Data Assimilation and Machine Learning

120th Data Assimilation and Prediction Science Seminar

2026年7月2日(木) 14:00 - 16:00

Upmanu Lall (Professor, Columbia University, USA)
Mengqian Lu (Professor, Hong Kong University of Science and Technology, Hong Kong)
Hyun-Han Kwon (Professor, University of Seoul, Republic of Korea)

Speaker: Professor Lall (Columbia University) Title: "Taming the Storm: Can We Predict, Engineer, and Reduce Losses from Climate and Weather Extremes?" Abstract: Climate and weather extremes — storms, heat waves, floods, droughts, and compound events — have become the defining natural hazard challenge of the 21st century. Their growing frequency and intensity are overwhelming engineered infrastructure, disrupting global supply chains, and propagating risks across societies through teleconnections that no single country can insulate itself against. While climate change mitigation through decarbonization remains an urgent priority, even optimistic emissions trajectories leave us facing decades of increasing exposure. Climate adaptation efforts — improved infrastructure design, financial instruments, early warning systems — are essential but are constrained by limited data, deep uncertainty in future projections, and the diffuse question of who bears responsibility for action. This talk argues that a third pillar is emerging and demands serious scientific and institutional attention: Climate Stabilization, or the deliberate modification of developing weather and climate extremes to reduce their societal impact. Rather than waiting for disasters to unfold and recovering afterward, this paradigm asks whether the physical dynamics of the atmosphere offer leverage points — windows in time and space — where strategically placed, small perturbations could redirect the trajectory of an extreme event. This is the core idea of Weather Jiu-Jitsu: exploiting the inherent instabilities and nonlinear sensitivities of atmospheric circulation to achieve large-scale redirection of an extreme using energy borrowed from the circulation itself, not brute-force external forcing. J The talk will address the foundational questions this agenda raises for a forecasting and Earth science community: What physical mechanisms enable or constrain atmospheric steering? How can ensemble prediction systems, adjoint methods, and emerging AI tools be harnessed to identify intervention points and compute impact outcomes with spatial specificity? What are the data and modeling gaps? How do we frame the ethical and governance dimensions as this moves from laboratory curiosity to potential operational deployment and commercial application?  I will sketch a research roadmap integrating chaos-informed perturbation theory to AI-enabled adaptive control optimization that builds on AI-accelerated impact forecasting to provide the foundation for Climate Stabilization as a rigorous scientific enterprise and, within a decade, a viable business with measurable returns to investors and societies alike. We hope that this will stimulate discussion with RIken's Moonshot Goal 8 program, which is exploring similar scientific and technological frontiers. Speaker: Professor Mengqian Lu (Hong Kong University of Science and Technology) Title: Bridging Climate Data to Actionable Decision-Making Across Industries Abstract: Extreme weather is escalating—impacting infrastructure, supply chains, and profitability across the world. At the same time, sustainability targets demand that businesses go green without sacrificing growth. The question is no longer if climate risk matters, but how to act on it—quickly and smartly. This talk presents climate solutions that combine advanced climate modeling with AI to deliver industry-specific, actionable insights. Developed at HKUST through the Center for Climate Resilience and Sustainability (CCRS) and the World Sustainable Development Institute (WSDI), this AI–dynamical hybrid system is already being applied across key sectors, including renewable energy, Arctic logistics, and disaster risk management. These tools enable organizations to make faster, more informed decisions under uncertainty. Backed by UNESCO and the WMO, this Research-to-Operation (R2O) framework turns complex climate data into operational tools that drive resilience, reduce losses, and uncover new opportunities. Real-world case studies will be shared to spark cross-sector collaboration between science, business, and policy. Speaker: Professor Hyun-Han Kwon (University of Seoul) Title: Bayesian Mixture Extreme-Value Modeling of Nonstationary Extreme Precipitation Across U.S. Regions Abstract Extreme precipitation is a major driver of flood risk, infrastructure stress, and climate-related disaster losses. However, annual maximum rainfall often reflects multiple physical mechanisms, including frontal or convective systems, tropical-cyclone-related rainfall, and transitional atmospheric regimes. Treating these extremes as samples from a single homogeneous process can obscure how regional rainfall risks are changing. This talk presents an ongoing study of nonstationary extreme precipitation using a Bayesian mixture extreme-value model. The model represents annual maximum daily precipitation as a combination of latent low- and high-intensity rainfall regimes, with time-varying component behavior and regime probabilities. This allows changes in return levels to be separated into contributions from baseline rainfall intensity, high-intensity event magnitude, and the probability of entering an extreme-producing regime. The framework is applied to long-term U.S. station records across the Southeast/Gulf, Mid-Atlantic, Northeast, and inland-control regions. Tropical-cyclone proximity and ERA5-based atmospheric diagnostics are used as external physical evidence, rather than as imposed predictors in the likelihood, to interpret the latent high-intensity regime and its regional variability. The broader goal is to move extreme-value analysis beyond stationary design estimation toward mechanism-aware and decision-relevant understanding of climate risk. By linking Bayesian uncertainty quantification, hydrometeorological interpretation, and regional comparison, this work provides a basis for improved infrastructure planning, impact-based forecasting, and future AI-enabled climate risk services.

会場: Hybrid Format (RIKEN R-CCS room C107 and Zoom)

イベント公式言語: 英語

Physics-based eruption forecasting at Kīlauea volcano using an Ensemble Kalman Filter

2026年4月7日(火) 13:00 - 14:30

Kyle R. Anderson (Research Geophysicist, California Volcano Observatory, U.S. Geological Survey (USGS), USA)

Today, most forecasts of volcanic eruptions are based on expert opinion, making them fundamentally subjective. Such forecasts have often proven successful but have clear limitations. Novel quantitative forecasting techniques have shown promise in experimental settings (hindcasting) but face numerous operational challenges and most have rarely if ever been applied to real-world eruptions (forecasting). In this talk I will discuss efforts to forecast a remarkable ongoing series of more than 40 high lava fountain eruptions at Kīlauea volcano, Hawaii, using a simple physics-based model in an Ensemble Kalman Filter (EnKF) data assimilation algorithm. Using this method, which is believed to be the first implementation of a physics-based EnKF eruption forecast, the times of Kīlauea’s lava fountain eruptions can be forecast days to weeks in advance. The method assimilates geodetic data to constrain the evolving state of the system, provides insight into the eruption mechanism and rate of magma supply to the volcano, and produces fully probabilistic forecasts. These forecasts are combined with other information, including forecasts based on machine learning algorithms, to derive forecast windows, which are disseminated to the public and to partner agencies for hazards mitigation activities. In this way, novel eruption forecasting tools are continually developed which serve an important public need while also improving understanding of the volcanic system.

会場: Hybrid Format (RIKEN R-CCS room C107 and Zoom)

イベント公式言語: 英語

Satellite Data Assimilation for Numerical Weather Prediction (NWP)

2026年3月31日(火) 14:30 - 16:00

Martin Weissman (Professor, Department of Meteorology and Geophysics, University of Vienna, Austria)

Satellite data assimilation for NWP has made tremendous progress over the past decades. Most of the assimilated observations in global NWP systems are nowadays satellite radiances from passive sensor in the infrared and microwave range. Additionally, GPS radio occultation provides information on upper-level humidity that serves as important uncalibrated anchoring information for humidity. Nevertheless, there are still significant limitations, especially in terms of lacking wind information that controls atmospheric dynamics in global NWP as well as in terms of using cloud-affected radiances in regional, convection-permitting NWP. My presentation will cover recent progress of my group in these fields.

会場: Hybrid Format (RIKEN R-CCS room C107 and Zoom)

イベント公式言語: 英語

Application of a one-dimensional scheme to model diurnal water temperature fluctuations near the surface of a stratified lake

2026年3月27日(金) 10:30 - 12:00

John Craig Wells (立命館大学 理工学部 環境都市工学科 教授 / 理化学研究所 計算科学研究センター (R-CCS) データ同化研究チーム 客員主管研究員)

When simulating the atmosphere across various scales, accurately resolving the diurnal warming of sea and lake surfaces is a critical requirement. For example, regional atmospheric models must correctly simulate air-water temperature gradients to successfully capture mesoscale circulations such as sea and lake breezes. Often the SST (or Lake Surface Temperature LST) applied to the atmospheric simulator is modelled using a “slab model” of a certain thickness and thermal mass. However slab models often predict diurnal variation of SST poorly. In this talk I will discuss preliminary results from “DiuSST”, recently proposed by R. Börner et al (2025; https://doi.org/10.5194/gmd-18-1333-2025) to provide boundary conditions for diurnally varying SST to atmospheric simulators. Börner et al ’s testing and validation of DiuSST was based on an ocean cruise that measured skin surface temperature with an infrared radiometer, and water temperature at 3m depth. By contrast I cross-check DiuSST results against near-surface temperature profiles in a stratified lake, Lake Biwa, that were recorded at nearshore and offshore locations during the early summer of 2021.

会場: Hybrid Format (RIKEN R-CCS room 107 and Zoom)

イベント公式言語: 英語

Taming the Butterfly: A New "Duality Principle" Turns Chaos into Control

2026年2月18日(水) 13:00 - 14:00

三好 建正 (理化学研究所 計算科学研究センター (R-CCS) データ同化研究チーム チームプリンシパル)

Data Assimilation (DA) is the backbone of modern weather forecasting. It integrates observational data into computer simulations to synchronize the model with nature. The Duality Principle posits that chaos control is mathematically the "twin" (dual) of DA. Data Assimilation: Uses observations to synchronize the Model to Nature. Chaos Control: Uses interventions to synchronize Nature to a desired Model ("target trajectory"). "The butterfly effect has long been a symbol of unpredictability," says Dr. Miyoshi. "But I asked a simple question: If a butterfly's wings can change the future, does that not imply that with the right, tiny push, we could choose a better future?" Instead of suppressing the chaotic system with massive force, this method acts like mathematical judo—leveraging the system's inherent instability. By applying minute, calculated "interventions" (analogous to the butterfly's flap), the system can be guided toward a "target trajectory"—for instance, shifting real-world conditions just enough to align with a model-simulated scenario where a typhoon causes no damage. Once synchronized, control becomes much easier to maintain. This study establishes the theoretical foundation for "Control Simulation Experiments" (CSE), a framework previously proposed by Miyoshi’s team. It provides a roadmap for future disaster prevention research, moving beyond passive prediction to active mitigation. Beyond meteorology, this general framework is expected to serve as a universal tool for studying interventions in various chaotic systems, from ecosystems to economics. Following the seminar, we will hold an informal discussion (brainstorming) on data assimilation with quantum computing in the same room from 2-4 pm.

会場: セミナー室 (359号室) 3階 359号室とZoomのハイブリッド開催

イベント公式言語: 英語

LEVERAGING EARTH OBSERVATIONS WITH MACHINE-LEARNING APPROACHES FOR WATER CYCLE MONITORING

2026年1月13日(火) 10:30 - 12:00

Victor Pellet (Professor, Laboratoire de Météorologie dynamique (LMD-X), Ecole Polytechnique, France)

Earth observation satellites provide unprecedented information to monitor the different components of the water cycle, from soil moisture to river dynamics. However, fully exploiting these observations remains challenging due to sensor limitations, data heterogeneity, complex physical processes, and spatio-temporal resolution constraints. This seminar provides an overview of machine-learning approaches that accompany and enhance remote sensing for water cycle analysis. It illustrates how statistical and machine-learning techniques can improve the exploitation of Earth observation (EO) data at different processing levels, from Level 1 to Level 4. Four examples are presented: (i) compressing hyperspectral information to reduce observation dimensionality, (ii) improving the retrieval of soil moisture from space by exploiting spatial patterns and handling missing data, (iii) harmonizing multi-source EO at the global scale for consistent water cycle monitoring, and (iv) modeling river dynamics using data-driven approaches. Together, these examples highlight the potential of machine-learning techniques to better integrate observations and improve our understanding of hydrological processes.

会場: 計算科学研究棟

イベント公式言語: 英語

DA Seminar: Machine learning for precipitation estimation and forecasting / Analysis of a Long-Lived Supercell: Life Cycle and Severe Weather Patterns in Northern Buenos Aires Province

2025年11月13日(木) 10:30 - 12:00

フアン・ルイス (理化学研究所 数理創造研究センター (iTHEMS) 数理展開部門 予測科学研究チーム 客員研究員)
ルチアーノ・ヴィダル (理化学研究所 数理創造研究センター (iTHEMS) 数理展開部門 予測科学研究チーム 客員研究員)

Title: Machine learning for precipitation estimation and forecasting Speaker: Dr. Juan Ruiz (University of Buenos Aires – CONICET) Abstract: Estimating and forecasting precipitation is essential for a wide range of human activities as well as for disaster prevention. In this talk we will discuss the application of deep neural networks to the estimation of precipitation with high time and spatial resolution, combining remote sensors and numerical weather predictions. The proposed models show that these information sources can be effectively combined to improve the accuracy of real-time precipitation estimates. Additionally, we will present the application of deep neural networks as a postprocessing tool for short-range deterministic and ensemble-based numerical weather predictions and for the quantification of their uncertainty. The performance of the machine-learning models in the quantification of the uncertainty is close to that achieved by the dynamical ensembles and can be even better in the presence of a model. Title: Analysis of a Long-Lived Supercell: Life Cycle and Severe Weather Patterns in Northern Buenos Aires Province Speaker: Dr. Luciano Vidal (National Meteorological Service, Argentina) Abstract: This work presents a detailed analysis of a long-lived convective supercell that affected the northern Buenos Aires province, Argentina, on March 19, 2024. The primary objective is to characterize its life cycle and associated severe weather patterns using an integrated multi-sensor approach. This methodology combines data from satellite imagery with documentation of surface damage caused by large hail and intense winds. The storm exhibited a remarkable longevity, traveling approximately 400 km over 5.5 hours and impacting a total of 11 municipalities before its dissipation. Throughout its trajectory, the supercell generated significant damage due to large hail and severe wind gusts that, in some areas, exceeded 150 km/h. Furthermore, the storm ultimately affected the Sarandí-Santo Domingo basin (the pilot basin of the Argentine-Japanese SATREPS/PREVENIR project) by generating flash floods. The results of this analysis provide crucial information for the improvement of forecasting and early warning systems for severe weather events in the region.

会場: Hybrid Format (RIKEN R-CCS room 107 and Zoom)

イベント公式言語: 英語

Temporal Evolution of Crustal Stress at Volcanoes During Periods of Unrest

2025年10月14日(火) 10:30 - 12:00

Eric Newland (Research Fellow, Faculty of Mathematical & Physical Sciences, University College London, UK)

Eruptions that occur at volcanoes after periods of quiescence are difficult to forecast. Pathways that connect the source to the surface may have become sealed. The pressurisation of the source leads to the deformation of the crust. Initially the crust deforms elastically, strain is accommodated via ground movement and elastic strain energy is stored to the crust. Then, the deformation transitions to inelastic where strain is accommodated via brittle failure (volcano-tectonic event), and elastic strain energy is transferred from the crust. We present a novel method to estimate the temporal evolution of elastic strain energy and bulk stress during periods of unrest. We consider the transfer of energy using measurements of surface deformation and seismic activity. We evaluate the temporal evolution of crustal bulk stress and investigate the progression of deformation in the crust. We apply our method to the unrest at the Campi Flegrei caldera, Italy from 2011-2024, and the eruption of Sierra Negra, Galapagos, 2018. Our calculations reveal that the bulk stress follows a characteristic progression, in which the stress initially increases linearly with time prior to the onset of significant seismicity, consistent with elastic deformation. We then observe a transition to inelastic deformation, when rate of elastic strain energy lost via fracturing increases and eventually exceeds the rate of elastic strain energy transferred to the crust. This results in a decrease in the bulk stress stored in the crust with time, indicating a progressive weakening of the crustal material due to seismicity-induced damage. Comparison with laboratory experiments show the behaviour is consistent with bulk failure in extension and the potential formation of new pathways in the crust. Finally, we demonstrate how our method, along with the understanding of eruption precursors gained from the results, can be used to constrain deformation regimes at reawakening volcanoes after extended repose and to evaluate the hazard posed during periods of unrest.

会場: Hybrid Format (RIKEN R-CCS room 107 and Zoom)

イベント公式言語: 英語

Data Assimilation for the Vicsek model

2025年9月25日(木) 13:00 - 14:00

Tomoharu Takaki (東京大学 大学院情報理工学系研究科 修士課程)

会場: 計算科学研究棟 R311 (メイン会場) / via Zoom

イベント公式言語: 英語

Steps between the Lorenz96 models and the real world (TBD)

2025年9月19日(金) 13:00 - 14:00

雨宮 新 (理化学研究所 数理創造研究センター (iTHEMS) 数理展開部門 予測科学研究チーム 研究員)

会場: 計算科学研究棟 R511 (メイン会場) / via Zoom

イベント公式言語: 英語

Ensemble transform Kalman filter (ETKF) extensions for near-bound variables: Results from simple aerosol data assimilation experiments

2025年9月16日(火) 9:00 - 10:30

Jiang Richard Liang (慶応義塾大学 博士研究員)

Traditional data assimilation (DA) methods approximate the error distributions using Gaussian probability density functions (PDFs). However, the error distributions of some variables, such as clouds, precipitation, and aerosols, could be better approximated by gamma and inverse-gamma PDFs. For such bounded variables, the error standard deviation will likely increase with the distance of the unknown true value from its bound. To properly include these error distributions, a previous study by C. Bishop invented a method called the GIG filter, which is based on gamma and inverse-gamma distributions. We compared the performance of this new method and the traditional DA method with cycled DA experiments using a new tracer model based on the Lorenz-96 model. The GIG filter's performance is better for assimilating near-bound variables in our experiments.

会場: Hybrid Format (RIKEN R-CCS room 107 and Zoom)

イベント公式言語: 英語

Covariance Localization Local ensemble transform Kalman filter (LETKF)

2025年9月10日(水) 13:00 - 14:00

ウナシシュ・モンダル (理化学研究所 数理創造研究センター (iTHEMS) 数理展開部門 予測科学研究チーム 特別研究員)

会場: 計算科学研究棟 R511 (メイン会場) / via Zoom

イベント公式言語: 英語

Ensemble transform Kalman filter (ETKF)

2025年9月3日(水) 13:00 - 14:00

岩中 達郎 (理化学研究所 計算科学研究センター (R-CCS) データ同化研究チーム リサーチアソシエイト)

会場: 計算科学研究棟 R511 (メイン会場) / via Zoom

イベント公式言語: 英語

Perturbed Observation

2025年8月20日(水) 13:00 - 14:00

会場: 計算科学研究棟 R511 (メイン会場) / via Zoom

イベント公式言語: 英語

Extended Kalman filter: Lecture 2

2025年8月13日(水) 13:00 - 14:00

会場: 計算科学研究棟 (メイン会場) / via Zoom

イベント公式言語: 英語

DA Seminar: Prof. Dai Yamazaki and Hannah Cloke

2025年8月6日(水) 15:00 - 16:30

ダイ・ヤマザキ (東京大学 生産技術研究所)
Hannah Cloke (Professor, Department of Meteorology, University of Reading, UK)

The seminar will be jointly given by Associate Professor Dai Yamazaki (The university of Tokyo) and Professor Hannah Cloke (University of Reading). Speaker 1: Associate Professor Dai Yamazaki (Institute of Industrial Science, The University of Tokyo) Title: How can we achieve fast and realistic simulation of river and flood dynamics on the global scale? Abstract: Modeling river hydrodynamics across continental-scale basins is challenging due to their inherently multiscale nature. On one hand, we must account for the water budget along river systems that extend over 1,000 km. On the other hand, water movement within channels and floodplains is governed by topographic features smaller than 100 meters. The global river model CaMa-Flood addresses this complexity by employing the Catchment-based Macro-scale Floodplain modeling approach (CMF approach). This method approximates the relationship between water volume, flood extent, and water depth through sub-grid scale parameterizations. These parameters, derived from high-resolution satellite-based digital elevation models (DEMs) and hydrography datasets, enable realistic simulation of river discharge and flood stages—without explicitly resolving small-scale floodplain dynamics. To further accelerate simulations, recent developments in CaMa-Flood have introduced several performance optimizations, including MPI/OpenMP parallelization, SIMD vectorization, sparse matrix implementation, and a GPU-enabled Python version. These enhancements make the model more suitable for large-scale and near-real-time applications such as global flood monitoring and climate impact assessment. Speaker 2: Professor Hannah Cloke (Department of Meteorology, University of Reading) Title: Preparing for floods in an uncertain future

会場: Hybrid Format (RIKEN R-CCS room 107 and Zoom) (メイン会場) / via Zoom

イベント公式言語: 英語