Aerodynamic loads are essential for structural safety assessment in aeronautical engineering, yet direct measurement or high-fidelity simulation is often challenging. During service, structural parameters are affected by multi-source uncertainties, leading to limited samples and complex spatial variability. This paper proposes a hybrid physics–data-driven reduced-order modeling (ROM) framework for aerodynamic load inversion under structural field uncertainty. The ROM maps high-dimensional aerodynamic loads into a low-dimensional feature space, transforming the inversion into a feature coefficient prediction problem. A physics-decoded neural network (PDNN) is then established, where the data-driven module captures the mapping between responses and feature coefficients, and the physics-constrained module ensures consistency with physical laws. To characterize field uncertainty, the interval Karhunen–Loève decomposition (IKLD) is employed to represent uncertain fields as a finite set of parameters, while an adaptive Kriging surrogate model (AKSM) propagates uncertainty to yield interval estimates of loads. Numerical and experimental results demonstrate that the proposed method achieves high accuracy, robustness, and practicality under sparse measurement, noise, and uncertain field conditions.