Desert cities are simultaneously faced with the stress of excessive heat and water resource shortage. Urban greening and strategic irrigation are proven effective heat mitigation strategies through shading and evapotranspiration. For sustainable urban development, smart urban irrigation schemes are required to maintain an intricate balance of water conservation with cooling efficiency, a challenge particularly acute in arid regions. While urban land surface models are capable of simulating these trade-offs, their computational complexity and steep learning curve hinder practical application in urban planning. In this study, we develop a machine learning-based protocol driven by a physical urban land surface model to optimize irrigation of urban trees in arid cities with field measurements. An artificial neural network surrogate was trained and validated, yielding high fidelity to the physical model simulations of canopy temperature (R2 = 0.972) and soil moisture (R2 = 0.989). We then adopted a genetic algorithm to find Pareto solutions by optimizing both the cooling and water use efficiencies of urban irrigation. The results of multi-objective optimization show that low-height trees with expansive crowns maximize shading-dominant cooling while dramatically reducing irrigation demand. These results challenge the reliance on water-intensive cooling strategies (e.g. lawns) and provide a scalable pathway to urban resilience to extreme heat and water scarcity.