Abstract: To address the prevalent issues in current propeller design, including high aerodynamic noise levels, the high computational cost of conventional optimization methods, and their tendency to become trapped in local optima, this paper proposed a low-noise propeller optimization design method that integrates multi-fidelity neural networks with deep reinforcement learning, while ensuring aerodynamic efficiency and meeting thrust requirements. B-spline curves were used to parameterize the propeller chord length, twist angle, and sweep distributions, thereby constructing a unified high-dimensional continuous design space. By combining low-fidelity data based on the BEMT/frequency-domain method with high-fidelity data generated using the CFD/FW-H method, an adaptive multi-fidelity neural network surrogate model was developed, where learnable weights were employed to dynamically fuse information from different fidelity levels. This surrogate model was then adopted as a virtual environment for reinforcement learning. A TD3 (Twin Delayed Deep Deterministic Policy Gradient)-based deep reinforcement learning algorithm was introduced to formulate the propeller geometric-parameter optimization as a continuous Markov decision process, enabling efficient global search through agent–environment interactions. Furthermore, a deep ensemble strategy was adopted to quantify the predictive uncertainty of the surrogate model, and a conservative evaluation scheme based on confidence bounds was incorporated into the reward function to improve optimization stability; the iterative samples were screened by selecting the geometric parameters corresponding to the 10% lowest predictive uncertainty, thereby enhancing the reliability of the optimization results. The results show that, compared with the baseline propeller, the optimized propeller achieves a noise reduction of 2.24 dB in the required 90° direction under the design conditions, with a maximum reduction of 3.69 dB obtained in the 80° direction. This method provides a new approach for efficient low-noise propeller optimization and holds promising potential for engineering applications.