Accurate flood inundation modelling using a complex high-resolution hydrodynamic (high-fidelity) model can be very computationally demanding. To address this issue, efficient approximation methods (surrogate models) have been developed. Despite recent developments, there remain significant challenges in using surrogate methods for modelling the dynamical behaviour of flood inundation in an efficient manner. Most methods focus on estimating the maximum flood extent due to the high spatial-temporal dimensionality of the data. This study presents a hybrid surrogate model, consisting of a low-resolution hydrodynamic (low-fidelity) and a Sparse Gaussian Process (Sparse GP) model, to capture the dynamic evolution of the flood extent. The low-fidelity model is computationally efficient but has reduced accuracy compared to a high-fidelity model. To account for the reduced accuracy, a Sparse GP model is used to correct the low-fidelity modelling results. To address the challenges posed by the high dimensionality of the data from the low- and high-fidelity models, Empirical Orthogonal Functions (EOF) analysis is applied to reduce the spatial-temporal data into a few key features. This enables training of the Sparse GP model to predict high-fidelity flood data from low-fidelity flood data, so that the hybrid surrogate model can accurately simulate the dynamic flood extent without using a high-fidelity model. The hybrid surrogate model is validated on the flat and complex Chowilla floodplain in Australia. The hybrid model was found to improve the results significantly compared to just using the low-fidelity model and incurred only 39% of the computational cost of a high-fidelity model.

Chemical and biological composition of surface materials and physical structure and arrangement of those materials determine the intrinsic reflectance of Earth’s land surface. The apparent reflectance—as measured by a spaceborne or airborne sensor that has been corrected for atmospheric attenuation—depends also on topography, surface roughness, and the atmosphere. Especially in Earth’s mountains, estimating properties of scientific interest from remotely sensed data requires compensation for topography. Doing so requires information from digital elevation models (DEMs). Available DEMs with global coverage are derived from spaceborne interferometric radar and stereo-photogrammetry at ~30 m spatial resolution. Locally or regionally, lidar altimetry, interferometric radar, or stereo-photogrammetry produces DEMs with finer resolutions. Characterization of their quality typically expresses the root-mean-square (RMS) error of the elevation, but the accuracy of remotely sensed retrievals is sensitive to uncertainties in topographic properties that affect incoming and reflected radiation and that are inadequately represented by the RMS error of the elevation. The most essential variables are the cosine of the local solar illumination angle on a slope, the shadows cast by neighboring terrain, and the view factor, the fraction of the overlying hemisphere open to the sky. Comparison of global DEMs with locally available fine-scale DEMs shows that calculations with the global products consistently underestimate the cosine of the solar angle and underrepresent shadows. Analyzing imagery of Earth’s mountains from current and future spaceborne missions requires addressing the uncertainty introduced by errors in DEMs on algorithms that analyze remotely sensed data to produce information about Earth’s surface.

Earth System Models’ complex land components simulate a patchwork of increases and decreases in surface water availability when driven by projected future climate changes. Yet, commonly-used simple theories for surface water availability, such as the Aridity Index (P/E0) and Palmer Drought Severity Index (PDSI), obtain severe, globally dominant drying when driven by those same climate changes, leading to disagreement among published studies. In this work, we use a common modeling framework to show that ESM simulated runoff-ratio and soil-moisture responses become much more consistent with the P/E0 and PDSI responses when several previously known factors that the latter do not account for are cut out of the simulations. This reconciles the disagreement and makes the full ESM responses more understandable. For ESM runoff ratio, the most important factor causing the more positive global response compared to P/E0 is the concentration of precipitation in time with greenhouse warming. For ESM soil moisture, the most important factor causing the more positive global response compared to PDSI is the effect of increasing carbon dioxide on plant physiology, which also drives most of the spatial variation in the runoff ratio enhancement. The effect of increasing vapor-pressure deficit on plant physiology is a key secondary factor for both. Future work will assess the utility of both the ESMs and the simple indices for understanding observed, historical trends.

Water storage plays an important role in mitigating heat and flooding in urban areas. Assessment of the water storage capacity of cities remains challenging due to the inherent heterogeneity of the urban surface. Traditionally, effective storage has been estimated from runoff. Here, we present a novel approach to estimate effective water storage capacity from recession rates of observed evaporation during precipitation-free periods. We test this approach for cities at neighborhood scale with eddy-covariance based latent heat flux observations from fourteen contrasting sites with different local climate zones, vegetation cover and characteristics, and climates. Based on analysis of 583 drydowns, we find storage capacities to vary between 1.3-28.4 mm, corresponding to e-folding timescales of 1.8-20.1 days. This makes the storage capacity at least one order of magnitude smaller than the observed values for natural ecosystems, reflecting an evaporation regime characterised by extreme water limitation.

Tsunami deposits provide information for estimating the magnitude and flow conditions of paleotsunamis, and inverse models have potential for reconstructing hydraulic conditions of tsunamis from their deposits. The majority of the previously proposed models are based on oversimplified assumptions and possess some limitations. We present a new inverse model based on the FITTNUSS model, which incorporates nonuniform and unsteady transport of suspended sediment and turbulent mixing. The present model uses a deep neural network (DNN) for the inversion method. In this method, forward model calculations are repeated for random initial flow conditions (e.g., maximum inundation length, flow velocity, maximum flow depth and sediment concentration) to produce artificial training data sets of depositional characteristics such as thickness and grain size distribution. The DNN was then trained to establish a general inverse model based on artificial data sets derived from the forward model. Tests conducted using independent artificial data sets indicated that this trained DNN can reconstruct the original flow conditions from the characteristics of the deposits. Finally, the model was applied to a data set of 2011 Tohoku-Oki tsunami deposits. The predicted results of flow conditions were verified by the observational records at Sendai plain. Jackknife resampling was applied to estimate the precision of the result. The estimated results of the flow velocity and maximum flow depth were approximately 5.4\pm0.140 m/s and 4.11\pm0.152 m, respectively after the uncertainty analysis. The DNN shows promise for reconstruction of tsunami characteristics from its deposits, which would help in estimating the hydraulic conditions of paleotsunamis.

The spatiotemporal patterns of precipitation are critical for understanding the underlying mechanism of many hydrological and climate phenomena. Over the last decade, applications of the complex network theory as a data-driven technique has contributed significantly to study the intricate relationship between many variable in a compact way. In our work, we conduct a study to compare an extreme precipitation pattern in Ganga River Basin, by constructing the networks using two nonlinear methods - event synchronization (ES) and edit distance (ED). Event synchronization has been frequently used to measure the synchronicity between the climate extremes like extreme precipitation by calculating the number of synchronized events between two events like time series. Edit distance measures the similarity/dissimilarity between the events by reducing the number of operations required to convert one segment to another, that consider the events’ occurrence and amplitude. Here, we compare the extreme precipitation patterns obtained from both network construction methods based on different network’s characteristics. We used degree to understand network topology and identify important nodes in the networks. We also attempted to quantify the impact of precipitation seasonality and topography on extreme events. The study outcomes suggested that the degree is decreased in the southwest to the northwest direction and the timing of peak precipitation influences it. We also found an inverse relationship between elevation and timing of peak precipitation exists and the lower elevation greatly influences the connectivity of the stations. The study highlights that Edit distance better captures the network’s topology without getting affected by artificial boundaries.

The predicted Antarctic contribution to global-mean sea-level rise is one of the most uncertain among all major sources. Partly this is because of instability mechanisms of the ice flow over deep basins. Errors in bedrock topography can substantially impact the projected resilience of glaciers against such instabilities. Here we analyze the Pine Island Glacier topography to derive a statistical model representation. Our model allows for inhomogeneous and spatially dependent uncertainties and avoids unnecessary smoothing from spatial averaging or interpolation. A set of topography realizations is generated representing our best estimate of the topographic uncertainty in ice sheet model simulations. The bedrock uncertainty alone creates a 5% to 25% uncertainty in the predicted sea level rise contribution at year 2100, depending on friction law and climate forcing. Pine Island Glacier simulations on this new set are consistent with simulations on the BedMachine reference topography but diverge from Bedmap2 simulations.

The freshwater ecosystems around the world are degrading, such that maintaining environmental flow (EF) in river networks is critical to their preservation. The relationship between streamflow alterations and, respectively, EF violations, and freshwater biodiversity is well established at the scale of stream reaches or small basins (~<100 km²). However, it is unclear if this relationship is robust at larger scales even though there are large-scale initiatives to legalize the EF requirement. Moreover, EFs have been used in assessing a planetary boundary for freshwater. Therefore, this study intends to carry out an exploratory evaluation of the relationship between EF violation and freshwater biodiversity at globally aggregated scales and for freshwater ecoregions. Four EF violation indices (severity, frequency, the probability to shift to violated state, and probability to stay violated) and seven independent freshwater biodiversity indicators (calculated from observed biota data) were used for correlation analysis. No statistically significant negative relationship between EF violation and freshwater biodiversity was found at global or ecoregion scales. While our results thus suggest that streamflow and EF may not be an only determinant of freshwater biodiversity at large scales, they do not preclude the existence of relationships at smaller scales or with more holistic EF methods (e.g., including water temperature, water quality, intermittency, connectivity etc.) or with other biodiversity data or metrics.

The tandem rise in satellite-based observations and computing power has changed the way we (can) see rivers across the Earth’s surface. Global datasets of river and river network characteristics at unprecedented resolutions are becoming common enough that the sheer amount of available information presents problems itself. Fully exploiting this new knowledge requires linking these geospatial datasets to each other within the context of a river network. In order to cope with this wealth of information, we are developing Veins of the Earth (VotE), a flexible system designed to synthesize knowledge about rivers and their networks into an adaptable and readily-usable form. VotE is not itself a dataset, but rather a database of relationships linking existing datasets that allows for rapid comparison and exports of river networks at arbitrary resolutions. VotE’s underlying river network (and drainage basins) is extracted from MERIT-Hydro. We link within VotE a newly-compiled dam dataset, streamflow gages from the GRDC, and published global river network datasets characterizing river widths, slopes, and intermittency. We highlight VotE’s utility with a demonstration of how vector-based river networks can be exported at any requested resolution, a global comparison of river widths from three independent datasets, and an example of computing watershed characteristics by coupling VotE to Google Earth Engine. Future efforts will focus on including real-time datasets such as SWOT river discharges and ReaLSAT reservoir areas.

Concurrent temperature and precipitation extremes during Indian summer monsoon generally have signicant effects on agriculture, society and ecosystems. Due to climate change, frequency and spatial extent of concurrent extremes have changed, and there is a need to advance our understanding in this domain. Quantication of individual extremes (temperature and precipitation) during the summer monsoon season and its teleconnections to climate indices have been studied comprehensively. But, less attention is devoted to the quantication of concurrent extremes and its teleconnections to climate indices. In this study, concurrent extremes (dry/hot and wet/cold) based on mean monthly temperature and total monthly precipitation during the Indian summer season from 1951 to 2019 over the Indian mainland are investigated. Next, the study uses wavelet coherence analysis to unravel the teleconnections of the spatial extent of concurrent extremes to climate indices (Nino 3.4, WEIO SST and SEEIO SST). Results show that the frequency of wet/hot concurrent extremes has increased signicantly, while the frequency of wet/cold concurrent has decreased for the time window 1985 to 2019 relative to 1951-1984. Also, a statistically signicant increase (decrease) in the spatial extent exists in concurrent dry/hot (wet/cold) extremes during the July, August and September months. The ndings of this study could advance our understanding of changes in concurrent extremes during the Indian summer monsoon due to climate change.

Despite the proliferation of computer-based research on hydrology and water resources, such research is typically poorly reproducible. Published studies have low reproducibility due to incomplete availability of data and computer code, and a lack of documentation of workflow processes. This leads to a lack of transparency and efficiency because existing code can neither be quality controlled nor re-used. Given the commonalities between existing process-based hydrological models in terms of their required input data and preprocessing steps, open sharing of code can lead to large efficiency gains for the modeling community. Here we present a model configuration workflow that provides full reproducibility of the resulting model instantiations in a way that separates the model-agnostic preprocessing of specific datasets from the model-specific requirements that models impose on their input files. We use this workflow to create large-domain (global, continental) and local configurations of the Structure for Unifying Multiple Modeling Alternatives (SUMMA) hydrologic model connected to the mizuRoute routing model. These examples show how a relatively complex model setup over a large domain can be organized in a reproducible and structured way that has the potential to accelerate advances in hydrologic modeling for the community as a whole. We provide a tentative blueprint of how community modeling initiatives can be built on top of workflows such as this. We term our workflow the “Community Workflows to Advance Reproducibility in Hydrologic Modeling’‘ (CWARHM; pronounced “swarm”).

Harsh subsurface environment limits robust workability of on-site instrumentation to be leveraged to track solid Earth’s dynamics. Distributed fiber-optic sensing technology (DFOS) allows long-period in-situ real-time detection of crustal geoenergy exploration-induced underground motions. Here, we first deployed 300-m-long fiber-optic cables behind casing of an actual injection well via single-ended, hybrid Brillouin-Rayleigh backscatterings interrogator to distributed monitor water injection test between two adjacent wells in onshore Mobara, Japan. Detailed DFOS recordings over the entire borehole visualized clear-cut spatiotemporal strain responses from one water injection. Potential injected water-transport footprint and impacted zone reasonably coincided with those of analogy-based strain fronts. Our study thus further uncovered that injection volume and injection pressure significantly dominated water injection-driven strain magnitude and coverage.

The shoreline development index – the ratio of a lake’s shore length to the circumference of a circle with the lake’s area – is a core metric of lake morphometry used in Earth and planetary sciences. In this paper, we demonstrate that the shoreline development index is scale-dependent and cannot be used to compare lakes with different areas. We show that large lakes will have higher shoreline development index measurements than smaller lakes of the same characteristic shape, even when mapped at the same scale. Specifically, the shoreline development index increases by about 14% for each doubling of lake area. These results call into question previously reported patterns of lake shape. We provide several suggestions to improve the application of this index, including a bias-corrected formulation for comparing lakes with different surface areas.

Free alternate bars are large-scale, downstream-migrating bedforms characterized by an alternating sequence of three-dimensional depositional fronts and scour holes that frequently develop in rivers as the result of an intrinsic instability of the erodible bed. Theoretical models based on two-dimensional shallow water and Exner equations have been successfully employed to capture the bar instability phenomenon, and to estimate bar properties such as height, wavelength and migration rate. However, the mathematical complexity of the problem hampered the understanding of the key physical mechanisms that sustain bar formation. To fill this gap, we considered a simplified version of the equations, based on neglecting the deformation of the free surface, which allows us to: (i) provide the first complete explanation of the bar formation mechanism as the result of a simple bond between variations of the water weight and flow acceleration; (ii) derive a simplified, physically based formula for predicting bar formation in a river reach, depending on channel width-to-depth ratio, Shields number and relative submergence. Comparison with an unprecedented large set of laboratory experiments reveals that our simplified formula appropriately predicts alternate bar formation in a wide range of conditions. Noteworthy, the hypothesis of negligible free surface effect also implies that bar formation is fully independent of the Froude number. We show that this intriguing property is intimately related to the three-dimensional nature of river bars, which allows for a gentle lateral deviation of the flow without significant deformation of the water surface.

Alkalinization of natural waters by the dissolution of natural or artificial minerals is a promising solution to sequester atmospheric CO$_2$ and counteract acidification. Here we address the alkalinization carbon capture efficiency (ACCE) by deriving an analytical factor that quantifies the increase in dissolved inorganic carbon in the water due to variations in alkalinity. We show that ACCE strongly depends on the water pH, with a sharp transition from minimum to maximum in a narrow interval of pH values. We also compare ACCE in surface freshwater and seawater and discuss potential bounds for ACCE in the soil water. Finally, we present two applications of ACCE. The first is a local application to 156 lakes in an acid-sensitive region, highlighting the great sensitivity of ACCE to the lake pH. The second is a global application to the surface ocean, revealing a latitudinal pattern of ACCE driven by differences in temperature and salinity.

Water retention curves (WRCs) and hydraulic conductivity functions (HCFs) are critical soil-specific characteristics necessary for modeling the movement of water in soils using the Richardson-Richards equation (RRE). Well-established laboratory measurement methods of WRCs and HCFs are not usually unsuitable for simulating field-scale soil moisture dynamics because of the scale mismatch. Hence, the inverse solution of the RRE is used to estimate WRCs and HCFs from field measured data. Here, we propose a physics-informed neural networks (PINNs) framework for the inverse solution of the RRE and the estimation of WRCs and HCFs from only volumetric water content (VWC) measurements. Unlike conventional inverse methods, the proposed framework does not need initial and boundary conditions. The PINNs consists of three linked feedforward neural networks, two of which were constrained to be monotonic functions to reflect the monotonicity of WRCs and HCFs. Alternatively, we also tested PINNs without monotonicity constraints. We trained the PINNs using synthetic VWC data with artificial noise, derived by a numerical solution of the RRE for three soil textures. The PINNs were able to reconstruct the true VWC dynamics. The monotonicity constraints prevented the PINNs from overfitting the training data. We demonstrated that the PINNs could recover the underlying WRCs and HCFs in non-parametric form, without a need for initial guess. However, the reconstructed WRCs at near-saturation–which was not fully represented in the training data–was unsatisfactory. We additionally showed that the trained PINNs could estimate soil water flux density with a broader range of estimation than the currently available methods.

Fracture networks are abundant in subsurface applications (e.g., geothermal energy production, CO2 sequestration). Fractured reservoirs often have a very complex structure, making modeling flow and transport in such networks slow and unstable. Consequently, this limits our ability to perform uncertainty quantification and increases development costs and environmental risks. This study provides an advanced methodology for simulation based on Discrete Fracture Model (DFM) approach. The preprocessing framework results in a fully conformal, uniformly distributed grid for realistic 2D fracture networks at a required level of precision. The simplified geometry and topology of the resulting network are compared with input (i.e., unchanged) data to evaluate the preprocessing influence. The resulting mesh-related parameters, such as volume distributions and orthogonality of control volume connections, are analyzed. Furthermore, changes in fluid-flow response related to preprocessing are evaluated using a high-enthalpy two-phase flow geothermal simulator. The simplified topology directly improves meshing results and, consequently, the accuracy and efficiency of numerical simulation. The main novelty of this work is the introduction of an automatic preprocessing framework allowing us to simplify the fracture network down to required level of complexity and addition of a fracture aperture correction capable of handling heterogeneous aperture distributions, low connectivity fracture networks, and sealing fractures. The graph-based framework is fully open-source and explicitly resolves small-angle intersections within the fracture network. A rigorous analysis of changes in the static and dynamic impact of the preprocessing algorithm demonstrates that explicit fracture representation can be computationally efficient, enabling their use in large-scale uncertainty quantification studies.

Changes in the magnitude and frequency of river flows have potential to alter sediment dynamics and morphology of rivers globally, but the direction of these changes remains uncertain. A lack of data across spatial and temporal scales limits understanding of river flow regimes and how changes in these regimes interact with river bank erosion and floodplain deposition. Linking characteristics of the flow regime to changes in bank erosion and floodplain deposition is necessary to understand how rivers will adjust to changes in hydrology from societal pressures and climatic change, particularly in snowmelt-dominated systems. We present a lidar dataset, intensive field surveys, aerial imagery and hydrologic analysis spanning 60 years, and spatial analysis to quantify bank erosion, lateral accretion, floodplain overbank deposition, and a floodplain fine sediment budget in an 11-km long study segment of the meandering gravel bed East River, Colorado, USA. Stepwise regression analysis of channel morphometry in nine study reaches and snowmelt-dominated annual hydrologic indices in this mountainous system suggest that sinuosity, channel width, recession slope, and flow duration are linked to lateral erosion and accretion. The duration of flow exceeding baseflow and the slope of the annual recession limb explain 59% and 91% of the variability in lateral accretion and erosion, respectively. This strong correlation between the rate of change in river flows, which occurs over days to weeks, and erosion suggests a high sensitivity of sedimentation along rivers in response to a shifting climate in snowmelt-dominated systems, which constitute the majority of rivers above 40° latitude.

Deep meteoric waters comprise a key component of the hydrologic cycle, transferring water, energy, and life between the earth's surface and deeper crustal environments, yet little is known about the nature and extent of meteoric water circulation. Using water stable isotopes, we show that maximum circulation depths of meteoric waters across North America vary considerably from 1 to 5 km, with the deepest circulation in western North America in areas of greater topographic relief. Shallower circulation occurs in sedimentary and shield-type environments with subdued topography. The amount of topographic relief available to drive regional groundwater flow and flush saline fluids is an important control on the extent of meteoric water circulation, in addition to permeability. The presence of an active flow system in the upper few kilometers of the Earth's crust and stagnant brines trapped by negative buoyancy offers a new framework for understanding deep groundwater systems.

Tropical rivers constitute a major portion of the global aquatic C flux entering the ocean, and the Rio Negro is one of the largest single C exporters with a particularly high export of terrestrial C. We investigated the role of whitesand ecosystems (WSEs) in blackwater formation in the Rio Negro basin to develop novel constraints for the terrestrial carbon export from land to the aquatic continuum. To this end, we used ultrahigh resolution mass spectrometry (FT-MS, Orbitrap) to identify markers in dissolved organic carbon (DOC) from ground- and surface waters of two contrasting WSEs feeding Rio Negro tributaries, and compared them with known Rio Negro marker from two openly available FT-MS datasets. Tributaries were fed by a whitesand riparian valley connected to terra firme plateau, and a typical upland whitesand Campina. WSE-DOC molecular composition differed by 80% from plateau DOC, which was characterized by reworked, highly unsaturated N- and S-containing molecules. WSE-DOC contained mainly condensed aromatics and polyphenols. WSE samples differed by 10% in molecular DOC composition and also by their isotopic content (14C, 18O, 2H). Upland WSE-DOC was exported by fresh precipitation and had maximum age of 13 years, being five years older than riparian valley WSE-DOC. Unexpectedly, only markers from the upland WSE, which cover a small proportion of the landscape, were identical to Negro markers. Markers of the riparian valley WSE, which are widespread and known for high DOC export, surprisingly showed lower coverage with Negro markers. Analysis of robust matching WSE markers between FT-MS datasets by Pubchem suggested well-known plant metabolites (chromenes and benzofurans) as promising candidates for targeted approaches and calibration. Our results suggest that terrestrial DOC from upland WSEs is a main source of specific blackwater molecules missing in the regional ecosystem C balance, whereas C export from the riparian valley and especially from terra firme plateaus represents mainly recycled and transformed carbon not directly affecting the ecosystem C balance. Our study highlights the potential of high-resolution techniques to constrain carbon balances of ecosystems and landscapes. Comparisons of FT-MS datasets and complementary isotopic information shows high potential to identify robust molecular markers that link forests, soils, aquifers and aquatic systems, and are needed for a deeper understanding of the regional C cycle in tropical blackwater catchments.

Large boulders with a diameter of up to several tens of meters are globally observed in mountainous bedrock channel environments. Recent theories suggest that high concentrations of boulders are associated with changes in channel morphology. However, data are scarce and ambiguous, and process-related studies are limited. Here we present data from the Liwu River, Taiwan, showing that channel width and slope increase with boulder concentration. We apply two mass balance principles of bedrock erosion and sediment transport and develop a theory to explain the steepening and widening trends. Five mechanisms are considered and compared to the field data. The cover effect by immobile boulders is found to have no influence on channel width. Channel width can partially be explained by boulder control on the tools effect and on the partitioning of the flow shear stress. However, none of the mechanisms we explored can adequately explain the scattered width data, potentially indicating a long-timescale adjustment of channel width to boulder input. Steepening can be best described by assuming a reduction of sediment transport efficiency with boulder concentration. We find that boulders represent a significant perturbation to the fluvial landscape. Channels tend to adjust to this perturbation leading to a new morphology that differs from boulder-free channels. The general approach presented here can be further expanded to explore the role of other boulder-related processes.

High quality citizen science data can be instrumental in advancing science toward new discoveries and a deeper understanding of under-observed phenomena. However, the error structure of citizen scientist (CS) data must be well-defined. Within a citizen science program, the errors in submitted observations vary, and their occurrence may depend on CS-specific characteristics. This study develops a graphical Bayesian inference model of error types in CS data. The model assumes that: (1) each CS observation is subject to a specific error type, each with its own bias and noise; and (2) an observation’s error type depends on the error community of the CS, which in turn relates to characteristics of the CS submitting the observation. Given a set of CS observations and corresponding ground-truth values, the model can be calibrated for a specific application, yielding (i) number of error types and error communities, (ii) bias and noise for each error type, (iii) error distribution of each error community, and (iv) the error community to which each CS belongs. The model, applied to Nepal CS rainfall observations, identifies five error types and sorts CSs into four model-inferred communities. In the case study, 73% of CSs submitted data with errors in fewer than 5% of their observations. The remaining CSs submitted data with unit, meniscus, and unknown errors. A CS’s assigned community, coupled with model-inferred error probabilities, can identify observations that require verification. With such a system, the onus of validating CS data is partially transferred from human effort to machine-learned algorithms.

Geothermal activity in the Chilean Southern Volcanic Zone is strongly controlled by the regional Liquiñe-Ofqui Fault System (LOFS) and the Andean Transverse Faults (ATF). We analyzed fifteen thermal springs in the Liquiñe area to assess the origin and the main physicochemical processes related to the LOFS and ATF. Major, minor and trace elements identify two defined clusters spatially related to the regional fault systems. In both clusters, ionic relationships suggest that the principal hydrogeochemical processes are mainly dominated by water-rock interactions. Factorial analysis provided two factors: i) F1 (65.1%), saturated by Cl, HCO, Na, SiO, Li, B and Cs, represents water-rock interaction processes driven by temperature in presence of CO; ii) F2 (28.5%) represented by SO and Mo, represents a minor water-rock interaction enhanced by the presence of HS. Samples associated to the LOFS have high scores of both factors, while those from the ATF have only high factor 1 scores. Ionic ratios compared with literature data, clearly identify the samples spatially associated to the LOFS from the ones associated to the ATF with a fuzzy pattern. Water stable isotopes values suggest a meteoric origin with small deviations from local meteoric isotopic line. CO exchange with slightly high and low temperature water rocks interaction is present in most of the samples. Our results indicate that groundwater circulation along faults is a complex process where different constraints influence the final hydeogeochemistry and reaction intensity. Finally, the established processes at Liquiñe area are not upscalable at the whole Southern Volcanic Zone.

Machine learning (ML) based models have demonstrated very strong predictive capabilities for hydrologic modeling, but are often criticized for being black-boxes. In this paper we use a technique from the field of explainable AI (XAI), called layerwise relevance propagation (LRP) to “open the black box”. Specifically we train a deep neural network on data from a set of hydroclimatically diverse FluxNet sites to predict turbulent heat fluxes, and then use the LRP technique to analyze what it learned. We show that the neural network learns physically plausible relationships, including different ways of partitioning the turbulent heat fluxes according to moisture or energy limiting characteristics of the sites. That is, the neural network learns different behaviors at arid and non-arid sites. We also develop and demonstrate a novel technique that uses the output of the LRP analysis to explore how the neural network learned to regionalize between sites. We find that the neural network primarily learned behaviors that differed between evergreen forested sites and all other vegetation classes. Our analysis shows that even simple neural networks can extract physically-plausible relationships and that by using XAI methods we can learn new information from the ML-based methods.