid reservoir simulation, production performance analysis, and intelligent optimization decision-making.