<p>The mechanical properties of additively manufactured metals are critical for evaluating the reliability and applicability of printed components. However, their strong dependence on processing conditions necessitates extensive sample fabrication and characterization, which is time-consuming and cost-intensive. Machine learning provides a data-driven approach to efficiently learn and model the complex relationships among process parameters, microstructural characteristics, and material properties. In this study, 304 and 316L stainless steels fabricated by electron beam freeform fabrication were investigated, and a physics-enhanced machine learning framework was proposed for hardness prediction. Key microstructural features, namely the area fraction and average area of δ-ferrite, were quantitatively extracted from optical micrographs using OpenCV-based image processing techniques. In addition, physics-informed descriptors derived from solidification thermophysical constraints and phase composition information were incorporated to enhance the learning process. Among the evaluated models, the physics-enhanced RF model achieved the best predictive performance, with an R² of 0.919 and an RMSE of 4.761 HV on the test dataset, significantly outperforming the GBR, GPR, SVR and RF. SHAP analysis further revealed that the steel grade indicator, together with the area fraction and average area of δ-ferrite, plays a dominant role in governing hardness. The proposed framework effectively integrates data-driven modeling with physical mechanism constraints, providing an interpretable and accurate approach for hardness prediction in additively manufactured stainless steels.</p>