Machine learning has been widely applied to reservoir fluid prediction; however, its performance strongly depends on the quantity, representativeness, and class distribution of training samples. In practical exploration and development, labeled fluid samples are commonly derived from limited well-log data and are therefore insufficient to capture the range of fluid responses under complex reservoir conditions, restricting model generalization. To address this issue, this study proposes a rock-physics-guided training sample expansion method for reservoir fluid classification under limited well constraints. Elastic response samples under various reservoir conditions were generated using a rock-physics model, and distribution matching between model-generated samples and actual well-log data was performed in the feature space to improve their consistency with actual reservoir responses. A fully connected neural network was then used to establish the nonlinear relationship between fluid-sensitive attributes and reservoir fluid classes. Data-driven, model-driven, and model–data joint-driven training strategies were compared to evaluate the applicability and transferability of the expanded samples. The results show that the rock-physics-guided expanded samples achieved classification performance comparable to that of well-data-driven models and improved model stability under reservoir conditions broader than those represented by the limited training wells. Optimization of the training sample class distribution further improved the recognition of different fluid types. Applications to actual well-log and seismic data verified the effectiveness of the proposed method. These results indicate that machine learning–based fluid prediction under limited well constraints depends not only on network optimization but, more fundamentally, on constructing training samples with physical validity, distributional consistency, and adequate geological coverage.