Traditional alloy design typically relies on a trial‐and‐error approach, which is both time‐consuming and expensive. Whilst physical metallurgical (PM) models offer some predictive capabilities, their reliability is limited by errors accumulating across space scales. To address this, this study proposes a novel framework that combines PM knowledge graphs (PMKGs) with graph neural networks (GNNs) to predict the tensile properties of quenching and partitioning steels, using genetic algorithms for dual‐objective optimization. Compared to traditional artificial intelligence (AI) models, this framework shows significant advantages in predicting ultimate tensile strength (UTS) and total elongation (TEL) with higher accuracy and stability. Notably, the R2 for TEL prediction improved by approximately 15%. Furthermore, this framework successfully balances UTS and TEL, resulting in the design of alloys with superior overall properties. The designed alloys, with a composition of approximately 0.3 wt.% C, 3 wt.% Mn, 1.2 wt.% Si, and minor amounts of Cr and Al, achieve a UTS exceeding 1500 MPa and TEL near 20%, aligning with PM principles and validating the rationality and feasibility of this method. This study offers new insights into applying AI in complex multi‐objective alloy design, highlighting the potential of integrating expert knowledge with GNNs.