SWMMLIDopt: a tool for optimization of low-impact development (LID) measures using the SWMM model

Low-impact development (LID) measures, such as bio-retention cells, green roofs, and permeable pavements, are examples of local stormwater management that provide adaptation to and mitigation of the negative impacts of climate change and urbanization. Many of these solutions are often nature-based solutions (NBSs), e.g., bio-retention cells, green roofs, and in contrast to conventional stormwater management pipes, implementation of LID techniques reduces the risk of urban pluvial flooding while improving downstream water quality through evaporation, storage, and infiltration processes (Géhéniau et al. 2015; Hunt et al. 2006; Johannessen et al. 2018; Støvring et al. 2018). In addition to benefits regarding water quantity and quality, implementation of NBS provides co-benefits (Raymond et al. 2017). NBS with stormwater management capabilities mitigate urban heat island effects (Susca et al. 2011), reduce the energy consumption of buildings (Refahi & Talkhabi 2015), enhance a city's visual amenity (Jungels et al. 2013), provide health benefits (Tzoulas et al. 2007), and improve a city's biodiversity (Wooster et al. 2022).

Several studies have examined the hydrological impacts of implementing LID measures at the catchment scale. For example, Palla & Gnecco (2015) investigated the hydrological benefits of implementing green roofs and permeable pavements in an urban catchment in Genoa, Italy. They found that at least 5% of impermeable surface must be retrofitted by LID as source control measure to achieve notable hydrological benefits. Likewise, Hamouz et al. (2020) examined the impact of retrofitting green roofs in a small urban catchment. They reported that a minimum of 11% of roofs should be retrofitted by green roofs to obtain noticeable reduction of maximum flow at the catchment outlet. Similar results are reported by Fan et al. (2022) who studied the optimal placement of LID to reduce overflow runoff to provide pollution control. They found that LID measures should be installed downstream in the catchment to maximize the effect of the measure. Skaugen et al. (2020) tested the impact of implementing rain gardens in an urban catchment with 500 houses. They concluded that 60 rain gardens, with a unit area of 10 m², would be required to minimize the peak flow by 10%. Thus, the hydrological benefits of incorporating LID measures in urban stormwater systems are evident.

The spatial placement of LID measures in urban catchments was found by recent studies to significantly impact the robustness of a stormwater management plan. Le Floch et al. (2022) compared several scenarios for spatial placement of LIDs in a small urban catchment in Norway. They found that placing LID measures in the downstream parts of the catchment yielded the highest reduction of peak drainage outflows, while implementing LIDs in upstream sub-catchments was the best strategy for reducing flooding from manholes. A similar finding was reported in an earlier study by Giacomoni & Joseph (2017). Hou et al. (2020) and Ercolani et al. (2018) recommended placing LID measures near areas prone to flooding, while Yao et al. (2020) and Liang et al. (2020) found the placement of LID measures on areas directly connected to the drainage pipe network to be the most cost-effective strategy.

Hydrological models are robust tools for evaluating the performance of LID measures in urban catchments. The stormwater management model (SWMM) software (Rossman 2015) is one of the most commonly used tools among academia for analyzing the hydraulic-hydrological processes in urban catchments. SWMM is an open-access software that allows users to build models of urban catchments with LID measures and to perform rainfall–runoff modeling, water quality modeling, simulation of pipe flow, and snow modeling, among others (Rossman 2015). Various tools have been developed to extend the capabilities of the original SWMM software. For instance, Warsta et al. (2017) developed GisToSWMM5, which is a tool that automates the construction of SWMM models from GIS data. Leutnant et al. (2019) developed the SWMMR library to operate SWMM with the R programming language. Similarly, McDonnell et al. (2020) created a library (PySWMM) to use SWMM with the Python programming language. The SWMM2PEST (Lin et al. 2019) and the OSTRICH-SWMM (Shahed Behrouz et al. 2020) tools were developed for the automatic calibration of the SWMM model.

A handful of tools have been developed to optimize placement of LID measures at a catchment scale, based on coupling SWMM software with multi-objective optimization algorithms. These include the GIP-SWMM (Shojaeizadeh et al. 2021), the OSTRICH-SWMM (Macro et al. 2019) and the RSWMM-cost (Alamdari & Sample 2019). These tools are valuable for providing optimal placement of LID measures under various conditions. However, most of these tools require high competence in programming languages, which makes them less accessible for stormwater practitioners (i.e., engineers and city planners). For instance, a Python library must be used to prepare the optimization input for the OSTRICH-SWMM tool, and RSWMM-cost is developed as a series of codes in R. The GIP-SWMM tool can be operated from a simple user interface. However, it optimizes LID placements based on volume reduction rates of drainage flow, using mass balance calculations. Le Floch et al. (2022) concluded that the spatial placement of LID measures is not the most important factor, if the goal is to reduce the volume of the drainage flow. In contrast, LID placement is important to reduce peak drainage flow rates and manhole flooding (Giacomoni & Joseph 2017; Le Floch et al. 2022).

The implementation of LID measures is essentially a question of costs and currently either multi-objective optimization or a cost–benefit analysis has been used for assessing alternatives. Possantti & Marques (2022) developed a framework for larger watersheds to make long-term scenarios of NBS implementation, coupling hydrological modeling with hydro-economic benefits. Jin et al. (2024) used multi-objective optimization to find the optimal solution of LID and used both hydrological, ecosystem services and cost criteria. Eskandaripour & Soltaninia (2024) also used a multi-objective optimization approach of LID but focused on pollutant control and costs criteria rather than volume runoff and costs criteria. Both Wu et al. (2024) and Osheen et al. (2024) have used a cost–benefit analysis approach to select optimal LID implementation scenarios.

Several studies considering implementation of specific LIDs provide costs estimates (e.g. (Eskandaripour & Soltaninia 2024; Jin et al. 2024) will be relevant for the location where the study is conducted, but may lead to misleading results if used elsewhere. Another aspect is which type of costs to present. Khadka et al. (2021) noted that the estimation of additional costs for LID measures instead of absolute costs is useful to promote LIDs, because property developers could be less hesitant to choose LID measures if only the additional costs are given. Apart from these recent studies, combination of both hydrological performance and costs is limited in computational stormwater management design, and there is room for flexible and efficient LID optimization tools to support cost-effective planning.

The main objective of the current study is to introduce a new tool, SWMMLIDopt, for optimizing the placement and extent of LID measures in urban catchments using multiple criteria. The tool is based on SWMM software which is coupled with a multi-objective optimization algorithm in a user-friendly environment. A specific goal is to demonstrate the capabilities of the SWMMLIDopt tool in real urban catchment settings. For this purpose, SWMMLIDopt was employed to optimize LID measures for two urban catchments with varying degrees of complexity (i.e., number of sub-catchments) and practical optimization constraints. The studied LID types included green roofs, permeable pavements, bio-retention cells, and vegetative swales.

The ZEB Laboratory is part of the NTNU campus in the city of Trondheim, Norway. The climate is subpolar oceanic (Dfc), according to the Köppen-Geiger climate classification (Kottek et al. 2006), with an average annual precipitation of 1,070 mm. The area of the NTNU campus is about 50 ha, with a high degree of developed area (around 60%). The built-up area on the campus will be extended in the next decades, especially in the southern parts where the current imperviousness share is low. In this study, the southern part, representing 10% of the total campus area, was selected to demonstrate the use of the SWMMLIDopt tool for optimizing placement of LID measures using a simplified delineation of the catchment. The NTNU study catchment has well-defined boundaries and was divided into four sub-catchments, separated by roads and sidewalks, as shown in Figure B1 (Supplementary Material). Surface runoff from all four sub-catchments is directed to a road located in the middle of the area, which serves as a common flood pathway from all sub-catchments (shown as open drainage pipe in the Figure B1, Supplementary Material). The catchment's outlet is close to the ZEB Laboratory, where the stormwater is drained to the municipal pipe network. The pipe network is outdated and has limited hydraulic capacity, and the downstream infiltration potential is low. Hence, most of the stormwater should be managed locally and within the catchment boundaries. Municipal regulation in Norway will provide a discharge permit for the ZEB Laboratory that limits the discharge to the municipal pipe network related to a design rainfall event with a return period of 20 years. The design rainfall event used for the ZEB Laboratory case is shown in Figure B3-A (Supplementary Material).

The generated SWMM model for VK consists of a detailed sub-catchment discretization based on land cover types and contains 612 sub-catchments representing dominant land cover (Figure 2). The SWMM model has the following areal distribution: 38% vegetation, 19% roofs, 14% parking plots, 14% walkways, 6% roads, 6% sand/gravel, 2% stone/tile pavers, and 0.4% is covered with bare rock. A more detailed description of the VK catchment and the urban hydrological measurements can be found in Sillanpää (2013). The VK SWMM model has been calibrated and validated in an earlier study against outflow measurements in Vallikallio (e.g., Tuomela et al. 2019). Figure B3-B shows the rainfall event used in the optimization for the Vallikallio catchment. This event has been estimated to have a return period of 1.3 years.

The SWMM is a suite of methods to simulate urban hydrological processes and stormwater quality (Rossman & Huber 2016). The model embeds a spatial description of urban sub-catchments, which supports both detailed description of homogeneous surfaces (Krebs et al. 2014) and description of heterogeneous sub-catchments (Guan et al. 2015) with averaged surface properties. The model simulates evaporation from water retained on surfaces, infiltration, and generation of surface runoff. Snow and groundwater routines are also available for the simulation of snow processes and groundwater storage. Runoff is routed laterally between sub-catchments and to the stormwater drainage network. The stormwater network model supports simulation of open channel and pipe network systems. SWMM is supplied with computation tools to simulate retention and detention of water in LID measures. The LID structures are parameterized as combination of layers with user-specified properties and boundary conditions allowing drainage or infiltration from the structures (Khadka et al. 2021).

SWMM has been benchmarked in case studies both for small urban catchments (Krebs et al. 2014) and for detailed studies of selected LID measures (Hamouz et al. 2020). The applications and the extensive SWMM manuals (Rossman 2015; Rossman & Huber 2016) provide recommendations for parameterization of model components, which has been the starting point for the LID optimization in the current study.

In the Vallikallio and NTNU catchments, the stormwater network hydraulics are simulated with the dynamic wave routing method with a time step of 5 s. Infiltration processes in the previous sub-catchments are calculated with the modified Horton method. The calibrated SWMM model parameterization for the Vallikallio catchment was adopted from Tuomela et al. (2019), while the parameters of NTNU catchment were taken from SWMM manual.

Currently, there are a total of eight different LID measures incorporated in the SWMM engine that can be assigned to a sub-catchment. This study uses the four most typically applied and researched ones: green roof (GR), permeable pavement (PP), bio-retention cell (BR), and vegetative swale (VS). The LID parameters were selected from the SWMM manual (Rossman 2015) and are summarized in Table 1.

The exact plans for the NTNU campus development are not completed yet, giving very few constraints to the selection of the LID measures for the ZEB Laboratory case. Hence, the only general constraints applied were that the LIDs should not exceed 80% of the area of each sub-catchment and that green roofs can only be placed in the sub-catchment near the outlet that contains most of the buildings (Figure B1, Supplementary Material).

In contrast, the Vallikallio catchment in Finland is already fully developed with a separate stormwater network system (Figure 2). This sets logical constraints to the possible application and extent of the LID measures. Several logical expressions were incorporated according to the land use type (Table 2). Additional criteria were set for each land use and LID combination: (1) the maximum GR and PP share is 80% of the sub-catchment area; (2) BR and VS can only cover up to 20% of the sub-catchment total area; (3) the total combined LID percentage of a sub-catchment area cannot exceed 80%.

SWMMLIDopt operates from a simple, easy-to-follow user interface that prepares the optimization input, executes the multi-objective optimization, based on a reduction of peak drainage flow rates and total costs, and finally visualizes the optimization results. SWMMLIDopt was developed using several R packages as stated in Table 3. The user interface of SWMMLIDopt, displayed in Appendix A, Supplementary Material.

SWMMLIDopt aims at finding values of (i.e., usage of each LID measure i at each sub-catchments ⁠) based on the two objective functions and the constraints. The user can select an upper limit for each LID control (e.g., ⁠, where ⁠. is the upper bound on the usage of the ith LID control).

The non-dominated sorting genetic algorithm (NSGA-II) (Deb et al. 2002) is used in SWMMLIDopt, which is widely used for multi-objective optimization. The multi-objective optimization aims at obtaining a set of non-dominated scenarios of LID measures from all available scenarios. This set of scenarios is referred to as the Pareto front.

A scenario X dominates (i.e., is better than) another scenario Y if the following two conditions are fulfilled:

• a. X is not worse than Y in any objective (i.e., cost and peak outflow)

• b. X is better than Y in at least one objective.

If these conditions do not hold, scenarios X and Y are said to be non-dominated to each other..

The steps of the NSGA-II algorithm are as follows: (i) The algorithm generates a random set of scenarios with LID measures, based on defined limits on each LID measure defined by the user; (ii) The cost and peak outflow associated with each scenario from (i) are determined; (iii) The algorithm ranks the scenarios from step (ii) based on their level of non-domination; (iv) Half of the scenarios from step (iii) are selected as ‘Parent scenarios’ by repeatedly choosing the best of two randomly selected scenarios from the generation according to their level of non-domination; (v) The Parent scenarios from (iv) are used to create ‘Offspring scenarios’ through selection, crossover and mutation operations (Deb et al. 2002). These offsprings are combined with their Parent to form a new generation; (vi) The NSGA-II algorithm repeats steps (i–v) until the maximum number of generations, set by the user, is reached. The final generation is then presented as the Pareto optimal set of scenarios of LID measures.

Different optimization metrics, such as runoff volume, maximum peak runoff and costs, may be used to optimize the selection and placement of LID measures. Other studies have used similar metrics, e.g., flood volume and total costs (Liu et al. 2019) and flood risk reduction rate, life cycle costs and land occupied (Wang et al. 2021). In the ZEB Laboratory case, the important hydrological metric is the maximum peak runoff to meet municipal discharge permit regulations, whereas in the Vallikallio case the optimization should make sure that the maximum peak runoff does not exceed the capacity of the existing pipe network. Hence, the maximum peak runoff and LID costs have been selected as optimization metrics to demonstrate SWMMLIDopt.

The costs of the LID measures were adopted from Khadka et al. (2021), who used German design guidelines and cost estimates to quantify the cost of constructing LID measures. In this case, the cost is a difference between the expenses of implementing LID measures and the expenses of implementing a conventional measure. The conventional measure was the construction of conventional impervious urban surfaces, such as asphalt on ground or flat roof on buildings, with runoff to a drainage pipe system. It should be noted that the costs by Khadka et al. (2021) did not include maintenance costs, only the additional investment costs. The additional investment cost for LID measures used in this optimization study are shown in Table B1, Supplementary Material. It should be noted that SWMMLIDopt facilitates the use of local cost data by providing a simple user interface to include this for each of the LID solutions. This will increase the relevance of the results for the decision-maker and take into account national/local costs.

Scenario #100 in Figure 2(c) can be considered as a medium peak outflow reduction scenario (approximately 450 l/s) associated with average additional costs. Figure 3(c) shows that for scenario #100 most of the LIDs are located in the lower left sub-catchment, whereas Figure 3(d) for scenario #1 the LIDs are placed in all sub-catchments and in all available areas applicable for LID installments.

Municipal regulation for runoff from the ZEB Laboratory will result in a discharge permit that allows maximum peak discharge of 400 l/s to the municipal pipe network for the design rainfall event with 20 years return period. Figure 3(a) shows that in order to achieve this requirement, additional investment costs of approximately 0.9 million euros must be expected using the selected LIDs.

SWMMLIDopt provides a user-friendly interface that allows the practitioners to set constraints, physical characteristics and costs relevant to the case in question. As demonstrated for the two catchments presented in this paper, SWMMLIDopt tackles varying degrees of spatial resolution and optimization constraints. Furthermore, the visualization and diagnose features built-in in the tool provides transparency in the optimization process, facilitating a comprehensive understanding and comparison of alternative LID scenarios. Hence, SWMMLIDopt fills the gap between suggested multi-objective Pareto-optimization strategies (Jin et al. 2024) and the existing tools, e.g. that currently target large-scale overarching planning scenarios (Possantti & Marques 2022).

The visualization capabilities of SWMMLIDopt are also useful to assess whether discharge limits are met, and if not, what will the costs be to meet the discharge limits. Alternatively, a practitioner can assess the estimated discharge reduction a given sum of investment in LID measure can result in. Furthermore, as the Pareto-front illustrates the relationship between costs and discharge reduction, its slope provides insights into whether significant discharge reductions per unit cost can be expected. One would typically expect a low reduction in discharge per unit cost when starting to install the measures (as shown in Figure 5(a), left part of the Pareto-front), indicating a minimum cost threshold before LID measures have a notable impact on catchment discharge. On the other hand, after achieving a certain discharge reduction, further reductions might require unreasonably high costs, highlighting the practical discharge limits of LID measures under local conditions and constraints.

SWMMLIDopt currently generates a multi-objective Pareto-front using two optimization criteria. However, additional criteria may be needed to account for the broader co-benefits of LIDs (Jin et al. 2024). The current version of SWMMLIDopt includes two objective functions, cost and peak outflow, but future updates are planned to incorporate additional objective functions to better account for LID co-benefits.

Although LID techniques are primarily used for their hydrological benefits (runoff peak and volume reduction), some types of LID measures are associated with additional environmental and socioeconomic benefits (El Hattab et al. 2018). Reported additional benefits are e.g. mitigation of urban heat island effects and greening of cities (Gunawardena et al. 2017), increased ecosystem services (Possantti & Marques 2022) and increased pollution control (Eskandaripour & Soltaninia 2024). The optimization algorithm may be extended to other objectives. However, adding such co-benefits requires that they are quantified, and that the quantification can be expressed in an equation form in the model. This is typically not possible for many of the co-benefits associated with LIDs. Furthermore, adding multiple additional objectives to the optimization will increase the required input, and may give excessive computation times rendering the tool impractical and ineffective in use.

Co-benefits of stormwater management by LID measures can, however, be considered in an integrated sustainability assessment (Helness et al. 2017, 2019) where several scenarios of LID measures from the Pareto-front can be compared with metrics that may also come from qualitative assessments by a decision-maker or through interviews with stakeholders. Such an inclusion of co-benefits of LIDs in the decision-making process will contribute to a level playing field when comparing LID measures with traditional measures. This ensures that the scenario chosen for implementation maximizes the overall benefits to a sub-catchment where LIDs are installed.

The optimization tool is useful when stormwater management is assessed in the light of climate change and its impacts on urban hydrology. SWMMLIDopt can then guide decision makers to explore which LIDs are able to mitigate future precipitation patterns best and where the optimal placements of such measures should be in the urban catchment. The climate warming in northern areas leads to strong shifts in the areas where air temperature is frequently near the freezing point. At the same time, drainage outflows are projected to increase in most seasons (Pang et al. 2022). Tamm et al. (2023) compiled recent high spatial and temporal resolution climate change scenarios for the Vallikallio site and simulated the climate change impacts on year-round water balance components including urban snow water equivalent and stormwater discharge. Di Natale et al. (2023) explored hydrological performance of LIDs under climate change, when LIDs were designed to reduce summer season runoff. They found that part of the LIDs’ capacity to reduce runoff is used to mitigate climate change when future precipitation volumes are increasing. On the other hand, the changing snowmelt and winter-spring runoff regime may lead to unexpected effects of LIDs on future hydrology. SWMMLIDopt provides a flexible tool to embed multiple optimization criteria for a more covering and systematic assessment of climate resilient LID designs in changing climatic conditions.

The strengths of the SWMMLIDopt tool, in comparison to similar tools reported in the literature, lie in its ability to offer intuitive visualization and comparative analysis of Pareto-optimal scenarios in a user-friendly environment. Furthermore, SWMMLIDopt's capability to diagnose and interpret optimization results (as demonstrated in Figure 7) provides greater transparency in the optimization process, facilitating a more comprehensive understanding of the results by stormwater practitioners.

While SWMMLIDopt offers a range of advantages, the SWMM model resolution may limit the possibility to differentiate between LID measures. This can be overcome by sufficient detail of the SWMM catchment model (Krebs et al. 2014). However, increasing spatial resolution leads to an increase of computation time required for optimization. This is primarily due to the properties of the NSGA-II algorithm that necessitates a large number of SWMM catchment model simulations to estimate the Pareto-front. This study used a single rainfall event approach to reduce the simulation time of the SWMM catchment models. It should be mentioned that such single rainfall simulations only allow for calculating event-based indicators such as peak reduction and peak delay for evaluating the performance of LID measures. However, many studies discussed limitations of these indicators (Stovin et al. 2017; Abdalla et al. 2021) such as their sensitivity to rainfall event characteristics and initial conditions. As such, performance indicators that rely on long-term continuous simulations were recommended in the literature (Hernes et al. 2020; Abdalla et al. 2021).

Due to the increasing computation cost in longer simulations and larger urban catchments, it can be challenging to produce performance indicators and optimize LID designs with SWMMLIDopt. (Shojaeizadeh et al. 2021) discuss this limitation of many current optimization tools. They propose an optimization tool for LID placement that relies only on one single simulation result of a SWMM catchment model, which significantly reduces optimization time. However, their methodology only uses volume reduction as a performance indicator for LID placement – which was found to be insensitive to the spatial arrangement of LID measures by (Le Floch et al. 2022).

To address the issue of a large number of required simulations, a more efficient optimization algorithm can be adopted for estimating the Pareto-front. For instance, the multi-objective Bayesian optimization algorithm (Emmerich et al. 2006) that requires significantly lower number of simulations in comparison to the NSGA-II algorithm (Binois & Picheny 2019). Another alternative is to use a computationally efficient surrogate model that represents the SWMM catchment model for the optimization process. This idea was recently proposed and tested by (Yang et al. 2023), who applied trained machine learning models as surrogate models for optimizing LID placement scenarios, which reduced the optimization time by 95%.

To control the computational demands, the optimization of LID designs can be split into (1) screening of a large number of design alternatives and (2) detailed assessment of pre-screened designs with easier inclusion of co-benefits in the analysis. The first step can be based on short-term simulations, whereas long continuous simulations are adopted in the second step. SWMMLIDopt is an efficient tool to the first step, and applicable to the second as well.

An open-source optimization tool, SWMMLIDopt, was introduced for computational assessment of LID alternatives parameterized with SWMM. The optimization was demonstrated for two northern catchments, one under construction with minimal constraints on LID design and one existing with retrofitted LIDs.

Once the cost information about LID construction was given, SWMMLIDOpt produced a pareto-optimal relationship between LID design costs and impact on catchment discharge over a design rainfall event. In the newly constructed ZEB Laboratory catchment, SWMMLIDopt indicated permeable pavements to be the most cost-effective, while for the retrofitted Vallikallio catchment, the optimization algorithm prioritized green roofs until most roofs in the catchment were retrofitted. The algorithm then favored the placement of permeable pavements in parking lots and walkways. The differences in the densities of bio-retention cells and vegetative swales were minor between the scenarios.

The results demonstrated the prospects of SWMMLIDopt for assessing the optimal combination of LID measures and where these should be placed in an urban catchment. The results for the two catchments demonstrated that SWMMLIDopt successfully deals with varying degrees of spatial resolution and optimization constraints. Optimization results also indicated that the sufficiently high resolution of the hydrological catchment model allows the user to rank the LID measures.

Co-benefits of stormwater management by LID measures should be included in the optimization criteria and the decision-making process to create a level playing field when comparing LID measures with traditional measures. Combined criteria ensure that the scenario chosen for implementation maximizes the overall benefits of LID installation.

The strengths of SWMMLIDopt lie in the visualization and comparative analysis of Pareto-optimal scenarios. The capability to diagnose and interpret optimization results provides transparency in the optimization process, facilitating a comprehensive understanding of the results by stormwater practitioners. However, there is a tradeoff between the required model resolution to differentiate between LID measures and the computation time required for optimization. The tradeoff is primarily due to the use of the NSGA-II algorithm that necessitates a large number of SWMM simulations to accurately estimate the Pareto-front. Development of more computationally efficient models to represent the catchment in the optimization and introduction of prescreening of LID designs before extensive optimization is suggested as a strategy to control computational demands.

This work was supported by the EviBAN project (Evidence based assessment of NWRM for sustainable water management). EviBAN was funded under the Water JPI 2018 Joint Call – WaterWorks 2017 on ‘Closing the Water Cycle Gap – Sustainable Management of Water Resources’, the Research Council of Norway (grant no. 300281), and the Academy of Finland (grant no. 326787).

The source codes and tools developed during this research are freely and publicly accessible under the GNU General Public License. They are hosted on GitHub at https://github.com/ElhadiMohsenAbdalla/SWMMLIDOPT and are also available through the EviBAN toolbox for adaptive stormwater management at https://www.sintef.no/projectweb/eviban/toolbox/.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.