Machine learning–enhanced multi-objective optimization of basin-scale water allocation in drylands

• RF-NSGA-III-Ada integrates machine learning prediction with multi-objective water allocation. • Adaptive initialization and operators improve the convergence and diversity of NSGA-III. • The Heihe application shows lower water use and emissions with higher economic efficiency. • The framework enables basin-scale water management under coupled economic and ecological constraints. Reconciling the conflict between rigid socio-economic demand and escalating scarcity is a critical challenge for sustainable water management in drylands. Conventional models are often limited in navigating high-dimensional sectoral trade-offs and nonlinear dynamics within the “water–ecology–economy” nexus. This study proposes RF-NSGA-III-Ada, a machine learning–enhanced framework that synergizes data-driven prediction with evolutionary optimization. Uniquely, the RF module quantifies socio-hydrological inertia to delineate a physically meaningful decision space, while the NSGA-III-Ada employs adaptive reference-point mechanisms to resolve the diversity-convergence dilemma. Applied to the prototypical Heihe River Basin, the framework optimizes the coupled allocation of surface, groundwater, and unconventional water sources. Results reveal a “stability-constrained efficiency” pathway: total water consumption decreases by 0.553 % (2025) and 1.172 % (2030), achieved through the framework’s precise structural optimization and marginal substitution strategies. Crucially, the optimized schemes effectively navigate the trade-offs between economic efficiency and social stability, achieving emission reductions without compromising the continuity of historical usage patterns. This framework offers a scalable paradigm for resolving the deadlock between structural rigidity and development in water-scarce regions globally.