Abstract: Objectives: Downward continuation is not only a key step in the fusion of multi-source gravity data but also plays a significant role in fields such as local gravity field refinement, Earth's internal structure inversion, and gravity-aided inertial navigation. However, the downward continuation process is inherently an ill-posed problem. Interference from observation noise, boundary effects, and the ill-conditioned nature of the coefficient matrix can amplify uncertainties in the continuation results. Methods: To address the challenges of high-frequency noise sensitivity in gravity downward continuation and the issues of slow convergence and susceptibility to local minima in traditional single-scale iterative methods, this study formulates downward continuation as a least-squares optimization problem. We propose an Adaptive Multigrid Regularized GaussNewton Method (AMG-RGN). The core of this method is the construction of a multi-scale correction framework adapted to the ill-posed characteristics. First, based on the theory that the optimal solution should be a fixed point of multigrid inversion, a coarse-grid correction functional incorporating a model discretization error compensation term is constructed. Simultaneously, a gradient consistency constraint is introduced to force the coarse-grid search direction to approximate the projection of the fine-grid gradient, ensuring the global convergence of the multi-scale iteration. Second, to address the differing signal and noise distributions across grid levels, an adaptive method based on energy ratio using Parseval's theorem is proposed. This method dynamically adjusts the regularization parameter according to the spectral energy variations of the solution, effectively balancing data fitting accuracy and model smoothness. Results: Using a triangular prism theoretical model and real airborne gravity data from the Gulf of Mexico, the proposed AMG-RGN method was compared against the stable Tikhonov Regularization Downward Continuation method (TRDC), the Vertical Derivative Taylor Series method (VDTS), and the Improved Derivative Iterative Downward Continuation method (IDIDC). The performance of these methods varied significantly under different noise levels. Specifically, the VDTS method was most severely affected by highfrequency noise, yielding poor accuracy across various metrics. While the IDIDC method showed some improvement, its accuracy remained limited by noise accumulation during iteration. The TRDC method demonstrated reasonable stability but sacrificed some high-frequency signals to suppress result divergence under high-noise conditions. In contrast, the AMG-RGN method achieved superior accuracy metrics compared to the other three methods. Furthermore, in terms of computational efficiency, benefiting from the accelerated convergence characteristic of the multigrid algorithm, the average computation time of the AMG-RGN method was only 2% to 5% of that required by the TRDC method, demonstrating a significant computational advantage. Conclusions: Based on the analysis of downward-continued gravity anomaly profiles and residual maps under different noise levels, the TRDC, VDTS, and IDIDC methods all exhibited varying degrees of fitting distortion in areas with sharp gravity anomaly changes. This oscillation phenomenon became significantly more severe as noise increased. In comparison, the AMG-RGN method not only recovered the peak characteristics of gravity anomalies with high precision but also preserved the details of local anomalies comprehensively, effectively reflecting the trend of gravity anomaly variations.