イクノ　ソウイチロウ   Soichiro Ikuno   生野　壮一郎    所属   コンピュータサイエンス学部 コンピュータサイエンス学科    職種   教授
言語種別	英語
発行・発表の年月	2026/04
形態種別	学術論文
査読	査読あり
標題	Jacobian-AIME: A Local and Functional Approach to Interpreting Nonlinear Maps for Physics-Informed Neural Networks
執筆形態	共著
掲載誌名	IEEE Access
掲載区分	国外
出版社・発行元	IEEE
巻・号・頁	14
著者・共著者	Kosuke Yano, Takafumi Nakanishi, Soichiro Ikuno
概要	We propose Jacobian-AIME, a novel explanation framework for Physics-Informed Neural Networks (PINNs). PINNs map coordinates to physical fields governed by PDEs. Existing explainers—perturbation-based methods such as Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), and global inverse approaches such as Approximate Inverse Model Explanations (AIME)—either sample off-manifold or average away local structure, limiting their physical fidelity. Jacobian-AIME focuses on what drives a prediction at a specific location. At each point, it returns one vector—the Local Feature Contribution Vector (LFCV)—that shows which input directions matter and by how much, while remaining consistent with the physics learned by the model. This provides sharp, stable, and local explanations rather than diffuse, global ones. In this study, numerical evaluations of the Jacobian-AIME method were performed using boundary value problems for the two-dimensional Poisson equation and lid-driven cavity flow as evaluation experiments. Jacobian-AIME demonstrated clearer and more stable contribution maps compared to LIME and SHAP. Furthermore, the LFCV visualizes the interactions between input points for specific output points. Additionally, Friedman and Wilcoxon tests were performed, demonstrating statistically significant differences compared to LIME and SHAP. Our results suggest that conventional XAI metrics such as AUC may be influenced by off-manifold perturbations, highlighting the need for physics-aware evaluation criteria; Jacobian-AIME consistently provides local, on-manifold inverse explanations at the query point. Jacobian-AIME is broadly applicable to differentiable scientific models. It offers a practical and principled path to physics-faithful, local interpretability for PINNs.
DOI	10.1109/ACCESS.2026.3681926
外部リンクURL	https://ieeexplore.ieee.org/document/11477892