CWatM provides a grid-based representation of terrestrial hydrology, applied in this instance at a spatial resolution of 5 arcmin (grid cells approximately 8 km wide near the Equator) and daily temporal resolution . CWatM distinguishes between six land cover types including forest, irrigated non-paddy cropland, irrigated rice paddy, impervious surface, water bodies and other land cover in simulating the water balance of each grid cell. CWatM includes processes relevant for high-altitude implementations, including snow, glacier and permafrost. Potential evaporation is calculated using the Penman–Monteith equations. Processes within soil layers include frost, infiltration, preferential flow, capillary rise, surface runoff, interflow and groundwater percolation.

The model simulates river streamflow using the kinematic wave routing approach and can simulate either naturalized streamflow or streamflow impacted by human activities including reservoirs, irrigation demand and water withdrawals and return flows by industrial and domestic sectors. Reservoir outflow in the model is a function of the relative filling of the reservoir, storage parameters and outflow parameters . Irrigation demand is a function of crop water demand, water availability and crop type (paddy or non-paddy) . Parameter calibration uses an evolutionary algorithm that optimizes a modified version of the Kling–Gupta efficiency (KGE) between the simulated and observed discharge .

MESSAGEix is an open-source, dynamic system optimization model developed for strategic energy planning . MESSAGEix is based on the original MESSAGE model that has been developed and applied widely over the past three decades to analyze scenarios of energy system transformation, both globally and in different geographic regions, under technical-engineering constraints and political–societal considerations, e.g., . A defining feature of MESSAGEix that distinguishes it from other energy models in its class (e.g., OSeMOSYS, ; and MARKAL, ), and what leverages its widespread use as a nexus solutions tool is that it incorporates the ix modeling platform (ixmp): a back-end database and version control system that enables users to collaboratively develop, solve and visualize models using the open-source R and Python programming environments . This feature complements with the philosophy and design of CWatM, which utilizes similar open-access software (Python) as the main interface for collaborative model development and calibration. In this context, the NEST framework employs the reticulate package to integrate R and Python work environments .

As a bottom-up system optimization model, MESSAGEix includes resource consumption and capacity limitations at the technology level. Each form of technology modeled in MESSAGEix is defined and characterized by input–output efficiencies (the rate at which a particular commodity is consumed or produced during technology operation), economic costs (investment, fixed and variable components) and environmental impacts (greenhouse gas emissions, water consumption, etc.). Technology in this context can represent any process that transfers or transforms commodities, including natural systems such as rivers, aquifers and crops. By solving the following deterministic interregional and intertemporal linear programming (LP) problem, MESSAGEix minimizes the total cost for system capacity and operation over a future time period while meeting user-specified levels of demand and technical/policy constraints: 1minf(x)=∑r,tcr,tTxr,tδr,t;Ax≥b. In the above system of equations, the time period index is given by t and the region index is given by r. The solution vector containing the capacity and activity of the technologies is given by x. Economic costs are described in the cost coefficient vector of the objective function c. The discount rate associated with future cash flows is represented by δ. The set of constraints including the supply–demand balances, capacity limits, technology retirements and capacity additions, activity bounds and additional policies addressing environmental impacts is contained in the technical coefficient matrix A and right-hand side constraint vector b. The full set of equations is summarized in the online model documentation (https://messageix.iiasa.ac.at, last access: 1 February 2020). The single-objective LP formulation can also readily be transformed to handle multiple objectives, such as minimum total investments, emission level or other environmental indicators .

By linking the inputs and outputs of individual processes, energy, water and land decisions can be represented as a single system using the MESSAGEix modeling scheme. Thus, decisions impacting system design and operation over the planning horizon are made understanding the nexus interactions and will adapt the transformation pathways for each sector to avoid constraints and reduce trade-offs from the perspective of the objective function. Moreover, MESSAGEix supports spatially distributed systems modeling using a node-link representation, where commodities can be transferred between nodes based on the definition of dedicated technologies. It is therefore possible to explicitly represent the interplays between up- and downstream water users. Commodities are distinguished by the location (level) within the supply chain, enabling explicit accounting of associated efficiency losses and costs for grid and conveyance infrastructures. The temporal representation enables users to select the investment periods (e.g., annual) and sub-investment periods (e.g., sub-annual) over which supply, demand and system capacity must be balanced .

The reference system is the user-defined bottom-up representation of the technological system and its spatiotemporal delineation in MESSAGEix that defines interactions between technologies and the balance of commodity flows across the system . The reference system contains the portfolio of possible technologies and interventions (existing and future) and does not typically change across scenarios; the parameterization of data, including the constraints, is varied to compare how the system reacts to certain inputs, policies and objectives. Two broad categories of data are used to characterize the NEST reference system: (1) historical data on resource use and availability, existing and planned technology capacities; and (2) parametric data for technologies used in the optimization model, expressed as costs or consumption of resources per unit of production. These data are based on assumptions and can vary spatially and temporally.

The existing MESSAGEix core model does not represent sub-annual storage dynamics and associated capacity constraints. Previous work demonstrates specific approaches for integrating short-term (i.e., daily) storage dynamics into long-term energy system models similar to MESSAGEix ; yet, sequential seasonal storage dynamics are most critical to represent from the perspective of water resources management because of the important role reservoirs play in balancing seasonal hydrologic and demand variability, and the potential for future reservoir development to compete with other water uses during filling. To enable inclusion of seasonal reservoirs in NEST, sequential monthly sub-annual time steps are included in the MESSAGEix implementation and the core model is enhanced with the following set of equations merged into the existing technical coefficient matrix and right-hand constraint vector: 5ΔSn,c,l,y,m⋅Δtm+Sn,c,l,y,m+1-(Sn,c,l,y,m⋅λn,c,l,y,m)=0Sn,c,l,y,m-≤Sn,c,l,y,m≤Sn,c,l,y,m+Sn,c,l,y,m≤Zn,c,l,yΔSn,c,l,y,m≤ΔZn,c,l,y. In the above equations, n is the node where the storage is located, c is the commodity stored, l is the level in the supply chain the storage interacts with, y is the investment period (annual), and m is the operational periods (sub-annual). The storage level is given by S, whereas the change in storage is given by ΔS. The first set of inequality constraints is used to limit the storage level to within a specific range (S- is the lower bound and S+ the upper bound), for example, to include operating rules for reservoirs used for multiple purposes. The second and third inequality constraints are the capacity limitations both in terms of system size (Z) and rate of commodity transfer (ΔZ). Storage losses (i.e., evaporation and seepage) are given by the factor λ and computed as a function of the estimated evaporation from the hydrological model and a linear area–volume relationship . The sub-annual time period duration Δt converts the storage change calculated as a rate into a volume consistent with the storage level. To account for filling behavior and interannual variations, we ensure (1) the start and end levels are the same across years when no new storage capacity is added; and (2) when new storage capacity is added, it must be filled uniformly throughout the first 10 years, thus presenting an additional freshwater demand. Capacity additions are exogenously defined based on reported data; future work will consider the capacity limitations as control variables that can be expanded through increased investment in storage capacity.

To avoid integer (binary) decision variables associated with the choice of whether or not to plant a specific crop in a specific area, an additional set of minimum utilization constraints is defined for crops included in MESSAGEix. This forces the optimization to maintain the growing schedule over the course of the year, while balancing the total land area across crop types. Further adjustments to the core model are needed to ensure the physical balance of EWL resources. Specifically, the existing MESSAGEix core model constrains resource supply to be greater than or equal to resource demand. This setup enables the model to spill excess resource production when beneficial to the overall operating costs of the system. However, this configuration poses challenges when accounting for inflows into the system to effectively size infrastructure capacity. For example, when considering wastewater return flows as a specific commodity that should be managed using wastewater treatment technologies, it is crucial to ensure a complete commodity balance across all time periods. Otherwise, the model would be able to exclude inflows to avoid building wastewater treatment capacity. To reconcile inconsistencies and to ensure a physical balance of EWL resources, we define a new set of supply–demand balance equality constraints in the enhanced MESSAGEix core model used in NEST.

Finally, for computational efficiency, we developed a set of tools in the R programming interface that enable users to rapidly prototype new models during the testing phase by selectively managing interactions with ixmp. We found that for the case study described in Sect.  that the new approach cuts model instance generation time by an order of magnitude. Importantly, the ixmp utilities can be optionally used so that once debugging is complete, models can readily be shared and modified using the powerful database utilities enabled with ixmp. All of the enhancements to the MESSAGEix model implemented in this paper can be obtained from the online repository for NEST (https://github.com/iiasa/NEST, last access: 1 February 2020).

To parameterize the model in terms of resources, technologies and demands, we used the data sources outlined in Table . Importantly, much of the data needed to run NEST can be obtained from open-access geospatial datasets with global coverage. Thus, NEST is readily adapted to other regions of the world. Nevertheless, it is important to emphasize the prioritization of approved local data, as well as use of the calibration steps that can be embedded in the framework that improve the performance of the model in terms of reproducing historical conditions. Moreover, it is important to stress the use of multiple climate models and RCP-SSP scenarios to bridge the range of uncertainties in the hydrological modeling and demand drivers.

We calibrated CWatM for the IRB at 5 arcmin resolution using the monthly streamflow data during 1995–2010 at the Besham station, in northern Pakistan. The Besham station is chosen because of its coverage of historical years; it incorporates the runoff from both glacial and seasonal snowmelt. However, multiple stations would be necessary to better represent regional heterogeneity (in particular lower versus upper basin). Future work will incorporate spatially distributed observations to improve the calibration. It is important to emphasize the complexity of the hydrology in the IRB and the difficulties in calibrating to observed data due to extreme elevation changes . For calibration, the CWatM simulations included human impacts on streamflow and a spin-up period of 5 years to allow long-term storage components to stabilize. Analysis of the initial calibration results showed that the calibration was mainly impacted by the ice melt coefficient and empirical shape parameter of the rainfall–runoff ARNO model for infiltration . Therefore, we ran a second calibration that searched for optimal values for only these two parameters. The calibrated parameter values are given in Table . The performance of the model after the two calibration runs is in Fig. . We then used the calibrated CWatM for the IRB for historical (1956–2005) and future (2006–2099) simulations using the downscaled meteorological inputs of the ISI-MIP2b project from four global climate models (GCMs: GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR, MIROC5) . The streamflow in CWatM was naturalized because human activity, and water withdrawals in particular, are represented and accounted for in the MESSAGEix framework. The resulting ensemble mean monthly runoff profiles for each riparian country's basin area and the Indus as a whole are depicted in Fig. . The total basin runoff matches closely with other reported data .

For implementation in MESSAGEix, the IRB is delineated into 24 BCUs using the basin and country administrative boundary datasets (Fig. ). Further disaggregation into the agroecological zones is not pursued in this case because of limited spatial variability in crop potential within the delineated BCUs. The planning horizon considers investment periods spanning 2020 to 2060 in 10-year time steps, and 2015 is parameterized as the base historical year (i.e., the initial starting point). Monthly sub-annual time steps are considered.

CWatM is run with fixed spatial and temporal resolution as mentioned in previous sections. Therefore, performance is not affected by the final scale of the optimization model. Running times are on the order of few hours on personal computers. The MESSAGEix component is instead scale sensitive: increasing the number of BCU or the temporal resolution increases the size of the matrix of the LP optimization significantly. In the current configuration, the cplex solver in the GAMS model reduces the system of equations to a LP matrix of approximately 1 million ×1 million elements and solves in less than 30 min on personal computers. For each policy scenario described in the following sections, CWatM is only run once for each SSP and RCP combination, while additional policies are only implemented and run in the optimization model.

With most of the land area dedicated to crop production, we simplify the reference system by limiting the land-use options to crop land choices and limit the crop types to fertilized options. The SSP2 (middle-of-the-road) socioeconomic scenario is explored in the analysis and the ensemble mean climate scenario across the RCP climate models is used for climate forcing.

Urban and rural population and per capita income for SSP1, 2 and 5 projected for 2050 are compared to 2010 values for each riparian country's part of the IRB in Fig. . It can be seen that rapid urbanization and growth in income levels are projected in the scenarios, and these changes translate into increased consumption of water, energy and crops in the modeling framework. Figure S1 in the Supplement depicts the corresponding sectoral exogenous demands for the SSP2 scenario. Note that results for China are not included because the existing and projected population growth in this region is very low, and thus the consumption has negligible impact on the downstream resources. Electricity demands increase most dramatically across countries due to the rapid increases in GDP and the assumption that electrification is supporting economic development. Water demands increase more gradually due to less influence of economic growth, although for India the manufacturing sector water uses increases significantly due to the existing water intensity. Corresponding projections of the population with and without access to pre- and post-treatment of freshwater are generated based on the GDP projections.

Canals play an important role in enabling the Indus Waters Treaty and are mapped to specific BCUs using the data in Table S1 in the Supplement. Operational constraints are also added to force the linkages to transfer water between routes, in line with the Indus Waters Treaty. The Indira Gandhi Canal is considered as a constraint on flows originating from the particular BCU where the inlet is found. Similarly, an urban water transfer to Karachi near the Indus Delta is included as an additional demand. The capacities of other water diversion infrastructures (surface and groundwater) for each sector are estimated from the historical withdrawals. The energy source for groundwater pumping is also identified, where diesel generators dominate in Pakistan and Afghanistan, and electricity is used predominately in India.

The existing and planned capacity of power generation in the IRB is depicted Fig. S2. Hydropower is the main source of generation capacity in the basin, with the basin regions of Pakistan also hosting significant amount of fossil generation. A number of large-scale hydropower projects are also planned in the region (Table S2) that were not found in the global databases, and these projects have also been included in the model. For projects with an opening date before 2025, it is assumed they are operational in 2020; all other projects are assumed to be operational in 2030. For hydropower projects with storage, the storage capacity is added to the BCU level storage in the year it becomes operational, with the filling of the reservoir averaged over the first 10 years of operation, as described previously. Existing storage capacity includes 26.4 km3 in Pakistan, 22.2 km3 in India and 0.6 km3 in Afghanistan. Operating rules are derived for the largest existing dams based on the historical reported releases for 2016 and 2017. Approximately 45 GW of additional hydropower potential is estimated using the approach described in Sect. , mainly in the Upper Indus Basin. The assessed solar and wind potential greatly exceeds the electricity demand, with most of the wind potential focused mainly in the Indus Delta region. Tapping the solar and wind potential, however, requires investment in transmission and flexible assets (e.g., storage). In the Supplement, we provide the variable capacity factor of solar and wind potential aggregated for each BCU (Variable_capacity_factor.xlsx).

The performance of the different crop types considered in the model in terms of yields is presented in Fig. S3 for irrigated and non-irrigated options. Crop categories are set according to the main types of crops grown in the region, with some aggregation of crop types occurring to simplify the number of decision variables. The maximum productivity on a per hectare basis demonstrates that irrigation significantly boosts crop productivity in many locations, enabling less land to be used. As mentioned previously, land for each crop type in each BCU is constrained based on suitability and total area. Certain crops also are performing better than others in some regions, while some crops are not available entirely in some regions. The historical crop yields are harmonized to historical irrigation water use by calculating the required irrigation to support the historical crop areas using Eq. () and then calibrating the irrigation intensities such that the withdrawals match with the reported irrigation deliveries in aggregated to the BCU scale.

The parameterized NEST model of the IRB is applied within a scenario analysis in which a baseline (business-as-usual) scenario and a multi-objective scenario achieving multiple SDG indicators by 2030 are compared. The SSP2 information is used to parameterize population and economic indicators in each scenario. The business-as-usual scenario assumes the continuation of existing policies (e.g., Indus Waters Treaty) and aims at cost minimization with limited environmental constraints such as emission or infrastructure access targets. Conversely, the SDG implementation pursues a vision of economic growth (poverty eradication) jointly combined with reducing resource access inequalities and the environmental impacts of infrastructure systems. It is important to emphasize that the SDG scenario is not exploring all of the individual targets and indicators but rather a limited set relevant for water, energy and land systems that are also well represented in the NEST framework. The main features of the baseline and multiple-SDG scenarios are summarized in Table . The scenarios are simulated by solving NEST under the different implementations. Additional sensitivity analysis is performed to highlight uncertainties in the modeling framework.

Figure  shows the balance of surface water in the model for a specific subcatchment (BCU) in Pakistan with planned storage expansion in 2030. Inflows in the region from upstream river flows or from internal runoff are subject to strong seasonal variations (Fig. a). Urban, rural and industry water demands are assumed to be constant through the year (therefore not shown in Fig. ), while water requirements for agriculture and power plants' cooling are instead endogenous and thus variable during the year (Fig. c). Supply and demand are not constrained to specific water sources; Fig.  depicts a case in which water requirements for agriculture are supplied by groundwater (Fig. b). Most surface water is indeed outflowing from the region to downstream nodes (Fig. d) due to environmental flow requirement and it is in turn used for hydroelectric generation (Fig. e). Noticeably, storage absorbs the high inflow peaks in the months of April and June, and releases high outflows in July (Fig. f). However, it is not straightforward to directly link the reservoir-level changes to hydropower generation or other regional water requirements under the conjunctive management strategy. Storage regulation appears in this case to mostly be serving downstream water demands as opposed to supporting hydropower potential.

Seasonality effects embedded in the model input are mostly related to water availability, renewable energy capacity factors and crop water requirements and productivity. Figure  shows outputs of the model that are affected by the abovementioned seasonal variations. Electricity generation fluctuations in hydropower generation are mostly compensated by nuclear, imports or natural gas. Similarly, the time for crop cultivation, growth and yield is season specific, taking into account precipitation and crop coefficient seasonality. Other studies have looked at the role of hydropower in the region with a nexus perspective, considering both electricity production and water management . Whilst the results from the Indus Basin Model Revised (IBMR) and NEST could be compared, if similar scenarios were run, it must be noted that IBMR only focuses on a subregion of the basin network with higher spatial detail, while NEST includes a more complete representation of energy demands, supply and water–energy linkages.

We present a comparison between the baseline and the multiple-SDG scenarios. Figure a depicts the yearly average new investment portfolio and associated average operational costs for each scenario. To achieve the sustainability goals, investment costs approximately double, while operational costs increase by about 30 % (these include fixed costs, variable costs of operation and costs of electricity imports). To meet the targets for wastewater treatment and the share of renewable energy (solar and wind and geothermal), a large portion of the new investments are dedicated to technology development and a shift in power plant type. Hydropower emerges as an important option in the baseline because of the planned expansions and unexploited potential quantified in the assessment.

As a consequence of the environmental flow policy in SDG6, multiple sectors need to adapt to lower water availability. The agriculture sector is particularly impacted due to its high share of total water demands and expands to non-irrigated areas to avoid water withdrawals. However, this implies lower yields and so more area is needed to support the same production and at higher operational costs due to the lower productivities. Additionally, there is increased investment into more efficient irrigation technologies, especially where most of the available arable land is cultivated and production still needs to be boosted to maintain agricultural supplies.

Figure b compares the nexus interactions at the basin scale for each scenario. The multiple-SDG scenario displays an almost 5-fold increase (from 500 to 2500 GWh yr-1) in energy requirements for water management (mostly pumping, treatment and arrangement of new canals). This is to support increased water access in the municipal sector and massively expanded wastewater treatment capabilities in urban areas but still represents less than 2 % of total electricity generation projected in 2020. A combined GHG emission target ensures the increased demands are met without increasing carbon emissions.

These results demonstrate the value of interconnection across EWL sectors in terms of chain reaction in investments (i.e., expanding piping distribution also requires expansion in electricity production and distribution), synergies (investing in irrigation efficiency implies saving in water distribution for irrigation) and trade-offs, as it is clearly not possible to minimize costs and resource use across all sectors to achieve the SDGs.

The sustainability scenario includes multiple policy objectives across different sectors, which are considered simultaneously by the model. Specific policy objectives can thus be analyzed individually or in combination. Cross-sectoral implications are not necessarily the same when assessing multiple policies at the same time or individually. However, to additionally understand the implication of each single SDG policy on the water, energy and land systems, we tested each policy independently (as in Table ).

Figure  depicts the electricity generation, water withdrawal by source and the land use for agriculture in India and Pakistan from 2020 to 2050 in all the scenario permutations tested. The baseline scenario assumes that enough water is present in the basin to meet increasing energy, water and food demands, while fulfilling the Indus Waters Treaty allocations, but neglecting the additional environmental flow standards, water efficiency guidelines and infrastructure access constraints present in the SDG6 case. The second row of plots depicts the sectoral changes induced by the multiple sustainability policies. Intuitively, constraining the use of surface water for environmental purposes has most impact on cross-sectoral activities in Pakistan because it is the most downstream country and thus faces the greatest challenge in meeting increasing water demands while concurrently allocating more flow to ecosystems when water is already scarce. In fact, its hydroelectric potential is significantly reduced and the main water source left is renewable groundwater. This has a large impact on the agriculture system, where both India and Pakistan expand cultivated land with rainfed crops, to adapt to water scarcity

For this case study, we do not consider land-use change to other types of land, such as forests.

It is crucial to note that in India the total available land for agriculture is already utilized in the base year in most of the modeled regions due in part to the Indus Waters Treaty obligations (which allow India to use a limited amount of western river waters for irrigation). Thus, to fulfill increasing food demand and reduce the water consumption per hectare in the SDG scenario, an uptake in more efficient irrigation technologies is observed. Importantly, the basin-wide water accounting framework enables the applied water efficiency policies to account for the complex interactions between irrigation water losses and groundwater availability, to ensure that a combination of surface and non-renewable groundwater sources are conserved.

Looking at single scenarios separately helps to understand what policy drives the specific changes and what sector is mostly affected.

SDG2. Most of the existing flood irrigation systems are substituted by drip and sprinkler technologies. This reduces the water demand for irrigation. For further analysis, the authors intend to add other SDG2-related targets concerning changes in food demand, import, export and shifts to different types of crops.

Integrated assessment models are subject to different types of uncertainty, which can cumulate and therefore require particular attention. Uncertainty can be broadly divided in data or parametric uncertainty, which is given by data sources, often represented as distribution or numerical ranges, and assumption uncertainty, occurring when dealing with future scenarios in the scope of policy analysis .

Here, we present an example of scenario uncertainty propagation between the two different models in NEST and a simplified parametric sensitivity analysis to demonstrate the need for a more thorough study.

Figure a shows different level of monthly total runoff from the CWatM using different climate models and under two different climate scenarios (RCP2.6 and 6.0). We notice major diversity in trend given by different climate models, while climate scenarios imply changes mostly in the eighth and ninth months of year 2020. When running the optimization model in NEST, outcomes carry the uncertainty from the hydrological model and cumulate it with other types of uncertainty. Figure b shows total cost for the Indus region where the uncertainty of different SSP assumptions is added for the previous set of climate scenarios. We notice how SSP assumptions more greatly affect total cost compared to either the climate model or RCP (each bundle of same-color lines includes runs with all climate scenario and RCP assumptions). However, looking at SSP2 and 1, with reduced stress caused by population growth, climate uncertainty is more significant than for SSP5.

Figure  illustrates different output changes under the baseline scenario in response to an arbitrary variation in input parameters (-25 %, +25 %). Intuitively, some outputs such as groundwater extraction and energy production are strongly affected by the variation of sector-related parameters, irrigation and power plants' efficiency, respectively. The plot also shows some significant cross-sectoral feedbacks. For instance, changes in energy efficiency and investment cost impact groundwater and surface water withdrawals by up to 5 %, and irrigation efficiency strongly impacts fossil-fueled energy production. Land use seems not to be sensitive to the input parameters considered here. Looking at total cost alterations, multiple parameters induce variations comparable with the scenario uncertainty described above. This preliminary analysis suggests that further and more thorough study is needed, adopting more realistic uncertainty ranges or data distributions.

Structural uncertainty also typically characterizes complex models such NEST . We expect that omitting some feedbacks or expanding some modules could distort some of the model responses. Further work will focus on exploring this type of uncertainty.

Increasing spatial and temporal resolution might be helpful to focus on subregions and identify possible critical areas with higher detail. However, it brings greater computational challenges associated with using classical mathematical programming methods. In this context, scaling of the input–output coefficients to ensure fast solution times can be challenging for nexus models, because many cross-sectoral interactions require definition of input–output coefficient ranges covering multiple orders of magnitude. Future work may need to explore heuristics or other emulations as an alternative approach to classical optimization methods in order to integrate and optimize the vast amounts of geospatial data increasingly available and promoting the use of ultrahigh-resolution models for infrastructure planning.

From a hydrological perspective, some limitations of the current NEST formulation include the use of static land-use maps in the development of the water resource potentials. Dynamic land-use maps could be used in future work using the optimal solutions from MESSAGEix. An important next step involves downscaling water- and land-use results to the spatial scale used in the hydrological model, improving the visualization and analysis of results, as well as enabling spatially explicit calculation of water availability and demands to represent dynamic changes of water and land use consistently across the two models in NEST. The assessment of groundwater could also be improved by including lateral groundwater flows and by changing the representation of aquifer recharge to a non-linear model. A major constraint in modeling hydrological processes is the linear formulation of the optimization model which limits dynamic representation of key sustainability indicators as continuous model decision variables (e.g., water quality).

Finally, assumptions on boundary conditions, such as costs of imports (of food, electricity or water), are important for simplistic assumptions (e.g., electricity imports in Fig. ). Future work could improve the representation of boundary conditions with supply–cost curves or by linking with market models representative of the system outside the study area. Linking with global and regional integrated assessment models through the common commodity markets could improve the expected import–export response in NEST under scenarios of global change and explore different scenarios of basin self-sufficiency and resilience to external shocks.