The adsorption and diffusion behaviors of clusters on surfaces play critical roles in numerous important applications. Potential-based molecular dynamics simulations are a powerful tool to study these behaviors at the atomic scale. However, conventional potentials typically parametrized using bulk or surface properties, fail to accurately describe the intricate surface behavior of clusters due to the complexity of their atomic environments. Here, we develop a specialized machine learning potential (MLP) for describing Al clusters on surfaces, which is related to wide-ranging applications. The MLP development was performed using a workflow that is based on an adaptive iterative learning method and incorporates initialization, generalization, and specialization modules. By utilizing accurate data from density functional theory (DFT) calculations, the MLP achieves an impressive level of accuracy that closely approximates DFT while maintaining a high computational efficiency. The MLP successfully predicts the surface behavior of different Al clusters and a wide range of basic properties of the Al bulk and surfaces. Remarkably, despite being trained without data from Al x (x = 4−6, 12), the MLP accurately predicts the adsorption and diffusion properties of these clusters. This work highlights the capability of MLPs in the large-scale investigation of the surface phenomena of different clusters and provides a robust methodology for constructing accurate MLPs tailored to intricate surface systems.