Objective: The accurate prediction of surface roughness is a major technical challenge in aircraft grinding processes. This difficulty primarily arises from the complex nonlinear influences of factors such as material properties, process parameter coupling, and dynamic disturbances. These factors have long made the modeling and prediction of surface roughness in grinding processes extremely challenging. Conventional physical models often oversimplify the grinding process and fail to capture its inherent complexity, while data-driven models are susceptible to noise in measurement data, thereby limiting their generalization capabilities. To overcome the difficulties in achieving both high accuracy and robustness, this study proposes a novel surface roughness prediction model based on a physics-informed neural network (PINN). Methods: This study utilizes a robotic grinding experimental platform to collect data encompassing various process parameters and the resulting post-grinding surface roughness. Initially, logarithmic transformation and normalization are applied to process parameters so as to enhance the model's training efficiency and numerical stability. Subsequently, a PINN with a multilayer perceptron (MLP) architecture is constructed, incorporating a physical power-law mechanism (PLM) governing surface roughness into a loss function. This design enables the synergistic optimization of both the physical constraints and data-driven learning, with the loss function comprising a data fitting error and a physical model error. A dynamic weighting strategy is further introduced into the loss function: during the early stages of training, considerable emphasis is placed on physical consistency by assigning higher weights to the physics-based term, thereby mitigating the effect of data noise. As training progresses, the weight of the data-driven component is gradually increased, allowing for the model to capture the complex nonlinear relationships among process parameters and improve prediction accuracy. Finally, the performances of the PINN, MLP, and PLM are systematically compared in training efficiency, prediction accuracy, and generalization performance. Results: Experimental results indicate that the proposed PINN model offers clear advantages over baseline models in training efficiency, prediction accuracy, and generalization capability. Specifically, 1) the convergence speed of the PINN model is comparable to that of the PLM and is approximately 40% faster than that of the MLP. Moreover, the training process is more stable, and the final loss after convergence is the lowest among those of all models; 2) The absolute prediction error of the PINN model does not exceed 0.03 μm, with an average relative error of only 3.263%. This not only satisfies the stringent process-tolerance requirements of aviation maintenance but also results in overall prediction accuracy superior to those of both the PLM and MLP; 3) In comparison with the MLP, the PINN model exhibits considerably less performance degradation on the test set, significantly mitigates overfitting, achieves a more uniform error distribution, and exhibits the least systematic bias, thereby showing superior robustness and generalization ability. Conclusions: By incorporating a physics-guided framework and a dynamic weight-adjustment strategy, the proposed PINN model achieves not only improved training efficiency and prediction accuracy but also enhanced generalization ability and robustness. In practical engineering applications, the PINN model achieves robust, high-precision surface-roughness prediction. Compared with traditional physical models and purely data-driven models, the PINN model delivers superior performance and fully meets the stringent requirements of the aviation industry with respect to intelligent manufacturing and surface-quality control.