LDLe Duc

Le Duc

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Le Duc has a total of 49 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Interval Data Assimilation: A Data Assimilation Scheme Where All Observations Are Intervals

    by Le Duc, Naoki Hiura, Yohei Sawada

    With the proliferation of non‐conventional observations from low‐cost sensors, citizen science networks, and binary detection systems, data assimilation increasingly confronts observations that take the form of inequalities or intervals rather than precise point values. Yet most existing methods assume point‐valued observations and cannot fully utilize inequality‐constrained data. This paper introduces Interval Data Assimilation (IDA), a unified Bayesian framework that seamlessly incorporates both equality and inequality observations within a single algorithmic structure. The key insight is that all observations can be represented as intervals: a point observation is a special case where the interval collapses to a single value, while an inequality observation that a variable is below a threshold corresponds to a one‐sided interval. We show that the logistic likelihood function uniquely satisfies the requirements for a well‐posed formulation, including convexity of the cost function and automatic quality control for outliers. Numerical experiments with the Lorenz‐96 and quasi‐geostrophic models demonstrate that IDA matches the Ensemble Transform Kalman Filter for conventional observations while achieving substantial improvements when inequality observations are incorporated. We further show that IDA handles observation biases effectively through interval representation and can be interpreted as a soft solver for quadratic programming problems, enabling applications in model predictive control. We also demonstrate that logistic regression emerges as a special case of IDA, offering a more principled justification for the logistic function and the binary cross‐entropy loss. These results make IDA valuable for atmospheric modeling and environmental monitoring while opening a bridge between data assimilation and machine learning.

  • Open Access

    Optimised initial conditions from machine-learning weather models: extended forecast skill is shadowing, not predictability

    by Le Duc, Yohei Sawada

    Optimising the initial condition of a machine-learning weather model by backpropagation has been shownto produce deterministic forecasts that remain skilful for about a month, well beyond the classicaltwo-week limit of atmospheric predictability. The correction has been interpreted as revealing errors inthe analysis it improves, implying that such models recover atmospheric information that conventionalanalyses miss. This interpretation cannot be tested against a real analysis, itself only an estimate ofthe state. We test it in a controlled twin experiment in which a differentiable primitive-equation modelprovides the truth and a neural emulator trained on it provides the machine-learning model, so that thesame optimisation runs through either dynamics against a known answer. The optimisation recovers agenuine part of the true initial error, more through the emulator than through the true model. Theoptimised initial condition is nevertheless no closer to the truth than the background, because theincrement also contains a large component that does not correct the state but improves the fit of theforecast trajectory. This component grows as the optimisation window lengthens, so a short window yieldsa better estimate of the state and a long one a worse. The forecast skill, by contrast, is nearlyindependent of the window and does not require the degraded initial condition. The extended skill istherefore real, but it does not measure how predictable the atmosphere is from its analysis: theoptimised initial condition is a better forecast trajectory, not a better estimate of the state.

  • Open Access

    Bayesian Flow Matching: Amortized Posterior Inference for Hydrological Parameter Estimation

    by Le Duc, Yohei Sawada

    Hydrological models must be calibrated from streamflow, yet equifinality, in which many distinct parameter sets reproduce the record almost equally well, makes a single best fit misleading and a posterior distribution over parameters the proper object of inference. Markov chain Monte Carlo, with the DREAM sampler the gold standard in hydrology, is the established route to that posterior, but it is not amortized, so its cost is repaid at every new record, and on concentrated posteriors it tends to lock onto one mode and report it over-confidently. This study proposes Bayesian Flow Matching (BFM), an amortized simulation-based method that learns, once, a conditional flow transporting a Gaussian to the parameter posterior, after which inference on any new record is a single forward pass. The flow is trained on prior simulations in three stages of increasing fidelity to the data, conditioned on the observed streamflow together with its forcing, and infers the hydrological and residual-error parameters jointly under a generalized likelihood. We compare BFM against DREAM on three conceptual models of 5, 11, and 16 parameters. BFM matches and, on the joint posterior, exceeds the calibration of DREAM, keeping the unidentifiable error-shape parameters honestly broad where the sampler is over-confident. Its streamflow forecasts equal or exceed DREAM on accuracy while producing predictive intervals several times sharper at equal or better reliability. These properties are preserved as the dimension grows, where DREAM's chains fail to converge without model-specific intervention, and once trained BFM infers in about half a second against multi-hour chains.

  • Open Access

    Real-time monitoring of urban flooding by ensemble Kalman filter

    by Juyoung Jo, Le Duc, Yohei Sawada

    The graphical abstract of the SWMM-DA framework integrating the physics-based SWMM with real-time data assimilation (DA). Real-time monitoring of urban flooding is challenging, as the required lead times are short and uncertainties in both hydrodynamic models and meteorological inputs substantially undermine monitoring accuracy. In particular, the complex networks of urban drainage systems (UDS) — governed by numerous physical parameters — make it difficult to maintain reliable monitoring performance during unprecedented storm events. In this study, we applied an Ensemble Kalman Filter (EnKF) to the widely used Storm Water Management Model (SWMM) to reduce these uncertainties. The proposed SWMM with data assimilation (SWMM-DA) system assimilates real-time sewer water level observations to sequentially update both model states and parameters. We evaluated the SWMM-DA system across, several severe storm events comparing its performance against a baseline SWMM using an existing manually calibrated parameter set. The key findings indicate that (1) jointly updating model states and parameters significantly improves urban flood monitoring accuracy without requiring prior parameter calibration from long-term historical observations, (2) shorter EnKF update windows are essential to maximize the potential of sequential data assimilation, demonstrating the feasibility of physics-based UDS models for real-time applications, and (3) the EnKF effectively compensates for model uncertainties by dynamically adjusting physical SWMM parameters.

  • Open Access

    Interval Data Assimilation: A Data Assimilation Scheme Where All Observations Are Intervals

    by Le Duc, Naoki Hiura, Yohei Sawada

    With the proliferation of non-conventional observations from low-cost sensors, citizen science networks, and binary detection systems, data assimilation increasingly confronts observations that take the form of inequalities or intervals rather than precise point values. Yet most existing methods assume point-valued observations and cannot fully utilize inequality-constrained data. This paper introduces Interval Data Assimilation (IDA), a unified Bayesian framework that seamlessly incorporates both equality and inequality observations within a single algorithmic structure. The key insight is that all observations can be represented as intervals: a point observation is a special case where the interval collapses to a single value, while an inequality observation that a variable is below a threshold corresponds to a one-sided interval. We show that the logistic likelihood function uniquely satisfies the requirements for a well-posed formulation, including convexity of the cost function and automatic quality control for outliers. Numerical experiments with the Lorenz-96 and quasi-geostrophic models demonstrate that IDA matches the Ensemble Transform Kalman Filter for conventional observations while achieving substantial improvements when inequality observations are incorporated. We further show that IDA handles observation biases effectively through interval representation and can be interpreted as a soft solver for quadratic programming problems, enabling applications in model predictive control. We also demonstrate that logistic regression emerges as a special case of IDA, offering a more principled justification for the logistic function and the binary cross-entropy loss. These results make IDA valuable for atmospheric modeling and environmental monitoring while opening a bridge between data assimilation and machine learning.

  • Open Access

    FORECASTING MULTI-HAZARD SCENARIOS WITH ENSEMBLE TROPICAL CYCLONE FORECASTS

    by Rezuanul Islam, Le Duc, Yohei Sawada

    Multi-hazard events resulting from tropical cyclones (TC) involve the simultaneous or sequential occurrence of various destructive disasters, including storm surges, heavy rainfall, and strong winds, which collectively pose significant threats to society (Needham et al., 2015; Takagi et al., 2022). While ensemble forecasts are widely employed in TC prediction, their potential for comprehensively assessing multi-hazard scenarios has been ignored (Titley et al., 2019). In this study, we introduce Pareto optimization based an innovative, efficient, and practical framework for analyzing multi-hazard scenarios, utilizing large ensemble TC forecasts comprising 1000 members to evaluate storm surge, rainfall, and wind-related risks. To exemplify the applicability of our approach, we conducted simulations for TC Hagibis that hit Central Japan in 2019 using the non- hydrostatic model from the Japan Meteorological Agency (JMA). The resulting atmospheric predictions served as inputs for modeling storm surges, rainfall patterns, and wind hazards. We anticipate that this assessment framework will empower forecasters, disaster management authorities, and communities to make well-informed decisions and enhance their resilience in the face of TC-related impacts.

  • Open Access

    Advancing Multi-Hazard Analysis: Worst Case Scenarios from Ensemble Tropical Cyclone Forecasts

    by Rezuanul Islam, Tsutao Oizumi, Le Duc, Takuma Ota, Takuya Kawabata, Yohei Sawada

    Ensemble forecasting is a powerful tool for supporting informed decision-making in managing multi-hazard risks associated with tropical cyclones (TCs). While TC ensemble forecasts are widely utilized in operational numerical weather prediction systems, their potential for disaster prediction remains underutilized. Here we propose a novel, efficient, and practical method to extract meaningful multi-hazard worst case scenarios (MHWCS) from a large ensemble TC forecast of 1000-members for the first time. We perform the simulation of TC Hagibis (2019) using the Japan Meteorological Agency's (JMA) nonhydrostatic model. The simulated atmospheric predictions were serving as inputs for JMA’s operational flood forecast model, as well as statistical storm surge and gust wind models. These models estimate river flooding, storm surge, and wind hazard intensities in Tokyo. By accounting for uncertainties in ensemble multi-hazard forecasts, we objectively demonstrate that Pareto-optimal solutions can effectively identify the meaningful MHWCS. These solutions illustrate complex trade-offs among competing hazard components across various forecast locations. Notably, the meaningful MHWCS do not necessarily represent the most extreme values for individual hazards but instead maximize hazard intensities relative to the ensemble mean, collectively leading to significant disaster impacts. Our findings further underscore the importance of evaluating Pareto-optimal solutions to assist risk managers in understanding how combinations of TC meteorological variables—such as track, translation speed, intensity, size, and rainfall—shape worst-case scenarios. For instance, meaningful MHWCS forecasts tend to exhibit moderate meteorological characteristics comparable to the ensemble mean, with variability in translation speed emerging as the strongest single predictor of single-hazard worst-case scenarios.

  • Open Access

    Realtime monitoring of urban flooding by ensemble Kalman filters

    by Le Duc, Juyoung Jo, Yohei Sawada

    Urban drainage models have been used in many cities for analysis, prediction, and control related to urban flooding. Many sources of uncertainties exist in these models comprising model parameters, meteorological forcings, and surface conditions. Thus, it is necessary to calibrate models before using them in reality. A common choice in calibration is to fit the model outputs with observations through many cases. This strategy is known as the offline mode in calibration and works on the stationary assumption of model parameters. If parameters vary in time, this method usually yields the climatological range of the parameters, which are not necessarily optimal in specific cases. In this study, instead of the offline model we follow the online mode in estimating model parameters by using an ensemble Kalman filter (EnKF). Furthermore, we estimate not only model parameters but also model states simultaneously utilizing the EnKF. Note that originally, EnKF is a data assimilation technique that is based on sampling in estimating any system states given observations, and later is used for the purpose of parameter estimation. The combination of EnKF and an urban drainage model is expected to lead to a real-time monitoring system for urban flooding similar to reanalysis systems in numerical weather prediction.

  • Open Access

    Advanced Impact-Based Forecast for a Severe Flood Event Using a 1000-member Ensemble Weather Prediction at Convective-Scale

    by Tsutao Oizumi, Takuya Kawabata, Le Duc, Kenichiro Kobayashi, Kazuo Saito, Takuma Ota


  • Living With Hyraxes: Biogeography and Comparative Ecology of West African Fornasinius Beetles

    by Stephanie N. Ajong, Luca Luiselli, Thomas Francis Lado, Gift Simon Demaya, Edem A. Eniang, Gabriel Hoinsoude Segniagbeto, Guillaume Koffivi Ketoh, Olivier Le Duc, Michele De Palma, Giovanni Amori, et al.

    Among the flower beetles (Scarabaeidae, Cetoniinae), the Goliathini comprise several genera of medium‐ and large‐sized beetles widely distributed in sub‐Saharan Africa. In this tribe, the genus Fornasinius Bertoloni, 1853, includes two species found in West Africa: F. higginsi (Westwood) and F. klingbeili Zöller, Fiebig,and Schulze. In this study, we present new data on the comparative ecology and biogeography of these two species, including sex ratio, population structure by size, monthly activity patterns and habitat features at two different spatial scales. These observations were conducted, mostly opportunistically, over a 20‐year period in Cote d'Ivoire, Ghana and Togo. Both species exhibited similar overall population structure and morphometrics. They were characterised by a male‐skewed adult sex ratio (with possible bias), male‐larger sexual size dimorphism and consistent population structure by size; the two species are seemingly parapatric (with the potential contact zone being separated in eastern Ghana by the Volta River and Volta Lake). Both are primarily found in the vicinity of trees that contained middens of hyraxes, such as Dendrohyrax dorsalis (F. higginsi), Dendrohyrax interfluvialis (F. klingbeili) and Procavia capensis (F. klingbeili). F. higginsi has been primarily observed in forest habitat, while F. klingbeili was found in both forests and Guinea savannahs. Occurrence sites of F. higginsi were predominantly characterised by higher tree cover, less bare areas and built‐up zones. Although these beetles may be locally abundant within hyrax middens, they typically occur in only a relatively small number of sites within their preferred habitat. Monthly activity patterns differed between the two species. Although adults of both species were observed throughout the year, sightings of F. higginsi were more common during the wet season, whereas F. klingbeili sightings were more frequent in the dry season. The observed differences in monthly activity patterns between these ecologically and morphologically similar species may help to minimise interspecific competition in potential—yet unknown—sites of syntopy. Although some Fornasinius populations might be locally threatened by deforestation or overhunting of hyraxes, the careful management of hyrax populations should adequately ensure the conservation of these two beetle species.