黑洞自旋简化反演中的可识别性诊断性逆PINN研究
A Diagnostic Inverse-PINN Study of Identifiability in Reduced Black-Hole Spin Inference
浏览论文内容
中文总结 AI 辅助
本研究提出混合物理信息逆框架，用PINN反演简化黑洞自旋，发现其存在弱可识别性问题，为相关实验提供诊断价值并指出后续研究方向。
中文摘要 AI 辅助
黑洞自旋是相对论天体物理学中的基本参数，影响吸积效率、喷流启动和微扰模式。本研究探讨用于简化黑洞自旋反演的混合物理信息逆框架，将自旋恢复表述为受简化标量角Teukolsky类方程约束的逆问题。Physics-Informed Neural Network（物理信息神经网络，PINN）近似角模式函数，自旋参数被视为与控制微分算子残差相关的可训练物理量。该框架在受控合成及含噪角模式构型下，针对多个参考自旋值进行评估。主要结果为诊断性而非确认性：PINN定性再现了角模式轮廓，但推断的自旋值集中在允许区间的上部，而非准确恢复全部参考自旋范围。这揭示了弱可识别性 regime，即准确的角轮廓重建和低残差损失未必意味着准确的自旋恢复。结果凸显了物理信息约束对简化黑洞自旋反演实验的诊断价值，同时表明需进一步开展可识别性分析、损失重加权、联合本征值推断、完整克尔微扰建模、现实不确定性模型构建及数值求解器验证。
英文摘要
Black-hole spin is a fundamental parameter in relativistic astrophysics, influencing accretion efficiency, jet launching, and perturbative modes. This work investigates a hybrid physicsinformed inverse framework for reduced black-hole spin inference. Spin recovery is formulated as an inverse problem constrained by a reduced scalar angular Teukolsky-like equation. A Physics-Informed Neural Network approximates the angular mode function, while the spin parameter is treated as a trainable physical quantity linked to the residual of the governing differential operator. The framework is evaluated under controlled synthetic and noisecontaminated angular-mode configurations across multiple reference spin values. The main result is diagnostic rather than confirmatory. The PINN reproduces angular-mode profiles qualitatively; however, the inferred spin values cluster near the upper part of the allowed interval rather than accurately recovering the full range of reference spins. This reveals a weak-identifiability regime in which accurate angular-profile reconstruction and low residual loss do not necessarily imply accurate spin recovery. The results highlight the diagnostic value of physics-informed constraints for reduced black-hole spin-inference experiments, while also showing the need for further identifiability analysis, loss reweighting, joint eigenvalue inference, full Kerr perturbation modeling, realistic uncertainty models, and validation against numerical solvers.
