Enhancing the heat resistance of cast Al-Si alloys is a critical prerequisite for expanding their service scope in automotive and aerospace sectors. To address the high cost and inherent inefficiency of conventional trial-and-error approaches for complex alloy composition design, this study proposes an interpretable machine learning strategy for developing heat-resistant cast aluminum alloys. This strategy employs a multi-step filtering process to identify synergistic microalloying element combinations, and leverages SHapley Additive exPlanations (SHAP) to elucidate the key intrinsic determinants and underlying mechanisms governing the high-temperature ultimate tensile strength (HTUTS) and high-temperature elongation (HTEL) of Al alloys, thus imparting explicit physical interpretability to the prediction model. Building on this, a novel heat-resistant cast Al-Si alloy with the composition Al-9.10Si-3.50Cu-0.50Mg-0.35Mn-0.25Zr-0.20Ti is rationally designed via coupling with the multi-objective coati optimization algorithm (MOCOA). The designed alloy delivers an exceptional HTUTS of 192.2 MPa at 300 ℃, outperforming all previously reported cast Al-Si alloys to date. Microstructural characterization confirms that the co-addition of Zr, Mn and Ti promotes the formation of thermally stable intermetallic phases, namely Al15(Fe, Mn)3Si2, (Al, Si)3ZrTi and Al3Zr. These phases synergistically impede dislocation motion at elevated temperatures, constituting the dominant mechanism underlying the remarkable enhancement in high-temperature strength. Meanwhile, the grain refinement induced by Zr and Ti, coupled with Mn-mediated mitigation of detrimental effects from brittle Fe-rich phases, accounts for the well-retained ductility. Overall, the findings of this work demonstrate the strong feasibility and high reliability of machine learning in multi-objective, performance-oriented alloy design, providing an efficient and interpretable methodological reference for the development of advanced heat-resistant cast Al-Si alloy.