From vulnerability exploration to adaptive policy design in semi-arid coastal basins: surrogate-assisted robust pathways building in the Quilimarí case, Chile.

Recent applications of Robust Decision Making (RDM) have demonstrated their value for exploring socio-hydrological vulnerabilities under deep uncertainty in water-scarce regions. A recent study in the semi-arid coastal Quilimarí basin (central Chile), used an RDM framework combining stakeholder engagement and an integrated WEAP-MODFLOW model to reveal critical trade-offs between agricultural production, drinking water security, groundwater depletion and saline intrusion under climate and development uncertainties. While this work provided valuable insights into system vulnerabilities and stressors, it remained focused on exploratory analysis rather than on the explicit design of adaptive strategies.

This study builds directly on the Quilimarí RDM case study and advances the framework toward adaptive policy design, introducing two methodological innovations. First, we extend the RDM approach by integrating the Direct Policy Search (DPS) framework  to identify robust and flexible water management strategies. Instead of evaluating a small set of predefined interventions, policies are formulated as adaptive decision rules that dynamically link observable system states such as groundwater levels, salinity thresholds or unmet drinking water demand, to management actions including abstraction restrictions, activation of alternative supplies or demand reallocation. This allows the systematic identification of pathways that evolve over time and remain robust across a wide ensemble of plausible hydroclimatic and socio-economic futures.

Second, to enable the computational requirement of DPS in data and process intensive basins, we develop a surrogate model that emulates the behavior of the full integrated surface-groundwater system. The original WEAP-MODFLOW model for Quilimarí, which explicitly represents groundwater dynamics, agricultural water use, and seawater intrusion, is approximated using an LSTM (Long Short-Term Memory), a type of Recurrent Neural Networks (RNN), trained on large ensembles of simulation outputs. The surrogate model preserves key nonlinearities and memory effects inherent to groundwater systems while reducing computational costs by several orders of magnitude, making large-scale adaptive policy search feasible.

The combined framework is applied to the Quilimarí basin to identify adaptive pathways that balance drinking water reliability, agricultural viability, and long-term groundwater sustainability under deep uncertainties as climate and land use change and growing population. Results show that the DPS using the surrogate model outperform static strategies identified in the original RDM analysis, particularly under severe drought and demand-growth scenarios, by avoiding maladaptation and reducing regret across objectives.

By explicitly linking the vulnerability exploration with the design of adaptive strategies, this study shows how RDM can be operationalized into implementable and flexible water management policies. The approach is transferable to other semi-arid coastal basins that face strong groundwater dependence, institutional constraints, and profound climate and other uncertainties.