发表机构
School of Engineering, Physics and Mathematics, Northumbria University（北安普顿大学工程、物理与数学学院）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
针对太阳日冕DEM反演不适定且计算密集的问题，提出结合蒙特卡洛Dropout的近似贝叶斯深度学习框架，实现逐像素不确定性量化和高效非负DEM重建，在合成与真实数据上验证了准确性与计算效率。
AI 中文摘要
准确估计太阳日冕等离子体的温度分布，即微分发射度量（DEM），对于理解日冕的热力学及其相关加热过程至关重要。然而，从多光谱观测（如NASA太阳动力学观测站（SDO）上的大气成像组件（AIA））中恢复DEM是一个数学上不适定的欠定问题，传统的基于正则化的反演方法计算密集且提供的不确定性量化有限。我们提出了一种用于DEM重建的深度学习框架，该框架结合了蒙特卡洛Dropout以执行近似贝叶斯推断，从而生成每个像素的DEM经验分布，用以表征学习到的反演中的认知或系统性模型不确定性。该网络采用双头架构进行训练，同时监督AIA图像重建和DEM保真度，并通过构造确保非负性。网络直接使用真实的SDO/AIA观测数据以及来自正则化反演的相应DEM解进行训练。我们针对合成热分布和真实日冕数据验证了性能，证明该方法能在多种等离子体条件下准确恢复热结构，同时相比传统反演技术保持显著的计算效率。该方法确保物理上合理、非负的解，并提供逐像素的不确定性，使其成为大规模太阳数据分析中可靠、高速且具有不确定性感知能力的工具。
英文摘要
Accurately estimating the temperature distribution of solar coronal plasma, known as the Differential Emission Measure (DEM), is vital for understanding the thermodynamics of the corona and associated heating. However, recovering the DEM from multispectral observations like those from the Atmospheric Imaging Assembly (AIA) on board NASA's Solar Dynamics Observatory (SDO) is a mathematically ill-posed, underdetermined problem, and traditional regularization-based inversion methods are computationally intensive and provide limited uncertainty quantification. We present a deep learning framework for DEM reconstruction that incorporates Monte Carlo Dropout to perform approximate Bayesian inference, yielding per-pixel empirical distributions over the DEM that characterize epistemic or systematic model uncertainty in the learned inversion. The network is trained with a dual-head architecture supervising both AIA image reconstruction and DEM fidelity, with non-negativity enforced by construction. The network is trained directly on real SDO/AIA observations along with the associated DEM solutions from regularized inversion. We validate performance against both synthetic thermal distributions and real coronal data, demonstrating accurate recovery of thermal structure across a range of plasma conditions, while maintaining significant computational efficiency over traditional inversion techniques. This method ensures physically legitimate, non-negative solutions and provides per-pixel uncertainties, making it a reliable, high-speed, and uncertainty-aware tool for large-scale solar data analysis.
Journal refThe Astrophysical Journal, 1007 (2026) 1-12
