<p>To achieve property-variation-driven alloy composition design, a novel strategy integrating machine learning for constructing an incremental model is proposed in this study. First, high-accuracy property prediction models were developed using a machine learning algorithm (<i>R</i>² = 0.95). Then, an incremental model (referred to as the P-C model) was innovatively established. It takes property variations and processing parameters as inputs and outputs composition variations, thereby enabling the inverse prediction of main elements variations. Bayesian optimization was further applied to design the contents of auxiliary elements. Ultimately, three new Cu-Ni-Si alloys with excellent comprehensive properties were successfully designed. The properties errors of all new alloys were &lt; 3.5% as verified by the property prediction models. Consistency between the decision-making of the model and metallurgical principles was further confirmed by the SHapley Additive exPlanations (SHAP). This strategy provides a new methodology for the inverse design of copper alloys and other metallic materials.</p>