ITS1.12/NH0.2 | Hybrid modeling of natural hazards: blending deep-learning, data-driven approaches and physics-based simulations
Hybrid modeling of natural hazards: blending deep-learning, data-driven approaches and physics-based simulations
Convener: Filippo GattiECSECS | Co-convener: Nishtha SrivastavaECSECS
| Mon, 24 Apr, 16:15–18:00 (CEST)
Room 0.94/95
Posters on site
| Attendance Mon, 24 Apr, 14:00–15:45 (CEST)
Hall X4
Orals |
Mon, 16:15
Mon, 14:00
Recent advances in the field of Artificial Intelligence, Machine Learning and Data Assimilation have been massively applied to model, anticipate, and predict the natural catastrophes, such as earthquakes, floods, landslides, volcanic eruptions, tsunamis, wildfires, glacier instabilities, in addition to multi-hazard and cascading effects due to climate change. However, the adopted algorithms require a solid inductive bias, provided by the physics of the phenomenon at stake (or at least the understanding of it). Furthermore, due to simplified assumptions, analytical models might encounter limits while modeling these natural catastrophes. Therefore, several hybrid strategies, utilizing the growing computational resources, are currently being developed, to achieve more flexibility and full synergy between numerical physics-based simulations, machine learning and data-driven approaches.
The hybrid modeling of natural hazards benefits from the interpretability of numerical simulations and from the extrapolation and generalization capabilities of advanced Machine Learning methods. This synergy leads to multi-fidelity predictive tools that leverage all the available knowledge on the phenomenon at stake. Moreover, to tackle lack of data and representation, observational databases can be integrated with the synthetic results for re-analysis and for training machine learning algorithms on never-before-seen disaster scenarios. This multidisciplinary session invites contributions addressing hybrid solutions to predict and to mitigate natural catastrophes. It also welcomes presentations on hybrid tools for vulnerability assessment.

Orals: Mon, 24 Apr | Room 0.94/95

Chairpersons: Filippo Gatti, Nishtha Srivastava
On-site presentation
Jake Lever, Sibo Cheng, and Rossella Arcucci

Twitter is increasingly being used as a real-time human-sensor network during natural disasters, detecting, tracking and documenting events. Current wildfire models currently largely omit social media data, representing a shortcoming in current models, as valuable and timely information is transmitted via this channel. By including this data as a real-time data source, we aim to help disaster managers make more informed, socially driven decisions, by detecting and monitoring online social media sentiment over the course of a wildfire event. This monitoring model is coupled to a real-time forecasting of wildfire dynamics.

Real-time forecasting of wildfire dynamics, which has attracted increasing attention recently in fire safety science, is extremely challenging due to the complexities of the physical models and the geographical features. Running physics-based simulations for large-scale wildfires can be computationally difficult. We propose a novel algorithm scheme, which combines reduced-order modelling (ROM), recurrent neural networks (RNN), data assimilation (DA) and error covariance tuning for real-time forecasting/monitoring of the burned area. An operating cellular automata (CA) simulator is used to compute a data-driven surrogate model for forecasting fire diffusions. A long-short-term-memory (LSTM) neural network is used to build sequence-to-sequence predictions following the simulation results projected/encoded in a reduced-order latent space. 

We implement machine learning in a wildfire prediction model, using social media and geophysical data sources with sentiment analysis to predict wildfire instances and characteristics with high accuracy. The geophysical data is satellite data provided by the Global Fire Atlas, and social data is provided by Twitter. In doing this, we perform our own data collection and analysis, comparing regional differences in online social sentiment expression.

The performance of the proposed algorithm has been tested in recent massive wildfire events in California.

How to cite: Lever, J., Cheng, S., and Arcucci, R.: Social & Physics Based Data Driven Methods for Wildfire Prediction, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15645,, 2023.

On-site presentation
Federica Gaspari, Federico Barbieri, Francesco Ioli, Livio Pinto, and Paolo Valgoi

The fragile geomorphological context of Italy sets a variety of natural challenges, ranging from seismic to hydrogeological risk. In such a complex territory, documenting the conditions of infrastructures is crucial for planning adequate strategies of maintenance through 3D modelling for structural analysis and digital twins’ implementation of structures like dams (Pagliari et al., 2016) or bridges (Gaspari et al., 2022). Geomatics, through periodical surveys using state-of-the-art technologies, reconstruct accurate 3D models of structures that results in the generation of dense pointclouds from which polygon meshes can be derived as well as in the model integration in Building Information Modeling (BIM) or Finite Element Method (FEM) environments for the computation of simulations and deformation monitoring or structural health assessment analysis in support of decision making.

Such data are generated through different approaches. A traditional methodology first implies the materialization and measurement of a topographic network in a local system with a total station and its subsequent georeferencing in a global coordinate reference system through a roto-translation based on Global Navigation Satellite System observations of ground control points. In the same framework, scans for the acquisition of dense pointclouds are defined through the adoption of a terrestrial laser scanner (TLS). Hence, the execution of planned drone flights, with nadiral and side view of the structure and its surrounding environment, serving as input for the generation of photogrammetric cloud through a robust Structure from Motion data processing.

Implementing open-source WebGL solutions like Potree supports the digital twin and data sharing with audiences of different technical backgrounds, committers concerned with the adoption of a monitoring platform for integrating products in different format as well as experts with non-geomatics expertise interested in further analysis of collected data through computer vision and deep learning approches that enrich the existent documentation. With a user-friendly interactive web platforms users are able to access the 3D model, make measurements and execute simple processing operation like cross-sections and clipping (e.g.

Since 2019, the dams of the Sila mountains in the Calabria region represented the case study for testing the described integrated approach. The present work concerns the integration of data from different sensors (TLS for indoor and outdoor environment, photogrammetric images and lidar from drone) for the generation of the digital twin of the arcuate-plan gravity dam of Trepidò. The dam digital twin of the dam and adjacencies consists of a pointcloud of 2594370 points, with adaptive density and average accuracy of 1-2 cm for the structure and 10 cm for the downstream vegetated sediment. It can be used to increase knowledge of the structure (built in 1930) and for structural analysis.




Gaspari, F., Ioli, F., Barbieri, F., Belcore, E., and Pinto, L. (2022): Integration of UAV-LiDAR and UAV-photogrammetry for infrastructure monitoring and bridge assessment, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2022, 995–1002,

Pagliari, D., Rossi, L., Passoni, D., Pinto, L., de Michele, C., and Avanzi, F. (2016). Measuring the volume of flushed sediments in a reservoir using multi-temporal images acquired with UAS, Geomatics, Natural Hazards and Risk, 8(1), 150–166,

How to cite: Gaspari, F., Barbieri, F., Ioli, F., Pinto, L., and Valgoi, P.: Integration of 3D surveying approaches for critical infrastructure digital twins in natural hazard-prone scenarios, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7773,, 2023.

On-site presentation
Azzeddine Soulaïmani, Azzedine Abdedou, Yash Kumar, and Pratyush Bhatt

Most science and engineering problems are modeled by time-dependent and parametrized nonlinear partial differential equations. Their resolution with traditional computational methods may be too expensive, especially in the context of predictions with uncertainty quantification or optimization, to allow for rapid predictions.  In this talk, we will overview data-driven methods aimed at representing high-fidelity computational models by means of reduced-dimension surrogate ones.  Different approaches will be presented for the uncertainty quantification for reliable predictions and forecasts in inundation problems.

Particularly, a non-intrusive reduced-order model based on convolutional autoencoders is proposed as a data-driven tool to build an efficient nonlinear reduced-order model for stochastic spatiotemporal large-scale physical problems. The method uses two-level autoencoders to reduce the spatial and temporal dimensions from a set of high-fidelity snapshots collected from an in-house high-fidelity numerical solver of the shallow-water equations. The encoded latent vectors, generated from two compression levels, are then mapped to the input parameters using a regression-based multilayer perceptron. The accuracy of the proposed approach is compared to the linear reduced-order technique-based artificial neural network (POD-ANN) on benchmark tests (the Burgers and Stoker's solutions) and a hypothetical dam-break flow problem over a complex bathymetry river. The numerical results show that the proposed nonlinear framework presents strong predictive abilities to accurately approximate the statistical moments of the outputs for complex stochastic large-scale and time-dependent problems, with low computational cost during the predictive online stage.

The caveat that remains is the long-term temporal extrapolation for problems marked by sharp gradients and discontinuities. Our study explores forecasting convolutional architectures (LSTM, TCN, and CNN) to obtain accurate solutions for time-steps distant from the training domain, on advection-dominated test cases. A simple convolutional architecture is then proposed and shown to provide accurate results for the forecasts. To evaluate the epistemic uncertainties in the solutions, the methodology of deep ensembles is adopted.


  • Bhatt, Y. Kumar and A. Soulaïmani. Deep Convolutional Architectures for Extrapolative Forecast in Time-dependent Flow Problems, DOI: 10.48550/arXiv.2209.09651.
  • Abdedou and A. Soulaïmani. Reduced-order modeling for stochastic large-scale and time-dependent problems using deep spatial and temporal convolutional autoencoders.
  • Jacquier, A. Abdedou, V. Delmas and A. Soulaimani. Non-intrusive reduced-order modeling using uncertainty-aware Deep Neural Networks and Proper Orthogonal Decomposition: Application to flood modeling. Journal of Computational Physics. Volume 424, 1 January 2021, 109854.
  • Abdedou and A. Soulaïmani. A non-intrusive reduced-order modeling for uncertainty propagation of time-dependent problems using a B-splines Bézier elements-based method and proper orthogonal decomposition: Application to dam-break flows. Computers & Mathematics with Applications. Volume 102, 15 November 2021, Pages 187-205.
  • Chaudhry and A. Soulaimani. A Comparative Study of Machine Learning Methods for Computational Modeling of the Selective Laser Melting Additive Manufacturing Process. Appl. Sci. 2022, 12(5), 2324;
  • Delmas and A. Soulaimani. Parallel high-order resolution of the Shallow-water equations on real large-scale meshes with complex bathymetries. Journal of Computational Physics. Volume 471, 15 December 2022, 111629


How to cite: Soulaïmani, A., Abdedou, A., Kumar, Y., and Bhatt, P.: Deep Convolutional Architectures for Uncertainty Quantification and Forecast in Inundation Problems, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-39,, 2023.

On-site presentation
Santosh Kumar Sasanapuri, Dhanya Chadrika Thulaseedharan, and Gosain Ashwini Kumar

Floods are one of the most devastating natural disasters in the world causing loss of human lives and property across the world. These losses can be minimized by accurate prediction of floods well in advance. However, 2D hydrodynamic models which are used for flood inundation modelling require high computational time and hence are unsuitable for development of real-time flood monitoring system in most cases. Therefore, a surrogate machine learning model named XGBoost Regressor (XBGR) is developed for flood inundation modelling. The developed model overcomes the constraint of high computational time required by 2D hydrodynamic models. The XGBR is developed to predict maximum flood depth map and is evaluated with the LISFLOOD-FP hydrodynamic model. The training data for the XGBR model is generated using the LISFLOOD-FP model. The surrogate model is trained on 21 flood events, tested on 4 and validated for 1 flood event. For better development of the surrogate model, physical characteristics of the study area are considered in the form of nine indices referred here as topographic variables along with the flood characteristic variables. However, to refrain the XGBR model from overfitting and decrease the training time, a feed forward feature selection method is used to select the best predictive topographic variables. Four topographic variables are selected after which there is no significant improvement in the model was found. Number of trees and learning rate parameters of XGBR model are parameterized which are having highest impact on the model performance. Mean absolute error (MAE) and root mean square error (RMSE) are used for evaluating model accuracy. For testing period, the average MAE and RMSE are 0.433 m and 0.780 m, respectively and for the validation event MAE and RMSE are 0.595 m and 0.960 m respectively. For evaluating the accuracy of the surrogate model on flood inundation extent, F1 score is used which is the harmonic mean of precision and recall. The F1 score is 0.908 for the testing events and is 0.931 for validation events. The higher value of F1 score (>0.9) indicates good accuracy of the XGBR model when validated using the hydrodynamic model.

How to cite: Sasanapuri, S. K., Chadrika Thulaseedharan, D., and Ashwini Kumar, G.: Using machine learning to emulate the hydrodynamic model for flood inundation modelling, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-554,, 2023.

On-site presentation
Fanny Lehmann, Filippo Gatti, Michaël Bertin, and Didier Clouteau

Physics-based deep learning experienced a major breakthrough a few years ago with the advent of neural operators. Beyond the traditional use of deep neural networks to predict the solution to a fixed Partial Differential Equation (PDE), these novel methods are able to learn the operator solution to a class of PDEs.

Comparisons and analyses of popular neural operators such as Fourier Neural Operator and DeepONet have been conducted for numerical case studies. However, they are still lacking for more realistic problems in complex settings.

In this study, we compare several neural operators to predict the propagation of seismic waves in heterogeneous media. Our database is composed of more than 12 million ground motion timeseries generated from 50,000 media. We quantify the accuracy of the neural operators, their memory requirements, and their dependence towards both the initial condition and the PDE parameters. We also propose insights on their possible extension to 3 dimensions.

How to cite: Lehmann, F., Gatti, F., Bertin, M., and Clouteau, D.: Advantages and promises of deep neural operators for the prediction of wave propagation, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5068,, 2023.

On-site presentation
Fatemeh Heidari, Qing Lin, Edgar Fabián Espitia Sarmiento, Andrea Toreti, and Elena Xoplaki

Early warning systems protect and support lives, jobs, land and infrastructure. DAKI-FWS, a German national project, aims at developing an early warning system to protect the German society and economy against extreme weather and climate events such as floods, droughts and heatwaves. With a seasonal temporal horizon, DAKI-FWS requires high resolution and bias corrected seasonal forecast of daily minimum and maximum temperatures, daily precipitation and wind speed. To derive such information, we have developed a deep neural network (DNN) approach to downscale and bias correct coarse resolution seasonal forecast ensembles on a 1 degree grid to a 1 arc minute grid.

The proposed DNN approach is here analyzed and compared with other machine learning approaches. Results show that such a deep learning technique can generate realistic, temporally consistent, and high-resolution climate information. The statistical and physical properties of the generated ensembles are analyzed using spatial correlation, cross validation and SVD. The DNN predicts extreme values that are very close to the observed values while preserving the physical relationships in the system as well as the trends in the variables.

How to cite: Heidari, F., Lin, Q., Espitia Sarmiento, E. F., Toreti, A., and Xoplaki, E.: Towards the development of an AI-based early warning system: a deep learning approach to bias correct and downscale seasonal climate forecasts, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9555,, 2023.

On-site presentation
Valeria Soto and Fernando Lopez-Caballero

The present paper focuses on the influence of Rayleigh and Love waves on the seismic structural performance of a simplified nonlinear beam structure representing a bridge column. The impact of surface waves in the structure is quantified directly by a coupled 3D SEM-FEM numerical wave
propagation simulation from the earthquake source to the structure using the Domain Reduction Method.
In the first step, ground motions, including basin-induced surface waves, are generated from a regional model containing the earthquake source and a simplified basin. Surface waves are extracted and characterized with the Normalized Inner Product (NIP) in terms of amplitude and
frequency content from ground motions at different locations inside the basin. In the second step, the seismic wavefield from the SE simulation is imposed in a FE model composed of a nonlinear structure placed over a portion of the basin sediments. The model considers soil-structure
interaction and structural non-linearity through a multifiber beam approach.
By placing the structure in different positions, the extracted surface waves and the structural damage can be linked to a specific location inside the basin. Therefore, the spatial variability of the structural damage and the surface wave characteristics can be quantified. Consequently, this work
evaluates if structural damage can be estimated only from typical ground motion intensity parameters or if other parameters associated with surface wave characteristics are necessary. The results show a correlation between obtained seismic damage with rotational components from
surface waves (torsion for Love waves and rocking for Rayleigh waves).

How to cite: Soto, V. and Lopez-Caballero, F.: Quantification of source- and basin-induced surface waves effects on the seismic performance of nonlinear structures, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11611,, 2023.

Virtual presentation
Naveen Ragu Ramalingam, Alice Abbate, Erlend Briseid Storrøsten, Kendra Johnson, Gareth Davies, Stefano Lorito, Marco Pagani, and Mario Martina

The hybrid modelling approach combining machine learning and physics-based simulation has been used in a variety of ways to study tsunami and improve our understanding of this complex natural hazard. They are broadly applied for (1) Tsunami forecasting and early warning systems and (2) Tsunami hazard and risk assessment including sensitivity, analysis uncertainty studies and inverse modelling for estimating the source. 

Rigorous evaluation of such a hybrid approach is constrained by the limited size of available simulation datasets which is important to guide their usage by practitioners. This study investigates the application of a hybrid tsunami modelling technique (Ragu Ramalingam et al., 2022, Ragu Ramalingam et al., 2022) which offers a computationally efficient approach for hazard assessment where large events-sets must be modelled typical of probabilistic tsunami hazard and risk assessment (PTHA/PTRA). We use a large tsunami simulation dataset for a coastal region of eastern Sicily, Italy and try to address the following question:

  • How to efficiently sample scenarios used to train the ML models?
  • Where and when are such methods accurate? 
  • How do they compare with other traditional modelling methods like Monte Carlo Sampling?

Additionally, the effort will deliver an open tsunami benchmarking dataset that can be utilised for further development, baseline comparison of various ML algorithms, and improved hyperparameter tuning.


Ragu Ramalingam, N., Johnson, K., Pagani, M., and Martina, M.: A hybrid ML-physical modelling approach for efficient approximation of tsunami waves at the coast for probabilistic tsunami hazard assessment, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5642,, 2022.

Ragu Ramalingam, N., Rao, A., Johnson, K., Pagani, M. and Martina, M. A hybrid ML-physical modelling approach for efficient probabilistic tsunami hazard and risk assessment, Proceedings of the 19th Annual Meeting of the Asia Oceania Geosciences Society (AOGS 2022), August 1-5, 2022, Virtual.

How to cite: Ragu Ramalingam, N., Abbate, A., Briseid Storrøsten, E., Johnson, K., Davies, G., Lorito, S., Pagani, M., and Martina, M.: Efficient Probabilistic Tsunami Hazard and Risk Assessment Using a Hybrid Modeling Approach: A Systematic Evaluation, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-16906,, 2023.

On-site presentation
Jonas Köhler, Wei Li, Johannes Faber, Georg Rümpker, Horst Stöcker, and Nishtha Srivastava

Usually, the earthquake catalog for a given region represents a collection of all detected and localized earthquakes and, thus, contains not only the main shocks, but also fore- and aftershocks. In order to perform an independent seismic event and seismic hazard analysis we require a catalog that, ideally, contains only mainshocks. Thus, the removal of dependent fore- and aftershocks from an earthquake catalogby declustering is a crucial step in seismic hazard analysis. Machine learning methods can potentially offer improvements in speed and accuracy in comparison to classical declustering approaches.

Here, we propose a hybrid approach to identify the temporal clusters of earthquakes from the catalogs of California (USGS) and Japan (ISC). We combine unsupervised 1-D clustering algorithms with seismologically informed methods and machine learning techniques. We use epidemic type aftershock sequence (ETAS) generated catalogs as well as classically declustered catalogs to benchmark the method.

How to cite: Köhler, J., Li, W., Faber, J., Rümpker, G., Stöcker, H., and Srivastava, N.: A hybrid approach for declustering of earthquake catalogs, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11462,, 2023.

Posters on site: Mon, 24 Apr, 14:00–15:45 | Hall X4

Chairpersons: Nishtha Srivastava, Filippo Gatti
Simon Thomas, Dan(i) Jones, Talea Mayo, and Devaraj Gopinathan

Storm surges can have devastating effects on coastal communities. These events, often caused by tropical cyclones, are difficult to simulate due to the challenging nature of process-based modelling and the relative paucity of data covering extreme tropical cyclone conditions. In order to make optimal use of existing physical models, we build an emulator to actively learn the relationship between tropical cyclone characteristics and maximum storm surge height.


We used the ADCIRC physical storm surge model, a reliable but costly tool, to simulate a series of representative tropical cyclones that typically affect the coast near New Orleans. These initial storms were sampled using Latin hypercube design, varying tropical cyclone characteristics such as the landfall speed, central pressure, and others. By running the ADCIRC model for each of these events, we were able to determine the maximum sea surface height caused by each simulated storm. Next, we trained a Gaussian process to fit the maximum sea levels at each point along the coast given the tropical cyclones' characteristics as input. Through active learning, we iteratively selected additional tropical cyclones to further improve the emulator’s accuracy. Finally, we evaluated the model's performance using a held-out test set of idealised tropical cyclones.


Our emulator approach allowed us to efficiently create a high-quality, low-cost statistical model that can potentially be used to predict the probability of future storm surge heights. Additionally, it allowed us to separate uncertainties in the input distribution of tropical cyclone characteristics from uncertainties in the model itself. By better understanding these sources of uncertainties, we can work towards more accurately assessing the potential impacts of future storms on coastal communities.

How to cite: Thomas, S., Jones, D., Mayo, T., and Gopinathan, D.: Tropical cyclone storm surge emulation around New Orleans, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-13337,, 2023.

Tatsuya Ishikawa

Solving partial differential equations (PDEs) stably and accurately is essential in simulation analysis of a variety of geophysical phenomena. Designing appropriate discretization schemes for PDEs requires careful and rigorous mathematical treatment and has been a long-term research topic. The computational efficiency is additionally a long-standing challenge when what-if hazard scenario analysis is considered. The data-driven discretization is a hybrid approach to combine machine learning and physics-based simulations, which provides a methodology to derive better discretization schemes from reliable references obtained typically using known stable schemes with higher resolution grids. As the resultant schemes may inherit the physics described by the PDEs, surrogate models employing them are expected to be in good agreement with expensive simulations. It is also argued that the learnt schemes by neural network models can exhibit similar characteristics to known sophisticated algorithms and outperform them in terms of accuracy. However, the method has currently been assessed with only limited examples and the detailed mechanisms of the learnt schemes are not well understood. In this presentation, thorough assessment and investigation of learning discretization schemes are conducted by applying the methodology to several types of differential equations with different learning models for the schemes. Whether the methodology has the potential to derive new schemes is also discussed.

How to cite: Ishikawa, T.: On learning discretization schemes of partial differential equations in geoscience, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11657,, 2023.

Gottfried Jacquet, Didier Clouteau, and Filippo Gatti

In the last decades, geophysicists have developed numerical simulators to predict earthquakes and other natural catastrophes. However, the more precise the model is, the higher the computational burden and the time to results. In addition, even if we could reproduce the phenomenon with more complex and more representative models, the underlying uncertainty would remain significantly high, affecting the reliability of the final prediction. In response to this challenge, we adopted a hybrid strategy, consisting into mixing physics-based numerical simulations and machine-learning. The goal is to transform synthetic earthquake ground motion, obtained via physics-based simulation, accurate up to a frequency of 5 Hz, into a broader-band prediction that mimics the recorded seismographs. In doing so, we factorize the latent representation of the seismic signal, by forcing an encoding that splits features into two parts: a low frequency one (0-1 Hz) and a high frequency one (1-20 Hz). In the following, we train a convolutional U-Net neural network and apply two different signal-to-signal translation techniques: pix2pix and BiCycleGAN. The latter strategies are compared with the prior work of Gatti et al., 2020, on the Stanford Earthquake Dataset (STEAD) showing their capability of mimicking recorded seismographs. We finally tested the two strategies on the synthetic time-histories obtained for the 2019 Le Teil earthquake (France).


How to cite: Jacquet, G., Clouteau, D., and Gatti, F.: Hybrid generation based on machine learning to enhance numerical simulation for earthquake, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5252,, 2023.