In coastal zones, a major objective of groundwater management is often to determine sustainable pumping rates which avoid well salinization. Understanding how model and climate uncertainties affect optimal management solutions is essential for providing groundwater managers with information about salinization risk and is facilitated by the use of optimization under uncertainty (OUU) methods. However, guidelines are missing for the widespread implementation of OUU in real-world coastal aquifers and for the incorporation of climate uncertainty into OUU approaches. An ensemble-based OUU approach was developed considering parameter, observation and climate uncertainty and was implemented in a real-world island aquifer in the Magdalen Islands (Quebec, Canada). A sharp-interface seawater intrusion model was developed using MODFLOW-SWI2 and a prior parameter ensemble was generated containing multiple equally plausible realizations. Ensemble-based history matching was conducted using an iterative ensemble smoother which yielded a posterior parameter ensemble conveying both parameter and observation uncertainty. Sea level and recharge ensembles were generated for the year 2050 and were then used to generate a predictive parameter ensemble conveying parameter, observation and climate uncertainty. Multi-objective OUU was then conducted, aiming to both maximize pumping rates and minimize the probability of well salinization. As a result, the optimal trade-off between pumping and the probability of salinization was quantified considering parameter, historical observation and future climate uncertainty simultaneously. The multi-objective, ensemble-based OUU led to optimal pumping rates that were very different from a previous deterministic OUU and close to the current and projected water demand for risk-averse stances. Incorporating climate uncertainty into the OUU was also critical since it reduced the maximum allowable pumping rates for users with a risk-averse stance. The workflow used tools adapted to very high-dimensional, nonlinear models and optimization problems to facilitate its implementation in a wide range of real-world settings.

Seawater intrusion is a major challenge for groundwater management in coastal zones, which are under pressure due to population growth, sea level rise and changes in climate (Michael et al., 2017; Jiao and Post, 2019). Numerical models are often relied on to support groundwater management using either advective–dispersive solute transport codes, which simulate mixing between freshwater and saltwater and are computationally demanding, or sharp-interface codes, which consider freshwater and saltwater to be immiscible but have significantly shorter simulation times. These numerical models can provide insight into freshwater availability under current and projected conditions and are frequently combined with optimization algorithms to determine sustainable pumping rates which avoid well salinization (Ketabchi and Ataie-Ashtiani, 2015). To provide information to groundwater managers about salinization risk, optimizations in coastal zones should recognize the uncertainty in model predictions (Werner et al., 2013). However, because of the high computational costs associated with both advective–dispersive solute transport models and stochastic uncertainty quantification, optimization under uncertainty (OUU) approaches have mostly been implemented in synthetic or simplified real-world cases (Rajabi and Ketabchi, 2017; Mostafaei-Avandari and Ketabchi, 2020). Guidance is missing for the implementation of OUU in real-world coastal settings (Ketabchi and Ataie-Ashtiani, 2015).

Real-world coastal OUU applications have generally used advective–dispersive solute transport models in combination with surrogate models, evolutionary algorithms and stochastic uncertainty quantification accounting for the uncertainty of a few model parameters (e.g., between 2 and 11 parameters in Sreekanth and Datta, 2014; Rajabi and Ketabchi, 2017; Lal and Datta, 2019; Mostafaei-Avandari and Ketabchi, 2020; Han et al., 2021). These applications were also not systematically preceded by parameter estimation (i.e., automated calibration). However, accounting for model parameter uncertainty accurately and robustly can require very high numbers of parameters (White, 2018). Furthermore, conducting data assimilation prior to OUU is essential, since it allows one to update parameter uncertainty estimates with observation uncertainty. In Coulon et al. (2022), the real-world OUU approach used a sharp-interface model, sequential linear programming (SLP) and first-order-second-moment (FOSM) uncertainty analysis accounting for both parameter and observation uncertainty (including 60 parameters and 162 observations). The approach was preceded by parameter estimation (Coulon et al., 2021) conducted using the widely used Gauss–Levenberg–Marquardt (GLM) algorithm. However, the GLM and SLP algorithms can be computationally costly to apply in very highly parameterized and nonlinear models, which can limit the implementation of this workflow in other coastal settings. Furthermore, FOSM-based uncertainty analysis relies on strong assumptions about model linearity and Gaussian distributions of uncertainty and therefore can only provide approximations of model predictive uncertainty. Stochastic, ensemble-based approaches are needed to provide more reliable uncertainty estimates. Developing an ensemble-based approach for coastal OUU adapted to highly parameterized, nonlinear models would facilitate the implementation of OUU methods in real-world coastal settings.

Coupling climate uncertainty with model parameter uncertainty and understanding the consequences for groundwater management was identified as a prospective topic of seawater intrusion research (Werner et al., 2013). However, current OUU approaches have only considered parameter uncertainty (e.g., in hydraulic conductivity, recharge, porosity and longitudinal dispersivity – Mostafaei-Avandari and Ketabchi, 2020), while climate change projections were incorporated into optimizations as discrete scenarios (e.g., optimization under a projected sea level rise scenario or under four projected recharge scenarios in Roy and Datta, 2018, and Zhao et al., 2021, respectively). The uncertainty associated with climate projections has therefore not been considered, although in the field of hydrology, climate uncertainty is often evaluated using ensembles of climate projections (Mustafa et al., 2019; Al Atawneh et al., 2021). Generating ensembles of climate projections and combining them with model parameter ensembles, within an OUU framework, would enable evaluation of the coupled impacts of climate and parameter uncertainty on groundwater management solutions and lead to more robust decision-making.

The objective of this study was to provide a framework for stochastic, ensemble-based pumping optimization under uncertainty in an island aquifer, considering parameter, historical observation and future climate uncertainty and using methods adapted to very high-dimensional, nonlinear models. An ensemble-based framework was implemented using a sharp-interface modeling approach. A sharp-interface model was first built using MODFLOW-SWI2 (Bakker et al., 2013), after which ensemble-based history matching was conducted using the iterative ensemble smoother PESTPP-IES (White, 2018). Sea level and recharge climate ensembles were generated for the year 2050 and incorporated into the model parameter ensemble. Multi-objective optimization under uncertainty was then conducted via PESTPP-MOU (White et al., 2022) using the ensemble to account for uncertainty in the simulated response to groundwater extraction within the optimization process. Results were compared to both the current water demand and water demand projections for the year 2050. This workflow was applied to a real-world island aquifer in the Magdalen Islands (Quebec, Canada) and was entirely scripted in Python to enhance the transparency and reproducibility of the analyses (e.g., White et al., 2020a; Fienen and Bakker, 2016).

Map view of the Grande Entrée numerical model with the seaward extent, geological formations and locations of pilot points, model observations (including electrical resistivity tomography and time-domain electromagnetic surveys) and pumping wells (domestic wells not shown). The boundary conditions implemented in MODFLOW are a uniform recharge rate on land cells (RCH package), general head boundary conditions for sea cells (GHB package) and groundwater pumping at municipal wells (MNW2 package). Modified from Coulon et al. (2022).

The island of Grande Entrée forms part of the Magdalen Islands archipelago, which is located in the middle of the Gulf of Saint Lawrence (Fig. 1). Fresh groundwater is the only source of drinking water for the sparsely populated island communities and is contained in a lens overlying saline groundwater. The pumping wells of the water supply system are at risk of salinization, since freshwater abstraction leads to the upward migration of saline groundwater towards the wells, a process called saltwater upconing. Deterministic parameter estimation and SLP–FOSM-based OUU approaches were previously implemented on Grande Entrée to determine maximum pumping rates which avoid well salinization considering parameter and observation uncertainty and current climate conditions (Coulon et al., 2021, 2022). This study details the implementation of the ensemble-based approach for history matching and multi-objective (MO) OUU and the incorporation of climate uncertainty (specifically, sea level and recharge uncertainty) into the pumping optimization. Optimization results of the MO-ensemble approach are compared to those of the SLP–FOSM approach.

A detailed description of the study area is available in Coulon et al. (2021). The geology is comprised of a highly permeable and heterogeneous Permian sandstone (Fig. 1) along with Quaternary sand dunes and glacial sediments (mostly fine sand). A spatially distributed recharge representative of the period 1989–2019 was generated for the Magdalen Islands archipelago by Lemieux et al. (2022) using a SWB2 groundwater recharge model (Soil-Water-Balance-2, Westenbroek et al., 2018). A spatial average of 524 mm yr

A sharp-interface, 2D horizontal model with a 20 m

Prior parameter probability distributions of the uncertain model parameters, assumed to be normal or lognormal, described by the mean and the 95 % confidence interval (i.e., the mean

Instead of adopting a deterministic approach, which seeks the minimum error variance parameter set resulting in the best fit to observations, an ensemble-based approach was selected in which multiple, equally plausible realizations of parameter sets were generated (i.e., a parameter ensemble) and carried forward into the analyses (Anderson et al., 2015). The following definitions will be used throughout the paper. A realization is a parameter set; collectively, many realizations form a parameter ensemble. The prior parameter ensemble is the set of realizations that is generated prior to history matching using only expert knowledge to define the statistical distributions of parameters. The posterior parameter ensemble is the set of realizations obtained after history matching (i.e., the prior parameter ensemble updated by observations). The predictive parameter ensemble is the set of realizations obtained after incorporating climate ensembles into the posterior parameter ensemble, all other aquifer parameters remaining constant (i.e., the posterior parameter ensemble updated by climate projections). The size of an ensemble corresponds to the number of realizations comprising the ensemble. The term “ensemble” is used in line with the PEST

Summary of the ensemble-based workflow. An ensemble represents a sample of the probability distribution function (PDF). After generating a prior parameter ensemble by drawing from prior parameter PDFs, (1) ensemble-based history matching yielded a posterior parameter ensemble; (2) a predictive parameter ensemble was obtained by updating the posterior parameter ensemble with year 2050 sea level and recharge projections; and (3) optimization under uncertainty was conducted using both the posterior and predictive parameter ensembles to obtain maximum allowable pumping rates considering parameter and observation uncertainty and either neglecting or accounting for climate projections, respectively.

Fifty-eight model parameters were considered to be adjustable, including 56 hydraulic conductivity values, a uniform recharge rate and a homogeneous transverse dispersivity (Sect. 3.1).

Ensemble-based history matching was then conducted using PESTPP-IES (White, 2018), which implements an iterative ensemble smoother form of the GLM algorithm (Chen and Oliver, 2013). PESTPP-IES was selected because it enables history matching of highly parameterized models with a significantly lower computational effort than sampling methodologies such as Markov chain Monte Carlo. History matching was implemented in a Bayesian framework using the assumptions of multivariate Gaussian prior and posterior distributions. Over successive iterations, PESTPP-IES conditioned the prior parameter ensemble with the information contained in 20 freshwater head observations (extracted from shallow wells, deep open wells and pumping wells) and 142 freshwater–seawater interface elevation observations (derived from deep open wells, TDEM and ERT geophysical surveys, Fig. 1) by minimizing a model-to-measurement fit objective function. Observations were paired with random realizations of measurement noise, and the least-squares objective function was calculated as the sum of squared weighted differences between simulated and observed data (PEST^{®}). History matching yielded a posterior parameter ensemble of size

Prior, posterior and predictive parameter ensembles, described by the ensemble size, the mean and the 5th–95th percentile range. Statistics of the prior parameter ensemble are not identical to those of the prior parameter PDFs, because ensembles represent a sample of probability distributions. During history matching, the sea level,

The sea level and recharge ensembles contained in the posterior parameter ensemble were representative of the 2014–2019 calibration period (Table 2). To make model predictions for the year 2050 while accounting for climate projections, the ensembles were replaced with year 2050 sea level and recharge ensembles (Fig. 2). The climate change predictive simulations were then conducted under steady-state conditions, with boundary conditions representative of the 2050 sea level and recharge conditions. All the predictive simulations were conducted under steady-state conditions because the storage parameters were unconstrained by the history matching and therefore remained highly uncertain. Running model simulations through the predictive parameter ensemble yielded prediction ensembles for the year 2050 conveying parameter, observation and climate uncertainty.

The reference elevation used for the numerical model was the local mean sea level of the Magdalen Islands (Lemieux et al., 2022), and therefore the term “meters above sea level” (or “m a.s.l.”) is used in reference to this elevation. Using the relative sea level rise projections for the study area (Sect. 2) and the current sea level (0.014 m a.s.l.), a 0.204 m a.s.l. mean sea level was expected for 2050, with a normal distribution and a standard deviation of 0.11 m.

Processing of the climate ensembles generated by the SWB2 groundwater recharge model:

Seventy-two projections of daily precipitation and minimum and maximum air temperature were provided by the OURANOS consortium for the Magdalen Islands for the period 2021–2050 (Charron, 2016). These resulted from three emission scenarios (the representative concentration pathways RCP2.6, RCP4.5 and RCP8.5) being run through an ensemble of 24 global climate models (GCMs) of the Coupled Model Intercomparison Project – Phase 5 (CMIP5). These 72 climate projections were run through the SWB2 groundwater recharge model developed by Lemieux et al. (2022) for the Magdalen Islands (Sect. 2) assuming a constant land use, which generated 72 daily recharge projections for the period 2021–2050 (Fig. 3a).

A recharge ensemble for the year 2050 was then extracted (

A

The primary objective of the optimization was to maximize pumping rates in the well field while avoiding well salinization due to upconing. Salinization occurred if the 1 % seawater salinity contour simulated under the well (

Parameters ^{®}).

The optimization problem was solved using the NSGA-II nondominated-sorting genetic algorithm (Deb et al., 2002) implemented in PESTPP-MOU (White et al., 2022) using ensemble-based constraint uncertainty. PESTPP-MOU was selected because it implements a wide range of evolutionary algorithms, which are more effective than traditional optimization methods when solving highly nonlinear optimization problems typical of coastal environments (Ketabchi and Ataie-Ashtiani, 2015). Its use required very little modification to input files after having previously used the PEST

Schematic 1 % seawater salinity contour (

Summary of the main steps undertaken in each generation of the optimization as a result of all the PESTPP-MOU parameters (in green) being set to their default values (modified from Deb et al., 2002). The handling of chance constraints is not shown. More information on differential evolution (DE) and the non-dominated sorting genetic algorithms NSGA-II and PESTPP-MOU can be found in Storn and Price (1997), Deb et al. (2002) and PEST

As the number of realizations increases, the ensembles become more representative of the PDFs that they sample, but computational times increase. The most important factor influencing the outcome of the OUU procedure is whether the statistical moments of the constraint ensembles have converged relative to the ensemble size, since the probabilities of constraint violation are directly used in the optimization algorithm (Eqs. 4, 5). Analyzing the convergence of posterior

Scatter plots of observed vs. simulated data for

Examples of four realizations extracted from the posterior parameter ensemble. Posterior hydraulic conductivity fields (

All projections in the year 2050 sea level ensemble were greater than the current mean sea level (Fig. 8a), with a mean projected sea level rise of 0.2 m and an uncertainty (represented by the standard deviation) of 0.1 m. The

Current vs. year 2050 projected

The 50 % seawater salinity contour (

The OUU procedure for the posterior and predictive parameter ensembles required approximately 89 000 and 86 000 model simulations, respectively, which took a total of 320 h using 130 cores at 2.3 GHz; 34 and 35 Pareto optimal pumping scenarios were identified. The Pareto front, or the optimal trade-off between pumping and probability of well salinization, was identified (Fig. 10), and the optimal allocation of pumping in the well field was determined for a range of probabilities of well salinization (Fig. 11). As the maximum pumping rate in the well field (Fig. 10) and at individual wells (Fig. 11) increased, the probability of well salinization increased as well.

The results of the MO-ensemble approach (when neglecting climate projections) were first compared to those of the SLP–FOSM approach described in Coulon et al. (2022). The maximum allowable pumping rates determined by the MO-ensemble approach were lower for all but highly risk-tolerant stances (Fig. 10), and the difference between both approaches was largest at highly risk-averse stances. For example, for a 6 % probability of well salinization, the MO-ensemble approach found a maximum allowable pumping rate of 98 m

The consequences of neglecting or accounting for climate projections within the MO-ensemble approach were then examined. When climate projections were considered, the maximum allowable pumping rates were lower for risk-averse stances (Fig. 10). Therefore, when considering climate projections, pumping rates needed to be lowered to conserve identical probabilities of well salinization; for example, selecting a 10 % probability of salinization required reducing the total pumping in the well field from 165 to 60 m

Optimal trade-off between the total pumping rate in the well field and the probability of well salinization when neglecting (blue) or accounting for (red) climate projections (ensemble-based approach). The results obtained using the SLP–FOSM approach (neglecting climate change effects) are also shown. The current water demand and the probability density function associated with the projected (year 2050) water demand are superimposed. The

Optimal allocation of pumping in the well field as a function of the probability of well salinization when neglecting (blue) or accounting for (red) climate projections (ensemble-based approach). The results obtained using the SLP–FOSM approach (neglecting climate change effects) are also shown. The

The OUUs identified sets of maximum allowable pumping scenarios corresponding to different probabilities of salinization, from which the groundwater managers can choose depending on their attitude towards risk (Coulon et al., 2022). Different results were obtained when using a MO-ensemble approach or a SLP–FOSM approach to OUU. Since ensemble-based uncertainty quantification is more reliable than FOSM-based uncertainty estimates, the probabilities of well salinization determined by the MO-ensemble approach are more reliable. In the study area, the SLP–FOSM approach may have overestimated the maximum allowable pumping rates compared to the MO-ensemble approach and underestimated the probability of well salinization associated with pumping scenarios, especially for risk-averse stances. The SLP–FOSM approach would have led to the conclusion that the well field is able to supply both the current water demand and the highest water demand projections with a wide margin while maintaining a very low probability of salinization of 2 %. In comparison, the MO-ensemble approach found much lower maximum allowable pumping rates, and if the highest risk-averse stance were adopted (i.e., a 6 % probability of well salinization), the well field would barely be able to meet the current water demand and could only meet the higher water demand projections with an increased probability of well salinization. However, in the MO-ensemble approach, the complete convergence to the Pareto front requires many generations of the NSGA-II algorithm, which in practice is limited by computational constraints. Therefore, the maximum allowable pumping rates found by the MO-ensemble approach for the study area may also be suboptimal and overly conservative.

Incorporating climate projections into the parameter ensembles instead of moving directly from history matching to OUU was essential, since neglecting climate projections resulted in underestimating predictive uncertainties (Fig. 9) and therefore underestimating probabilities of well salinization for risk-averse pumping scenarios (Fig. 10). This was the case even though the climate projections hardly affected the median model predictions (Fig. 9): the simple increase in model predictive uncertainty due to additional consideration of climate uncertainty, which can be expected in other areas, was enough to impact the results of the OUU. Accounting for climate projections in the OUU decreased the maximum allowable pumping rates for users with a risk-averse stance and therefore led to more conservative pumping scenarios. When accounting for climate projections, if the highest risk-averse stance was adopted (i.e., an 8 % probability of well salinization), the well field would meet neither the current water demand nor the lowest water demand projections. Therefore, groundwater managers could either decide to meet the water demand but at higher probabilities of salinization (greater than 10 %–15 %) or to find supplementary sources of water. Multiplying the probability of well salinization by its consequences would allow one to characterize the risk of salinization and therefore assist decision-makers in their evaluation of different management options.

While parameter, observation and climate uncertainties were considered in this study, model conceptual uncertainty was neglected, and using a sharp-interface approach to simulate saltwater upconing could result in increased conceptual uncertainty. However, the posterior parameter values were physically plausible and consistent with the prior parameter distributions, and the information in the observations was appropriately assimilated; these are the two indicators available to detect the potential for conceptual model uncertainty issues. The Doherty and Christensen (2011) model pairing methodology could be used to more explicitly investigate the potential for conceptual model issues through pairing of a sharp-interface model with an advective–dispersive-based variable-density model. In the context of lateral seawater intrusion, methodologies have also been developed to optimize pumping using a coupled sharp-interface or advective–dispersive approach (e.g., Christelis and Mantoglou, 2018; Dey and Prakash, 2022), which could be explored in the context of freshwater lenses. This topic is discussed in more detail in Coulon et al. (2022).

Although the sea level and recharge projections were considered to be independent (Sect. 3.3), sea level rise projections linked to GCMs and RCPs can sometimes be found (e.g., James et al., 2021), and linked sea level recharge ensembles could be used in future OUUs. However, climate uncertainty is always dependent on the size and nature of the projections contained in the climate ensemble (e.g., the number and nature of the applied GCMs and RCPs), and the uncertainty is expected to increase with the ensemble size. Even if all the available GCMs were used, these would still represent a subset of possible future climates (Ray and Brown, 2015). Therefore, the optimal pumping scenarios should regularly be updated with the latest climate projections. Another approach could be to integrate robustness to climate uncertainty into the optimization, aiming to find pumping scenarios that are less sensitive to climate uncertainty and that perform well under a wide range of climate projections (Borgomeo et al., 2018).

This study considered climate change effects under steady-state conditions and explored the management options that would result from the groundwater system (i.e., the freshwater lens) equilibrating with the 2050 climate projections. In reality, since climate change effects are a transient process and this steady state may never be reached, this approach can be viewed as being conservative. When considering transient conditions, climate uncertainties might increase with time and interannual climate variability (e.g., variations in the frequency and intensity of extreme weather events) may also impact the optimization results.

Few examples were found in the existing literature of to how to merge the parameter and climate recharge ensembles (Sect. 3.3.2). These were obtained independently through different types of models: a calibrated groundwater flow model with a simplified representation of recharge (MODFLOW-SWI2) and a more complex, uncalibrated groundwater recharge model able to convert weather projections into recharge projections (SWB2). The groundwater recharge model could have been coupled to the groundwater flow model, with its uncertain parameters added to the history matching procedure, and the weather projections would then be run directly through the coupled models. However, the coupling of numerical models is often computationally demanding. In the approach that was selected, the

Recharge (

Ensemble-based history matching yielded a relatively large range of simulated values for each observation (Fig. 6); however, this could be explained by the cutoff of PESTPP-IES after the second iteration. Over successive PESTPP-IES iterations, the goodness of fit increases and the ensemble diversity (and therefore the posterior parameter variance) decreases (Sect. 3.2). While it is recommended to use a small number of iterations with the IES algorithm (PEST

Using an ensemble-based approach in combination with a population-based evolutionary algorithm such as NSGA-II required a significant number of model simulations. The use of parallel processing was critical to implementing the framework within reasonable computational times and was greatly facilitated by using the PEST

A fully scripted workflow was developed for ensemble-based history matching, incorporation of climate projections and pumping optimization under uncertainty considering parameter, historical observation and future climate uncertainty. The workflow was implemented in a real-world island aquifer. It allowed for the quantification of the optimal trade-off between pumping and the probability of well salinization considering parameter, observation and climate uncertainty simultaneously, letting groundwater managers choose the final pumping scenario depending on their attitude towards risk. Incorporating climate projections into the OUU allows groundwater managers to account for multiple sources of uncertainty simultaneously, i.e., uncertainty arising from the model itself (e.g., parameter and observation uncertainty) and climate uncertainty (e.g., sea level and recharge uncertainty).

The workflow used easily accessible, model-independent tools for ensemble-based history matching and multi-objective optimization under uncertainty that are applicable to high-dimensional, nonlinear models and to nonlinear optimization problems. The approach can therefore be implemented in a large range of coastal settings, including high-dimensional and nonlinear models, provided that model simulations are parallelized. The workflow could also be adapted to other management problems involving optimization under uncertainty. The multi-objective, ensemble-based approach led to much lower maximum allowable pumping rates than the previously applied sequential linear programming, FOSM-based approach (Coulon et al., 2022) for users with risk-averse stances. This was the case even though the numerical model and the optimization problem were relatively linear.

A method for merging parameter and climate recharge ensembles was suggested, and the effect of this coupling on management optimization was explored. Incorporating climate uncertainty into the workflow was critical, since it reduced the maximum allowable pumping rate for users with a risk-averse stance and since neglecting climate uncertainty resulted in underestimating the probabilities of well salinization. In the study area, when considering sea level and recharge projections to the year 2050, the well field would be unable to meet the current water demand or any of the year 2050 water demand projections while maintaining very low risks of well salinization. Conceptual uncertainty was not considered in the analysis, but its evaluation and coupling with the other sources of uncertainty would be useful as it could also impact the optimization results.

The Python scripts used to implement the workflow and produce the figures are available on GitHub at

All the data used by the Python-based workflow are also available on the GitHub repository

CC: conceptualization, formal analysis, methodology, visualization, writing. JTW: formal analysis, methodology, software, writing. AP: conceptualization, methodology, resources, writing. LG: formal analysis, methodology. JML: conceptualization, resources, supervision, writing.

The authors declare that they have no conflict of interest.

Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors.

The authors would like to thank John Molson for proofreading the manuscript and for his insightful comments as well as Sreekanth Janardhanan and an anonymous reviewer for their valuable comments. The authors would like to thank the OURANOS consortium and Marco Braun for supplying the daily precipitation and minimum and maximum air temperature climate simulations for the Magdalen Islands for the period 1950–2100.

This research has been supported by the Québec Ministère du Développement Durable, de l'Environnement et de la Lutte Contre les Changements Climatiques (project Acquisition de connaissances sur les eaux souterraines dans la région des Îles-de-la-Madeleine grant). A scholarship provided by the Canadian National Chapter of the IAH (IAH-CNC) to Cécile Coulon helped finance this study (2021 IAH-CNC Scholarship For Graduate Studies in Hydrogeology). Jeremy White was funded by INTERA Incorporated.

This paper was edited by Alberto Guadagnini and reviewed by Sreekanth Janardhanan and Leonardo Sandoval.