Machine learning (ML) models have become popular in the Earth Sciences for improving predictions based on observations. Beyond pure prediction, though, ML has a large potential to create surrogates that emulate complex numerical simulation models, considerably reducing run time, hence facilitating their analysis.

The behaviour of eco-geomorphological systems is often examined using minimal models, simple equation-based expressions derived from expert knowledge. From them, one can identify complex system characteristics such as equilibria, tipping points, and transients. However, model formulation is largely subjective, thus disputable. Here, we propose an alternative approach where a ML surrogate of a high-fidelity numerical model is used instead, conserving suitability for analysis while incorporating the higher-order physics of its parent model. The complexities of developing such an ML surrogate for understanding the co-evolution of vegetation, hydrology, and geomorphology on a geological time scale are presented, highlighting the potential of this approach to capture novel, data-driven scientific insights.

To obtain the surrogate, the ML models were trained on a data set simulating a coupled hydrological-vegetation-soil system. The rate of change of the two variables describing the system, soil depth and biomass, was used as output, taking their value at the previous time step and the pre-defined grazing pressure as inputs. Two popular ML methods, random forest (RF) and fully connected neural network (NN), were used. As proof of concept and to configure the model setup, we first trained the ML models on the output of the minimal model described in [1], comparing the ML responses at gridded inputs with the derivative values predicted by the minimal model. While RF required less tuning to achieve competitive results, a relative root mean squared error (rRMSE) of 5.8% and 0.04% for biomass and soil depth respectively, NN produced better-behaved outcome, reaching a rRMSE of 2.2% and 0.01%. Using the same setup, the ML surrogates were trained on a high-resolution numerical model describing the same system. The study of the response from this surrogate provided a more accurate description of the dynamics and equilibria of the hillslope ecosystem, depicting, for example, a much more complex process of hillslope desertification than captured by the minimal model.

It is thus concluded that the use of ML models instead of expert-based minimal models may lead to considerably different findings, where ML models have the advantage that they directly rely on system functioning embedded in their parent numerical simulation model.

How to cite: Pomarol Moya, O., Karssenberg, D., Immerzeel, W., Nussbaum, M., and Mehrkanoon, S.: Understanding geoscientific system behaviour from machine learning surrogates, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-11880, https://doi.org/10.5194/egusphere-egu24-11880, 2024.

While forecasting of climate and earth system processes has long been a task for numerical models, the rapid development of deep learning applications has recently brought forth competitive AI systems for weather prediction. Earth system models (ESMs), even though being an integral part of numerical weather prediction have not yet caught that same attention. ESMs forecast water, carbon and energy fluxes and in the coupling with an atmospheric model, provide boundary and initial conditions. We set up a comparison of different deep learning approaches for improving short-term forecasts of land surface and ecosystem states on a regional scale. Using simulations from the numerical model and combining them with observations, we will partially emulate an existing land surface scheme, conduct a probabilistic forecasts of core ecosystem processes and determine forecast horizons for all variables.

How to cite: Wesselkamp, M., Chantry, M., Kalweit, M., Pinnington, E., Choulga, M., Boedecker, J., Dormann, C., Pappenberger, F., and Balsamo, G.: Partial land surface emulator forecasts ecosystem states at verified horizons, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20863, https://doi.org/10.5194/egusphere-egu24-20863, 2024.

How to cite: Prasad, A., Merkouriadi, I., and Nummelin, A.: Exploring data-driven emulators for snow on sea ice , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-14957, https://doi.org/10.5194/egusphere-egu24-14957, 2024.

Urban warming in cities is increasingly exacerbated by the escalation of more frequent and severe heat extremes. Effectively mitigating overheating necessitates the adoption of a comprehensive, whole-system approach that integrates various heat mitigation measures to generate rapid and sustained efficacy in mitigation efforts. However, there remains a significant gap in the exploration of how to quantify the efficacy of mitigation strategies at the city-scale.

We address this research question by leveraging mesoscale Weather Research Forecasting (WRF) models alongside machine-learning (ML) techniques. As a showcase, ML models have been established for Zurich and Basel, Switzerland, utilizing seven WRF-output-based features, including shortwave downward radiation (SWDNB), hour of the day (HOUR), zenith angle (COSZEN), rain mix ratio (QRAIN), longwave downward radiation (LWDNB), canopy water content (CANWAT), and planetary boundary layer height (PBLH). Impressively, the resultant median R² values for T2 (2m temperature) predictions during heatwave and non-heatwave periods are notably high at 0.94 and 0.91 respectively.

Within the perspective of the whole-system approach, we quantify the impacts of reducing shortwave radiation absorption at ground surfaces, a potential result of a combination of both shading and reflective coating-based mitigation measures, through the utilization of ML models. Remarkably, a 5% reduction in the absorption of radiation at ground surfaces in Zurich could lead to a reduction in T2 by as much as 3.5 °C in the city center. During a heatwave in Basel, the potential for cooling is even more pronounced, with temperature decreases of up to 5 °C. These case studies in Zurich and Basel underscore the efficacy of utilizing WRF feature-trained ML models to quantify heat mitigation strategies at the city-scale.

How to cite: Zhao, Y., Wang, Z., Strebel, D., and Carmeliet, J.: Blending machine-learning and mesoscale numerical weather prediction models to quantify city-scale heat mitigation, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7581, https://doi.org/10.5194/egusphere-egu24-7581, 2024.

Non-linear, dynamic patterns are the rule rather than the exception in ecosystems. Predicting such patterns would allow an improved understanding of energy and nutrient flows in such systems. The Scientific Machine Learning approach Universal Differential Equation (UDE) by Rackauckas et al. (2020) tries to extract the underlying dynamical relations of state variables directly from their time series in combination with some knowledge on the dynamics of the system. This approach makes this kind of tool a promising approach to support classical modeling when precise knowledge of dynamical relationships is lacking, but measurement data of the phenomenon to be modeled is available.

We applied the UDE approach to a 22-year data set of the southern Baltic Sea coast, which constituted six different phytoplankton bloom types. The data set contained the state variables chlorophyll and different dissolved and total nutrients. We learned the chlorophyll:nutrient interactions from the data with additional forcing of external temperature, salinity and light attenuation dynamics as drivers. We used a neural network as a universal function approximator that provided time series of the state variables and their derivatives.

Finally, we recovered algebraic relationships between the variables chlorophyll, dissolved and total nutrients and the external drivers temperature, salinity and light attenuation using Sparse Identification of nonlinear Dynamics (SinDy) by Brunton et al. (2016).

The gained algebraic relationships differed in their importance of the different state variables and drivers for the six phytoplankton bloom types in accordance with general mechanisms reported in literature for the southern Baltic Sea coast. Our approach may be a viable option to guide ecosystem management decisions based on those algebraic relationships.

Rackauckas et al. (2020), arXiv preprint arXiv:2001.04385.

Brunton et al. (2016), PNAS 113.15: 3932-3937.

How to cite: Nieters, P., Berthold, M., and Vortmeyer-Kley, R.: Learning phytoplankton bloom patterns - A long and rocky road from data to equations , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19352, https://doi.org/10.5194/egusphere-egu24-19352, 2024.

Thanks to the recent progress in numerical methods, the application fields of artificial intelligence (AI) and machine learning methods (ML) are growing at a very fast pace. The EURAD (European Joint Programme on Radioactive Waste Management) community has recently started using ML for a) acceleration of numerical simulations, b) improvement of multiscale and multiphysics couplings efficiency, c) uncertainty quantification and sensitivity analysis. A number of case studies indicate that use of ML based approaches leads to overall acceleration of geochemical and reactive transport simulations from one to four orders of magnitude. The achieved speed-up depends on the chemical system, simulation code, problem formulation and the research question to be answered. Within EURAD-DONUT (Development and Improvement Of Numerical methods and Tools for modelling coupled processes), a benchmark is on-going to coordinate the relevant activities and to test a variety of ML techniques for geochemistry and reactive transport simulations in the framework of radioactive waste disposal. It aims at benchmarking several widely used geochemical codes, at generating high-quality geochemical data for training/validation of existing/new methodologies, and at providing basic guidelines about the benefits, drawbacks, and current limitations of using ML techniques.

A joint effort has resulted in the definition of benchmarks of which one is presented here. The benchmark system is relevant to the sorption of U in claystone formations (e.g. Callovo-Oxfordian, Opalinus or Boom clay). Regarding the chemical complexity, a system containing Na-Cl-U-H-O is considered as the base case, and a more complex system with the addition of calcium and carbonate (CO2) to change aqueous speciation of U. Parameters of interest, among others, are the resulting concentrations of U sorbed on edges (surface complexes), of U on ion exchange sites, and the amount of metaSchoepite, with the resulting Kd’s. Following aspects are discussed: (i) Streamline the production of high-quality consistent training datasets, using the most popular geochemical solvers (PHREEQC, ORCHESTRA and GEMS). (ii) The use of different methods (e.g. Deep Neural Networks, Polynomial Chaos Expansion, Gaussian Processes, Active Learning, and other techniques to learn from the generated data. (iii) Setup appropriate metrics for the critical evaluation of the accuracy of ML models. (iv) Testing the accuracy of predictions for geochemical and reactive transport calculations. 

How to cite: Laloy, E. and Montoya, V. and the EURAD-DONUT Team: Machine learning based metamodels for geochemical calculations in reactive transport models: Benchmark within the EURAD Joint Project, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-21545, https://doi.org/10.5194/egusphere-egu24-21545, 2024.

Global gridded crop models (GGCMs) have become state-of-the-art tools in large-scale climate impact and adaptation assessments. Yet, these combinations of large-scale spatial data frameworks and plant growth models have limitations in the volume of scenarios they can address due to computational demand and complex software structures. Emulators mimicking such models have therefore become an attractive option to produce reasonable predictions of GGCMs’ crop productivity estimates at much lower computational costs. However, such emulators’ flexibility is thus far typically limited in terms of crop management flexibility and spatial resolutions among others. Here we present a new emulator pipeline CROp model Machine learning Emulator Suite (CROMES) that serves for processing climate features from netCDF input files, combining these with site-specific features (soil, topography), and crop management specifications (planting dates, cultivars, irrigation) to train machine learning emulators and subsequently produce predictions. Presently built around the GGCM EPIC-IIASA and employing a boosting algorithm, CROMES is capable of producing predictions for EPIC-IIASA’s crop yield estimates with high accuracy and very high computational efficiency. Predictions require for a first used climate dataset about 45 min and 10 min for any subsequent scenario based on the same climate forcing in a single thread compared to approx. 14h for a GGCM simulation on the same system.

Prediction accuracy is highest if modeling the case when crops receive sufficient nutrients and are consequently most sensitive to climate. When training an emulator on crop model simulations for rainfed maize and a single global climate model (GCM), the yield prediction accuracy for out-of-bag GCMs is R2=0.93-0.97, RMSE=0.5-0.7, and rRMSE=8-10% in space and time. Globally, the best agreement between predictions and crop model simulations occurs in (sub-)tropical regions, the poorest is in cold, arid climates where both growing season length and water availability limit crop growth. The performance slightly deteriorates if fertilizer supply is considered, more so at low levels of nutrient inputs than at the higher end.

Importantly, emulators produced by CROMES are virtually scale-free as all training samples, i.e., pixels, are pooled and hence treated as individual locations solely based on features provided without geo-referencing. This allows for applications on increasingly available high-resolution climate datasets or in regional studies for which more granular data may be available than at global scales. Using climate features based on crop growing seasons and cardinal growth stages enables also adaptation studies including growing season and cultivar shifts. We expect CROMES to facilitate explorations of comprehensive climate projection ensembles, studies of dynamic climate adaptation scenarios, and cross-scale impact and adaptation assessments.

How to cite: Folberth, C., Baklanov, A., Khabarov, N., Oberleitner, T., Balkovic, J., and Skalsky, R.: CROMES - A fast and efficient machine learning emulator pipeline for gridded crop models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5852, https://doi.org/10.5194/egusphere-egu24-5852, 2024.

Deep learning models emerged as viable alternatives to rapid and accurate flood mapping, overcoming the computational burden of numerical methods. In particular, hydraulic-based graph neural networks present a promising avenue, offering enhanced transferability to domains not used for the model training. These models exploit the analogy between finite-volume methods and graph neural networks to describe how water moves in space and time across neighbouring cells. However, existing models face limitations, having been exclusively tested on regular meshes and necessitating initial conditions from numerical solvers. This study proposes an extension of hydraulic-based graph neural networks to accommodate time-varying boundary conditions, showcasing its efficacy on irregular meshes. For this, we employ multi-scale methods that jointly model the flood at different scales. To remove the necessity of initial conditions, we leverage ghost cells that enforce the solutions at the boundaries. Our approach is validated on a dataset featuring irregular meshes, diverse topographies, and varying input hydrograph discharges. Results highlight the model's capacity to replicate flood dynamics across unseen scenarios, without any input from the numerical model, emphasizing its potential for realistic case studies.

How to cite: Bentivoglio, R., Isufi, E., Jonkman, S. N., and Taormina, R.: Multi-scale hydraulic-based graph neural networks: generalizing spatial flood mapping to irregular meshes and time-varying boundary condition, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7681, https://doi.org/10.5194/egusphere-egu24-7681, 2024.

Seismology has witnessed significant advancements in recent years with the application of deep
learning methods to address a broad range of problems. These techniques have demonstrated their
remarkable ability to effectively extract statistical properties from extensive datasets, surpassing the
capabilities of traditional approaches to an extent. In this study, we present SAIPy, an open-source
Python package specifically developed for fast data processing by implementing deep learning.
SAIPy offers solutions for multiple seismological tasks, including earthquake detection, magnitude
estimation, seismic phase picking, and polarity identification. We introduce upgraded versions
of previously published models such as CREIME_RT capable of identifying earthquakes with an
accuracy above 99.8% and a root mean squared error of 0.38 unit in magnitude estimation. These
upgraded models outperform state-of-the-art approaches like the Vision Transformer network. SAIPy
provides an API that simplifies the integration of these advanced models, including CREIME_RT,
DynaPicker_v2, and PolarCAP, along with benchmark datasets. The package has the potential to be
used for real-time earthquake monitoring to enable timely actions to mitigate the impact of seismic
events.

How to cite: Srivastava, N., Li, W., Chakraborty, M., Cartaya, C. Q., Köhler, J., Faber, J., and Rümpker, G.: SAIPy: A Python Package for single station Earthquake Monitoring using Deep Learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19502, https://doi.org/10.5194/egusphere-egu24-19502, 2024.

Estimating the seismic hazard in earthquake-prone regions, in order to assess the risk associated to nuclear facilities, must take into account a large number of uncertainties, and in particular our limited knowledge of the geology. And yet, we know that certain geological features can create site effects that considerably amplify earthquake ground motion. In this work, we provide a quantitative assessment of how largely can earthquake ground motion simulation benefit from deep learning approaches, quantifying the influence of geological heterogeneities on the spatio-temporal nature of the earthquake-induced site response. Two main frameworks are addressed: conditional generative approaches with diffusion models and neural operators. On one hand, generative adversarial learning and diffusions models are compared in a time-series super-resolution context [1]. The main task is to improve the outcome of 3D fault-to-site earthquake numerical simulations (accurate up to 5 Hz [2, 3]) at higher frequencies (5-30 Hz), by learning the low-to-high frequency mapping from seismograms recorded worldwide [1]. The generation is conditioned by the numerical simulation synthetic time-histories, in a one-to-many setup that enables site-specific probabilistic hazard assessment. On the other hand, the successful use of Factorized Fourier Neural Operator (F-FNO) to entirely replace cumbersome 3D elastodynamic numerical simulations is described [4], showing how this approach can pave the way to real-time large-scale digital twins of earthquake prone regions. The trained neural operator learns the relationship between 3D heterogeneous geologies and surface ground motions generated by the propagation of seismic wave through these geologies. The F-FNO is trained on the HEMEW-3D (https://github.com/lehmannfa/HEMEW3D/releases) database, comprising 30000 high-fidelity numerical simulations of earthquake ground motion through generic geologies, performed by employing the high-performance code SEM3D [4]. Next, a smaller database was built specifically for the Teil region (Ardèche, France), where a MW 4.9 moderate shallow earthquake occurred in November 2019 [4]. The F-FNO is then specialized on this database database with just 250 examples. Transfer learning improved the prediction error by 22 %. According to seismological Goodness-of-Fit (GoF) metrics, 91% of predictions have an excellent GoF for the phase (and 62% for the envelope). Ground motion intensity measurements are, on average, slightly underestimated.

[1] Gatti, F.; Clouteau, D. Towards Blending Physics-Based Numerical Simulations and Seismic Databases Using Generative Adversarial Network. Computer Methods in Applied Mechanics and Engineering 2020, 372, 113421.
https://doi.org/10.1016/j.cma.2020.113421.

[2] Touhami, S.; Gatti, F.; Lopez-Caballero, F.; Cottereau, R.; de Abreu Corrêa, L.;Aubry, L.; Clouteau, D. SEM3D: A 3D High-Fidelity Numerical Earthquake Sim-ulator for Broadband (0–10 Hz) Seismic Response Prediction at a Regional Scale.Geosciences 2022, 12 (3), 112. https://doi.org/10.3390/geosciences12030112. https://github.com/sem3d/SEM

[3] Gatti, F.; Carvalho Paludo, L. D.; Svay, A.; Lopez-Caballero, F.-; Cottereau, R.;Clouteau, D. Investigation of the Earthquake Ground Motion Coherence in Het-erogeneous Non-Linear Soil Deposits. Procedia Engineering 2017, 199, 2354–2359.https://doi.org/10.1016/j.proeng.2017.09.232.[4] Lehmann, F.; Gatti, F.; Bertin, M.; Clouteau, D. Machine Learning Opportunities to Conduct High-Fidelity Earthquake Simulations in Multi-Scale Heterogeneous Geology.Front. Earth Sci. 2022, 10, 1029160. https://doi.org/10.3389/feart.2022.1029160.

[4] Lehmann, F.; Gatti, F.; Bertin, M.; Clouteau, D. Fourier Neural Operator Sur-rogate Model to Predict 3D Seismic Waves Propagation. arXiv April 20, 2023.http://arxiv.org/abs/2304.10242 (accessed 2023-04-21).

How to cite: Gatti, F., Lehmann, F., Gabrielidis, H., Bertin, M., Clouteau, D., and Vialle, S.: Deep learning generative strategies to enhance 3D physics-based seismic wave propagation: from diffusive super-resolution to 3D Fourier Neural Operators., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2443, https://doi.org/10.5194/egusphere-egu24-2443, 2024.

The Machine Learning Estimator for Ground Shaking Maps (MLESmaps) harnesses the ground shaking inference capability of Machine Learning (ML) models trained on physics-informed earthquake simulations. It infers intensity measures, such as RotD50, seconds after a significant earthquake has occurred given its magnitude and location. 

Our methodology incorporates both offline and online phases in a comprehensive workflow. It begins with the generation of a synthetic training data set, progresses through the extraction of predictor characteristics, proceeds to the validation and learning stages, and yields a learned inference model. 

MLESmap results can complement empirical Ground Motion Models (GMMs), in particular in data-poor areas, to assess post-earthquake hazards rapidly and accurately, potentially improving disaster response in earthquake-prone regions. Learned models incorporate physical features such as directivity, topography, or resonance at a speed comparable to that of the empirical GMMs. 

In this work, we present an overview of the MLESmap methodology and its application to two distinct study areas: southern California and southern Iceland

How to cite: Monterrubio-Velasco, M., Blanco, R., Callaghan, S., Bhihe, C., Pienkowska, M., Ejarque, J., and de la Puente, J.: Machine Learning Estimator for Ground-Shaking maps, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15914, https://doi.org/10.5194/egusphere-egu24-15914, 2024.

Physics-based simulations of earthquake ground motions prove invaluable, particularly in regions where strong ground motion recordings remain scarce. However, the computational demands associated with these simulations limit their applicability in tasks that necessitate large-scale computations of a wide range of possible earthquake scenarios, such as those required in physics-based probabilistic seismic hazard analyses. To address this challenge, we propose a neural-network approach that enables the rapid computation of earthquake ground motions in the spectral domain, alleviating a significant portion of the computational burden. To illustrate our approach, we generate a database of ground motion simulations in the San Francisco Bay Area using AxiSEM3D, a 3D seismic wave simulator. The database includes 30 double-couple sources with varying depths and horizontal locations. Our simulations explicitly incorporate the effects of topography and viscoelastic attenuation and are accurate up to frequencies of 0.5 Hz. Preliminary results demonstrate that the trained neural network almost instantaneously produces estimates of peak ground displacements as well as displacement waveforms in the spectral domain that align closely with those obtained from the wave propagation simulations. Our approach also extends to predicting ground motions for ‘unsimulated’ source locations, ultimately providing a comprehensive resolution of the source space in our chosen physical domain. This advancement paves the way for a cost-effective simulation of numerous seismic sources, and enhances the feasibility of physics-based probabilistic seismic hazard analyses. 

How to cite: Ramadan, F., Fry, B., and Nissen-Meyer, T.: Rapid Computation of Physics-Based Ground Motions in the Spectral Domain using Neural Networks, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18444, https://doi.org/10.5194/egusphere-egu24-18444, 2024.

How to cite: Montagna, C. P., Garg, D., Allegra, M., Cannavò, F., Currenti, G., Bruni, R., and Papale, P.: A digital twin for volcanic deformation merging 3D numerical simulations and AI, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19255, https://doi.org/10.5194/egusphere-egu24-19255, 2024.

Simulating nitrate fate and transport in freshwater is an essential part in water quality management. Numerical and data-driven models have been used for it. The numerical model SWAT simulates daily nitrate loads using simulated flow rate. Data-driven models are more flexible compared to SWAT as they can simulate nitrate load and flow rate independently. The objective of this work was evaluating the performance of SWAT and a deep learning model in terms of nutrient loads in cases when deep learning model is used in (a) simulating flow rate and nitrate concentration independently and (b) simulating both flow rate and nitrate concentration. The deep learning model was built using long-short-term-memory and three-dimensional convolutional networks. The input data, weather data and image data including leaf area index and land use, were acquired at the Tuckahoe Creek watershed in Maryland, United States. The SWAT model was calibrated with data over the training period (2014-2017) and validated with data over the testing period (2019) to simulate flow rate and nitrate load. The Nash-Sutcliffe efficiency was 0.31 and 0.40 for flow rate and -0.26 and -0.18 for the nitrate load over training and testing periods, respectively. Three data-driven modeling scenarios were generated for nitrate load. Scenario 1 included the flow rate observation and nitrate concentration simulation, scenario 2 included the flow rate simulation and nitrate concentration observation, and scenario 3 included the flow rate and nitrate concentration simulations. The deep learning model outperformed SWAT in all three scenarios with NSE from 0.49 to 0.58 over the training period and from 0.28 to 0.80 over the testing period. Scenario 1 showed the best results for nitrate load. The performance difference between SWAT and the deep learning model was most noticeable in fall and winter seasons. The deep learning modeling can be an efficient alternative to numerical watershed-scale models when the regular high frequency data collection is provided.

How to cite: Lee, J., Kim, D., Hong, S., Yun, D., Kwon, D., Hill, R., Pachepsky, Y., Gao, F., Zhang, X., Lee, S., and Cho, K.: Comparison of SWAT and a deep learning model in nitrate load simulation at the Tuckahoe creek watershed in the United States, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6622, https://doi.org/10.5194/egusphere-egu24-6622, 2024.