On the Effectiveness of Kolmogorov—Arnold Networks for Enhanced Oil Recovery Prediction in Polymer Flooding

Abstract

Enhanced oil recovery techniques like polymer flooding are vital for maximizing hydrocarbon extraction, but predicting recovery factors remains challenging due to complex reservoir dynamics. Traditional artificial neural networks offer predictive power but lack interpretability, limiting their practical utility in engineering applications. This study introduces Kolmogorov-Arnold Networks (KANs) as an interpretable alternative for modeling recovery factors in polymer flooding. Using a synthetically generated dataset of over 160,000 samples with key reservoir parameters, we trained KAN models through a series of experiments, varying network architectures, grid sizes, learning rates, and optimizers (Adam and LBFGS) to evaluate performance and interpretability. The best-performing KAN achieved a Test Mean Squared Error (MSE) of 0.000597 and a Test coefficient of determination$(\boldsymbol{R}^{2})$of 0.902 with only 1,885 parameters, closely matching the performance of a previous deep neural network model that used 43,265 parameters (Test$\mathbf{R}^{\mathbf{2}}$of 0.908). While KANs did not surpass the DNN in predictive accuracy, they offered comparable results with significantly reduced complexity and enhanced transparency. The modular structure and learnable activation functions of KANs provide insights into the decision-making process. This work contributes to the EOR field by demonstrating that KANs can effectively model complex reservoir behaviors while being transparent in their decision making.

More by Samson Dawit Bekele

• Separating Artificial-Viscosity Annealing from the L-BFGS Viscosity in Physics-Informed Neural Networks for Two-Phase FlowBig Data and Cognitive Computing · 2026
• Identifiability and Conditioning of Flux-Parameter Inversion for the Buckley–Leverett EquationMathematics · 2026
• Artificial-viscosity schedules for Buckley–Leverett PINNs: code and reproducibility archiveZenodo · 2026
• Physics-Informed Neural Networks for Coupled Non-Isothermal Two-Phase Flow in Porous Media: Training Pathologies, Remedies, and the Role of Gradient ConflictMachine Learning and Knowledge Extraction · 2026