Laser polishing serves as an effective post-processing technique to improve the surface finish of additively manufactured AlSi10Mg components, which are widely used in aerospace and automotive applications. However, predicting the resulting surface roughness remains challenging due to complex, nonlinear interactions among thermal input, process parameters, and material behavior. This study introduces a lightweight physics-augmented neural framework that integrates engineered features derived from laser–material interaction principles such as laser energy density, energy per unit length, and logarithmic energy density into a feedforward regression model. Unlike conventional black-box neural networks, the proposed hybrid approach enhances interpretability and generalization in data-scarce environments. Trained on a dataset of thirty samples, the model achieved an R² above 0.90 across training and validation sets, with a mean squared error below 0.2 on testing. The framework demonstrates a scalable and data-efficient route to surface quality prediction in laser polishing, supporting the development of intelligent, closed-loop control systems for additive manufacturing in Industry 4.0 environments.