Search

Infrasound detections of Mediterranean cyclones known as medicanes (for « Mediterranean hurricanes ») are demonstrated in low- and high- frequency ranges, respectively. We summarize the main findings of a recently published study [1]. We use data from the infrasound station IS48 of the International Monitoring System, in Tunisia, to investigate the infrasound signatures of these meso-cyclones, using a multi-channel correlation algorithm. We discuss cases of detections and non-detections, based on the state of the middle atmosphere and of the wind noise measured at the station. Detections and likely sources are discussed in light of other datasets, comprising satellite observations of deep convection [2] and cloud-to-ground lightning detections from a ground-based network. Detections of infrasound emitted by the cyclone-related swell are modelled using a microbarom source model [3] and are in agreement with observations, comforting the identification of the lower frequency sources. This multi-technology and modelling approach allows to discuss the various sources at plat that may contribute to the monitoring of such extreme meteorological events.[1] Listowski, C.; Forestier, E.; Dafis, S.; Farges, T.; De Carlo, M.; Grimaldi, F.; Le Pichon, A.; Vergoz, J.; Heinrich, P.; Claud, C. Remote Monitoring of Mediterranean Hurricanes Using Infrasound. Remote Sens. 2022, 14, 6162. https://doi.org/10.3390/rs14236162[2] Dafis, S.; Claud, C.; Kotroni, V.; Lagouvardos, K.; Rysman, J. Insights into the convective evolution of Mediterranean tropical-like cyclones. Q. J. R. Meteorol. Soc. 2020, 146, 4147–4169.[3] De Carlo, M.; Accensi, M.; Ardhuin, F.; Le Pichon, A. ARROW (AtmospheRic InfRasound by Ocean Waves): A new real-time product for global ambient noise monitoring. In Proceedings of the EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022

L'éruption explosive du volcan Hunga Tonga - Hunga Ha’apai (HTHH), le 15 janvier 2022, a produit la plus puissante explosion enregistrée depuis les explosions du Krakatau et du Tambora dans les années 1800, libérant une énergie équivalente à 110 mégatonnes de TNT. Les ondes générées se sont propagées dans le sol, et dans l’atmosphère jusqu’à l’ionosphère. L'onde atmosphérique la plus énergétique observée sur les baromètres correspond au mode de Lamb. De période supérieure à 2000 s, son amplitude est comparable à celle observée lors de l’éruption du Krakatau en 1883. L’empreinte des perturbations atmosphériques a été caractérisée à l’échelle planétaire par des réseaux de mesures au sol, à bord de satellites ou de plateformes aéroportées. L’analyse combinée de ces observations a permis d’évaluer les conséquences à court terme de l'éruption du HTHH. Les méthodes d'investigation géophysiques présentées dans cette note montrent l’apport d’analyses interdisciplinaires pour caractériser la réponse impulsionnelle des enveloppes fluides planétaires (atmosphère, océans et mers) à une éruption d’une intensité exceptionnelle.

International audience; In this paper, we analyze the relevance of the use of the shallow water model and the Boussinesq model to simulate tsunamis generated by a landslide. In a first part, we determine if the two models are able to reproduce waves generated by a landslide. Each model has drawbacks but it seems that it is possible to use them together to improve the simulations. In a second part we try to recover the landslide displacement from the generated wave. This problem is formulated as a minimization problem and we limit the number of parameters to determine assuming that the bottom can be well described by an empirical law.; Dans cet article, nous analysons la pertinence de l'utilisation de modèles intégrés suivant la verticale tels que le modèle de Saint-Venant et le modèle de Boussinesq pour simuler les tsunamis générés par un glissement de terrain. Dans une première partie, nous déterminons si les deux modèles sont capables de reproduire les vagues générées par un glissement de terrain. Chaque modèle présente des inconvénients mais il semble qu'il soit possible de les utiliser ensemble pour améliorer les simulations. Dans une deuxième partie, nous essayons de retrouver le déplacement du glissement de terrain à partir de la vague générée. Ce problème a été formulé comme un problème de minimisation et nous limitons le nombre de paramètres à déterminer en supposant que le fond peut être bien décrit par une loi empirique.

International audience; One of the features of Boussinesq-type models for dispersive wave propagation is the presence of mixed spatial/temporal derivatives in the partial differential system. This is a critical point in the design of the time marching strategy, as the cost of inverting the algebraic equations arising from the discretization of these mixed terms may result in a nonnegligible overhead. In this paper, we propose novel approaches based on the classical Lax–Wendroff (LW) strategy to achieve single-step high-order schemes in time. To reduce the cost of evaluating the complex correction terms arising in the Lax–Wendroff procedure for Boussinesq equations, we propose several simplified strategies which allow to reduce the computational time at fixed accuracy. To evaluate these qualities, we perform a spectral analysis to assess the dispersion and damping error. We then evaluate the schemes on several benchmarks involving dispersive propagation over flat and nonflat bathymetries, and perform numerical grid convergence studies on two of them. Our results show a potential for a CPU reduction between 35 and 40% to obtain accuracy levels comparable to those of the classical RK3 method.

On June 17 2017, the western coast of Greenland was the site of a tsunami which flooded several villages, killing 4 people and destroying 11 houses in the village of Nuugaatsiaq. This tsunami was triggered by a subaerial landslide which occurred in a fjord 32 km ENE of Nuugaatsiaq. This paper presents the numerical modeling of this landslide of $$\sim$$ 50 million $$\hbox {m}^{3}$$ and of the tsunami propagation from its source to Nuugaatsiaq. The landslide is considered as a granular flow under gravity forces and the water waves generated are related to the displacement of the sea bottom. The results obtained are similar in amplitude to our inferences from videos, i.e., three water waves between 1 and 1.5 m arriving at Nuugaatsiaq with a period of roughly 3 min, and are also in general agreement with the amplitude (1 m) resulting from deconvolution of oscillations recorded on a horizontal seismogram operating at Nuugaatsiaq (NUUG). According to the field survey performed by Fritz et al. (EGU General Assembly Conference Abstracts, Vol. 20 of EGU General Assembly Conference Abstracts, p 18345, 2018a) on July 2017, a second mass next to the landslide is threatening Karrat Fjord. A sensitivity study is realized on its volume, with 2, 7, 14 and 38 million $$\hbox {m}^{3}$$ reaching the sea. The shape of the water waves is found to be independent of volume, and linearity is observed between the volume and the water wave heights. Finally, the orientation of the slide does not seem to influence either the period or the shape of the generated water waves.

The South East Aegean Sea is seismically active with crustal and intermediate-depth earthquakes of magnitude that can exceed the magnitude Mw = 7. On July 20, 2017, a shallow normal faulting Mw = 6.6 earthquake occurred in the Gulf of Gokova, 10 km away from the cities of Kos and Bodrum. A tsunami hit the coasts of Turkey and Kos Island with waves of up to 2 m that damaged several harbors. Based on the field surveys and the tide-gauge records of Bodrum, Syros and Kos harbors, we performed tsunami simulations to constrain the source and analyze the North v/s South-dipping solutions for the fault plane. Depending on the source model, the simulations show different height distributions which provide complementary constrains on the source parameters, helping the characterization of the fault plane. The comparison between our maximum wave heights models and the field data favor the North-dipping fault scenario. When using only the constraint of the Bodrum tide-gauge, our models lead to a source with a 0.7 m fault slip, incompatible with the slip inferred from geodesy. When using the constrain from neighboring coastal wave heights and run-up recovered during field surveys, the best fit is obtained with a 2 m fault slip scenario consistent with the geodetic models. There is a 0.25 m mean difference between observed and synthetic water heights that could be due to the unaccuracy of the bathymetry data.

In the literature, OpenFOAM has been used to simulate landslide tsunamis, modeling the landslide as a solid or a two-phase flow. Here we present an approach using three phases (air, water and sediment) modeled as Newtonian fluids. This 3D model is validated against two benchmarks with deformable landslides: one subaerial (Viroulet et al. 2016) and one submarine (Grilli et al. 2017). These benchmarks are also run by a 2D depth-integrated model, AVALANCHE, recently used to reproduce the 2017 Karrat Fjord, Greenland and the 2018 Anak Krakatau, Indonesia events. Both models are able to reproduce either the water waves or the landslide behavior but not both at the same time.Considering OpenFOAM as a reference code, sensitivity studies on the slope angle, the landslide viscosity and the landslide initial submergence showed that AVALANCHE produces similar results for slope angles between 10 and 45°, for subaerial or close to the surface landslide and/or for low viscosity values. In the other cases (submarine landslides and higher viscosity values) results indicated that OpenFOAM should be preferred to a 2D depth-integrated model. References:Grilli, S., Shelby, M. & Kimmoun, O. (2017), ‘Modeling coastal tsunami hazard from submarine mass failures: effect of slide rheology, experimental validation, and case studies off the US East Coast’, Natural Hazards 86, 353-391.Viroulet, S., Sauret, A., Kimmoun, O. & Kharif, C. (2016), Tsunami waves generated by cliff collapse: comparison between experiments and triphasic simulations, in E. Pelinovsky & C. Kharif, eds, ‘Extreme Ocean Waves’, Springer International Publishing, Cham, pp. 173-190.

International audience; The granting process of all credit institutions is based on the probability that the applicant will refund his/her loan given his/her characteristics. This probability also called score is learnt based on a dataset in which rejected applicants are de facto excluded. This implies that the population on which the score is used will be different from the learning population. Thus, this biased learning can have consequences on the scorecard's relevance. Many methods dubbed "reject inference" have been developed in order to try to exploit the data available from the rejected applicants to build the score. However most of these methods are considered from an empirical point of view, and there is some lack of formalization of the assumptions that are really made, and of the theoretical properties that can be expected. In order to propose a formalization of such usually hidden assumptions for some of the most common reject inference methods, we rely on the general missing data modelling paradigm. It reveals that hidden modelling is mostly incomplete, thus prohibiting to compare existing methods within the general model selection mechanism (except by financing "non-fundable" applicants, which is rarely performed in practice). So, we are reduced to empirically assess performance of the methods in some controlled situations involving both some simulated data and some real data (from Crédit Agricole Consumer Finance (CACF), a major European loan issuer). Unsurprisingly, no method seems uniformly dominant. Both these theoretical and empirical results not only reinforce the idea to carefully use the classical reject inference methods but also to invest in future research works for designing model-based reject inference methods, which allow rigorous selection methods (without financing "non-fundable" applicants).

The arc of the Lesser Antilles is associated with a significant tectonic activity due to the subduction of the Atlantic oceanic plate under the Caribbean plate. Earthquakes in this context have the potential to trigger landslides and tsunamis due to the important vertical seafloor displacement. The historical tsunamigenic earthquakes in this region are rare, but the damages they may have generated before along the coasts show that they pose a considerable threat to the closest inhabited islands. The most recent tsunamigenic earthquake occurred in 2004 in the area of Les Saintes, along a normal fault system located in the back-arc of the subduction. This $$M_{w}=6.3$$ earthquake generated small waves with 2 m of run-up in several bays of Les Saintes, a group of islands in the South of Guadeloupe. A recent survey conducted in the source area using deep-sea vehicles revealed for the first time an important coseismic slip on the Roseau fault plane, attributed to the 2004 event, which had not been predicted in the seismic inversion models. This event and the dataset on the Roseau fault gives the opportunity to model precisely the earthquake, to compare the simulation results with the observations and to evaluate the impact of the rupture heterogeneity and rupture shallowness on the height of the tsunami waves. Extending our earlier work (Cordrie et al. in OCEANS 2019—Marseille, Jun 2019, Marseille, France, pp 1–9, 2019. https://doi.org/10.1109/OCEANSE.2019.8867447 ) and in order to avoid a loss of quality from the dataset in the modeling of the initial sea-surface deformation especially in shallow depth and near field context, a transfer function of the deformation from the seafloor to the sea surface and different numerical schemes were used. Results on the tsunami height distribution indicate some local tsunami amplification phenomena linked to the bathymetry or the coastline geometry and highlight the most endangered areas of the islands. The simulations give additional constraints on the source, show the impact of the slip heterogeneities on the tsunami and finally provide a complementary estimation of the 2004 coseismic slip intensity.

International audience; On the evening of December 22, 2018, the coasts of the Sunda Strait, Indonesia, were hit by a tsunami generated by the collapse of a part of the Anak Krakatau volcano. Hundreds of people were killed, thousands were injured and displaced. This paper presents a preliminary modeling of the volcano flank collapse and the tsunami generated based on the results of a 2D depth-averaged coupled model involving a granular rheology and a Coulomb friction for the slide description and dispersive effects for the water flow part. With a reconstructed total volume (subaerial and submarine) of the landslide of 150 million m 3 inferred from pre and post-collapse satellite and aerial images , the comparison of the simulated water waves with the observations (tide gauges located all around the strait, photographs and field surveys) is satisfactory. Due to the lack of information for the submarine part of the landslide, 2 Alexandre Paris et al. the reconstructed submarine slope is assumed to be approximately constant. A significant time delay on the results and particularly in the Bandar Lampung Bay could be attributed to imprecisions of bathymetric data. The sensitivity to the basal friction and to dispersive effects is analyzed through numerical tests. Results show that the influence of the basal friction angle on the simulated wave heights decreases with distance and that a value of 2°gives consistent results with the observations. The dispersive effects are assessed by comparing water waves simulated by a shallow water model and a Boussinesq model. Simulations with frequency dispersion produce longer wave periods and smaller wave amplitudes in the Sunda Strait and particularly in deep waters.

Two approaches are proposed to simulate tsunamis generated by granular landslides: a depth-averaged model, AVALANCHE, and laminar Navier–Stokes simulations using the OpenFOAM model. Both models are validated against two 2D benchmarks, a subaerial and a submerged one involving a triangle initial slide shape. In both models, the landslide is defined as a viscous fluid flowing downslope. Except in the first instants, both models can reproduce either the landslide behavior or the generated first water waves but cannot reproduce simultaneously both the landslide and the water surface. There is an overlap in the optimal viscosity range between both models. Sensitivity studies are carried out by varying the slope angle and the landslide submergence for the subaerial benchmark. The largest waves are obtained for initial landslide position close to the free surface. The height of the generated waves increases linearly with the slope angle and the landslide Reynolds number in the depth-averaged model. The relationships are more complex in the Navier–Stokes model. For low slide Reynolds numbers and with an initial slide close to the free surface, both models produce similar waves. Nine additional cases are performed with a k − e turbulence closure model and varying the submergence and the slope angle.

This paper aims at best describing the submarine landslide which induced partial submersion of the atolls of Mururoa and Fangataufa in 1979. More precisely, waves propagated along the south coast of Mururoa atoll and penetrated into its lagoon some minutes after the landslide triggering (t = 0 s), whereas a train of eight water waves reached the runway located on the north-east coast of Fangataufa (40 km south of Mururoa) between t = 7 min 30 s and t = 20 min. A numerical model based on shallow water equations is used to simulate the landslide as well as the associated tsunami. Saint-Venant equations are used to propagate the tsunami in coastal areas, whereas the offshore propagation is simulated by solving weakly nonlinear Boussinesq equations. Low- and high-resolution nested grids are used to simulate the tsunami propagation in deep sea and in shallow waters, respectively. Several scenarios have been tested to reproduce the observed water and run-up heights in the near and far fields. The best scenarios correspond to a landslide with a volume in the range (75–90 Mm3) (for a basal friction angle of 35°) and with a basal friction angle in the range (30°–40°) (for a volume of 80 Mm3). These results have been completed by a parametric study on the slide parameters.

In this paper, we present a new source assessment of the La Palma collapse scenario previously described and studied in Abadie et al. (2012). Three scenarios (i.e., slide volumes of 20, 40 and 80 km3) are considered, from the initiation of the slide to the water waves generation, using THETIS, a 3D Navier–Stokes model. The slide is considered as a Newtonian fluid whose viscosity is adjusted to approximate a granular behavior. After 5 minutes of propagation with THETIS, the generated water wave is transferred into FUNWAVE-TVD for 15 minutes of Boussinesq model simulation. Then, four different depth-averaged codes are used to propagate the wave to the Guadeloupe area, Europe and French coasts. Finally, the wave impact in terms of run-up is evaluated through direct computations in specific areas or using theoretical formulas. Although the wave source appears reduced due to the rheology used compared to former works, the wave impact is still significant for the maximum slide volume considered on surrounding islands and coasts, as well as on remote most exposed coasts such as Guadeloupe. In Europe and in France, the wave impact is moderate (for specific areas in Spain and Portugal) to weak (Atlantic French coast). The comparison between the different wave models in overlapping computational regions shows an overall agreement in terms of first wave amplitude and time of arrival, but differences appear in the trailing waves.

In the evening on December 22, 2018, the Sunda Strait, Indonesia, suffered a destructive tsunami that no one had seen coming. It killed more than 400 people, injured and displaced thousands more. Post event satellite images indicate that the tsunami was probably caused by the collapse of a part of the southwestern flank of the Anak Krakatau volcano. Assuming a landslide volume of 150 million m3 sliding into water as a granular flow under gravity forces, the simulated water waves are in general agreement with the tide gauges that recorded the tsunami.

Subduction zones are areas with significant tectonic activity, complex fault systems, volcanic edifices and growth of islands. Earthquakes in this context are associated with sudden subsidence or uplift of the sea floor with the potential to trigger landslides and tsunamis. A tsunami initiated by a megathrust earthquake can generate strong currents, and is able to carry massive quantities of sediments. The inhabited islands and coasts near subduction areas are threatened by those destructive earthquakes and tsunamis. The arc of the Lesser Antilles which formed due to the plunge of the Atlantic oceanic plate under the Caribean plate, is one of them. The tsunamigenic earthquakes in this region are rare, poorly documented and the only available data are qualitative (observations of coseismic displacement, intensity of the damages). Quantitative data is lacking as no seismograms or tide-gauge were set up before 1968. This lack of data increases the uncertainty on the existent fault models and the shortness of the catalog of seismicity does not enable a good analysis of the recurence cycle of the megathrust earthquakes. The most recent tsunamigenic earthquake occured in 2004 in the area of Les Saintes normal fault system. This Mw6.3 earthquake generated small waves with 2m of run-up in several bays of Les Saintes, a group of island in the South of Guadeloupe. A recent survey done on the fault scarp revealed important coseismic displacement at the surface which had not been predicted in the seismic source models resulting from the joint inversion of teleseismic and GPS data. Tsunami simulations of those different models of fault enable the calculation of the associated time arrivals and water heights of the waves. Compared to the reported observation, those results give complementary informations and additional constrains on the source parameters.

We study the rates of estimation of finite mixing distributions, that is, the parameters of the mixture. We prove that under some regularity and strong identifiability conditions, around a given mixing distribution with $m_{0}$ components, the optimal local minimax rate of estimation of a mixing distribution with $m$ components is $n^{-1/(4(m-m_{0})+2)}$. This corrects a previous paper by Chen [Ann. Statist. 23 (1995) 221–233]. ¶ By contrast, it turns out that there are estimators with a (nonuniform) pointwise rate of estimation of $n^{-1/2}$ for all mixing distributions with a finite number of components.

The trans-pacific propagation of the 2010 and 2015 Chilean tsunamis have been simulated by resolving both the Saint-Venant and the Boussinesq equations, these latter equations taking into account the frequency dispersion of waves. Numerical results show that the frequency dispersion simulation must be considered to reproduce the wave heights and arrival times recorded by several DART buoys in the Pacific Ocean. Several grid resolutions have been tested using either the Saint-Venant or the Boussinesq model (10′, 5′ and 2′). Only the Saint-Venant simulation at a 5′ resolution and the Boussinesq simulation at a 2′ resolution allow to reproduce the amplitudes and arrival times of the first waves. Numerical results indicate that frequency dispersion has a significant impact on the computed waveforms and suggest that the Saint-Venant code generates numerical dispersion similar to the physical frequency dispersion. It has been confirmed by one dimensional sensitivity tests. By comparing both events (2010 and 2015), we show that the frequency dispersion is stronger for the second event and it has been confirmed by the analysis of the Absolute Pressure Gages KSR1 and MRT1 spectrograms at some 17000 km from the source. Indeed, the main wave energy is contained in the first waves for the 2010 events and in the following waves for the 2015 event. Nevertheless, for the 2010 event, we also detect some late arrival waves at the APG KSR1, three hours after the first arrival time, with energy only two times lower than the first wave one.

Resume : La propagation trans-Pacifique du tsunami chilien de 2010 est simulee en resolvant soit les equations de Saint-Venant soit celles de Boussinesq, ces dernieres prenant en compte la dispersion frequentielle des vagues. Les resultats numeriques indiquent que la modelisation de la dispersion frequentielle est necessaire pour reproduire les hauteurs d’eau enregistrees par differentes bouees DART du Pacifique. Independamment du modele choisi, la resolution des grilles utilisees modifie significativement la forme du tsunami modelise. Des tests de sensibilite monodimensionnels montrent que le modele de Boussinesq converge a partir d’un pas d’espace de 5 km et que le modele de Saint-Venant, utilise avec un pas d’espace de 10 km, genere de la dispersion numerique proche de la dispersion frequentielle physique. Enfin, l’analyse des spectrogrammes montre que la dispersion frequentielle du tsunami a un impact non negligeable sur le temps d’arrivee, l’amplitude et la frequence des vagues se propageant en champ lointain. Elle doit donc etre prise en compte pour l’etude de l’amplification des vagues lorsque celles-ci atteignent les cotes, et notamment dans les zones habitees a risque. Mots cles : Dispersion frequentielle, Tsunami, Simulation numerique, Bathymetrie.

After earthquakes, landslides are the second cause for tsunami generation. A proper understanding is required to prevent future disaster or to develop early warnings.This can be achieved through physical models in laboratory or numerical models. In the last category, several models exist and can provide very similar results for a case study. Among them, depth-averaged models using for example shallow water or Boussinesq equations, can be opposed to Navier-Stokes models. The main objective of this PhD thesis is to compare these two modeling strategies with two specific models, a depth-averaged model, AVALANCHE, and a Navier-Stokes model, OpenFOAM. First, two benchmarks (a subaerial and submerged one) are used to calibrate the models. This highlighted that both models could reproduce the experimental data and that several combinations of parameters led to similar results. Second, sensitivity studies are carried out to evaluate the influence of the initial landslide position and the slope angle and to observe the behavior of the different equations (shallow water, Boussinesq or Navier-Stokes) during the wave generation and propagation phases. Finally, both models are applied to two real cases, the June 17, 2017, Karrat Fjord, Greenland, landslide and tsunami, and the December 22, 2018, Anak Krakatau, Indonesia, collapse and tsunami, and are intercompared.; Les effondrements gravitaires sont la deuxième cause de génération de tsunamis après les séismes. Il est important de bien les comprendre afin de prévenir de futures catastrophes ou de développer les systèmes d’alerte. Pour cela, des modèles analogiques, en laboratoire, ou numériques sont utilisés. Dans la deuxième catégorie, de nombreux modèles existent et peuvent produire des résultats similaires pour un cas donné. Parmi eux, les modèles intégrés sur la profondeur qui utilisent par exemple les équations de type shallow water ou de Boussinesq, peuvent être opposés aux modèles Navier-Stokes. L’objectif de cette thèse est de comparer ces deux stratégies à l’aide de deux modèles spécifiques un modèle intégré sur la hauteur, AVALANCHE, et un modèle Navier-Stokes, OpenFOAM. Tout d’abord, les deux modèles sont calibrés grâce à deux benchmarks, un glissement subaérien et un submergé. Cette étude a montré que les deux modèles pouvaient reproduire les données expérimentales et que plusieurs combinaisons de paramètres permettaient d’obtenir les mêmes résultats. Ensuite, des études de sensibilité sont réalisées afin d’évaluer l’influence de la position initiale du glissement et de la pente, et d’observer le comportement des différentes équations (shallow water, Boussinesq ou Navier-Stokes) pendant les phases de génération et de propagation de la vague. Enfin, l’application des deux modèles à deux cas réels, le glissement et le tsunami du 17 juin 2017 dans le Karrat Fjord au Groenland et le tsunami généré par l’effondrement du volcan Anak Krakatau le 22 décembre 2018, en Indonésie, permet de les comparer entre eux.

This manuscript deals with model-based statistical learning in the binary classification setting. As an application, credit scoring is widely examined with a special attention on its specificities. Proposed and existing approaches are illustrated on real data from Crédit Agricole Consumer Finance, a financial institute specialized in consumer loans which financed this PhD through a CIFRE funding.First, we consider the so-called reject inference problem, which aims at taking advantage of the information collected on rejected credit applicants for which no repayment performance can be observed (i.e. unlabelled observations). This industrial problem led to a research one by reinterpreting unlabelled observations as an information loss that can be compensated by modelling missing data. This interpretation sheds light on existing reject inference methods and allows to conclude that none of them should be recommended since they lack proper modelling assumptions that make them suitable for classical statistical model selection tools.Next, yet another industrial problem, corresponding to the discretization of continuous features or grouping of levels of categorical features before any modelling step, was tackled. This is motivated by practical (interpretability) and theoretical reasons (predictive power). To perform these quantizations, ad hoc heuristics are often used, which are empirical and time-consuming for practitioners. They are seen here as a latent variable problem, setting us back to a model selection problem. The high combinatorics of this model space necessitated a new cost-effective and automatic exploration strategy which involves either a particular neural network architecture or a Stochastic-EM algorithm and gives precise statistical guarantees.Third, as an extension to the preceding problem, interactions of covariates may be introduced in the problem in order to improve the predictive performance. This task, up to now again manually processed by practitioners and highly combinatorial, presents an accrued risk of misselecting a ``good'' model. It is performed here with a Metropolis-Hastings sampling procedure which finds the best interactions in an automatic fashion while ensuring its standard convergence properties, thus good predictive performance is guaranteed.Finally, contrary to the preceding problems which tackled a particular scorecard, we look at the scoring system as a whole. It generally consists of a tree-like structure composed of many scorecards (each relative to a particular population segment), which is often not optimized but rather imposed by the company's culture and / or history. Again, ad hoc industrial procedures are used, which lead to suboptimal performance. We propose some lines of approach to optimize this logistic regression tree which result in good empirical performance and new research directions illustrating the predictive strength and interpretability of a mix of parametric and non-parametric models.This manuscript is concluded by a discussion on potential scientific obstacles, among which the high dimensionality (in the number of features). The financial industry is indeed investing massively in unstructured data storage, which remains to this day largely unused for Credit Scoring applications. Doing so will need statistical guarantees to achieve the additional predictive performance that was hoped for.; Cette thèse se place dans le cadre des modèles d’apprentissage automatique de classification binaire. Le cas d’application est le scoring de risque de crédit. En particulier, les méthodes proposées ainsi que les approches existantes sont illustrées par des données réelles de Crédit Agricole Consumer Finance, acteur majeur en Europe du crédit à la consommation, à l’origine de cette thèse grâce à un financement CIFRE.Premièrement, on s’intéresse à la problématique dite de ``réintégration des refusés''. L’objectif est de tirer parti des informations collectées sur les clients refusés, donc par définition sans étiquette connue, quant à leur remboursement de crédit. L’enjeu a été de reformuler cette problématique industrielle classique dans un cadre rigoureux, celui de la modélisation pour données manquantes. Cette approche a permis de donner tout d’abord un nouvel éclairage aux méthodes standards de réintégration, et ensuite de conclure qu’aucune d’entre elles n’était réellement à recommander tant que leur modélisation, lacunaire en l’état, interdisait l’emploi de méthodes de choix de modèles statistiques.Une autre problématique industrielle classique correspond à la discrétisation des variables continues et le regroupement des modalités de variables catégorielles avant toute étape de modélisation. La motivation sous-jacente correspond à des raisons à la fois pratiques (interprétabilité) et théoriques (performance de prédiction). Pour effectuer ces quantifications, des heuristiques, souvent manuelles et chronophages, sont cependant utilisées. Nous avons alors reformulé cette pratique courante de perte d’information comme un problème de modélisation à variables latentes, revenant ainsi à une sélection de modèle. Par ailleurs, la combinatoire associé à cet espace de modèles nous a conduit à proposer des stratégies d’exploration, soit basées sur un réseau de neurone avec un gradient stochastique, soit basées sur un algorithme de type EM stochastique.Comme extension du problème précédent, il est également courant d’introduire des interactions entre variables afin, comme toujours, d’améliorer la performance prédictive des modèles. La pratique classiquement répandue est de nouveau manuelle et chronophage, avec des risques accrus étant donnée la surcouche combinatoire que cela engendre. Nous avons alors proposé un algorithme de Metropolis-Hastings permettant de rechercher les meilleures interactions de façon quasi-automatique tout en garantissant de bonnes performances grâce à ses propriétés de convergence standards.La dernière problématique abordée vise de nouveau à formaliser une pratique répandue, consistant à définir le système d’acceptation non pas comme un unique score mais plutôt comme un arbre de scores. Chaque branche de l’arbre est alors relatif à un segment de population particulier. Pour lever la sous-optimalité des méthodes classiques utilisées dans les entreprises, nous proposons une approche globale optimisant le système d’acceptation dans son ensemble. Les résultats empiriques qui en découlent sont particulièrement prometteurs, illustrant ainsi la flexibilité d’un mélange de modélisation paramétrique et non paramétrique.Enfin, nous anticipons sur les futurs verrous qui vont apparaître en Credit Scoring et qui sont pour beaucoup liés la grande dimension (en termes de prédicteurs). En effet, l’industrie financière investit actuellement dans le stockage de données massives et non structurées, dont la prochaine utilisation dans les règles de prédiction devra s’appuyer sur un minimum de garanties théoriques pour espérer atteindre les espoirs de performance prédictive qui ont présidé à cette collecte.