Volume 56

Interpretable Characterization of Carbon Sequestration Reservoirs Using Kolmogorov-Arnold Networks Jiayu Li, Xingwen Jiang, Siwei Meng, Yuwei L, Jizhou Tang

Abstract

In recent years, AI technology has been used to evaluate carbon sequestration reservoirs and caprocks, but the black-box nature of neural networks raises credibility concerns. This study employs the Kolmogorov-Arnold Network (KAN) to interpretably characterize carbon sequestration caprocks, including lithology identification and porosity prediction. Inspired by the Kolmogorov-Arnold representation theorem, KANs feature learnable activation functions on edges and univariate spline functions, enhancing both accuracy and interpretability. Smaller KANs achieve accuracy comparable to larger MLPs in data fitting and solving partial differential equations. For lithology identification, well log datasets from the Daniudi and Hangjinqi Gas Fields were used, with a KAN achieving a test accuracy of 0.806, surpassing traditional MLPs. For porosity prediction, datasets from the Gulf of Mexico wells were used, with a KAN achieving an MSE of 0.055. Fine-tuning and retraining derived a physical formula representing porosity based on well log data, elucidating the relationship between porosity and various parameters. This study demonstrates that KANs provide accurate and interpretable predictions, offering promising prospects for carbon sequestration site selection and reservoir characterization, thereby enhancing model credibility and advancing AI applications in geological sciences.

Keywords Kolmogorov-Arnold Networks, Carbon Storage, Lithology Identification, Porosity Prediction, Explainable AI