The unique properties of aluminum-based alloys, such as low density, high specific strength, and excellent resistance to oxidation and corrosion, enable the design of advanced metamaterials. In this work, we theoretically investigate the effect of alloying aluminum with magnesium and zirconium on its thermodynamic and mechanical properties. Since exploring the vast chemical compound space of these alloys through Density Functional Theory (DFT) calculations is computationally prohibitive, we developed a scalable and transferable machine learning interatomic potential (MLIP) capable of accurately calculating diverse properties of Al-Mg-Zr alloys. The MLIP was trained using an active learning technique based on ab initio molecular dynamics simulations, Bayesian statistics, and kernel ridge regression. This methodology ensures that the MLIP captures the effects of alloying concentration and atomic configurations up to the solubility limit, providing access to highly accurate physicochemical properties of a wide range of Al-based alloys at a reasonable computational cost. We expect this approach to enable efficient phase space exploration, offering a robust tool for designing advanced Al-based alloys with optimized properties.