2026-05-25 中国科学院（CAS）
＜関連情報＞
- https://english.cas.cn/newsroom/research-news/202605/t20260526_1159897.shtml
- https://www.sciencedirect.com/science/article/abs/pii/S0957417426010274
PI-DeepOKAN：出力ヘッド重み付けを用いた物理情報に基づく深層演算子コルモゴロフ・アーノルドネットワークによる多相流予測 PI-DeepOKAN: Physics-informed deep operator kolmogorov– arnold network with output-head reweighting for multiphase flow prediction
Xiaodi Zhang, Haibo Cheng, Dong Li, Jiahao Qiao, Xian’e Xiong, Liting Zhang, Minglu Hu
Expert Systems with Applications Available online: 30 March 2026
DOI:https://doi.org/10.1016/j.eswa.2026.132114
Abstract
Reservoir simulation provides a fundamental framework for analyzing subsurface flow processes. In recent years, surrogate modeling has been widely investigated to enhance the efficiency of reservoir simulation. However, existing data-driven reservoir surrogate models tend to exhibit limited generalization under small-sample settings. In this paper, we propose PI-DeepOKAN, a physics-informed deep operator Kolmogorov-Arnold network with output-head reweighting for multiphase flow surrogate modeling. The model integrates Kolmogorov-Arnold networks with learnable activations to improve expressive power. It further employs a dual-branch deep operator architecture enhanced by output-head reweighting, which adaptively reweights the geological and control branches for output-specific fusion. Physical constraints are incorporated into the loss as soft penalties to enhance physical consistency under limited observations. Validation results on two datasets show that the proposed method outperforms mainstream neural networks in terms of key metrics such as coefficient of determination R2, relative L2 error, and root mean square error (RMSE). Compared with best-performing baseline on the public faulted reservoir dataset, the proposed method reduces the relative L2 error by 65.2% for pressure and 63.3% for water saturation, and reduces RMSE by 64.7% for pressure and 63.4% for water saturation. These results indicate that PI-DeepOKAN is a promising approach for accurate and robust surrogate modeling of multiphase flow in complex reservoirs.
