2025 Spring Lectures in Climate Data Science
January 14, 2025 - May 22, 2025
- 12:00 – 1:30 pm (EDT)
- IN-PERSON at the Tang Family Hall (Rm 202) at the Columbia Engineering Innovation Hub (2276 12th Ave, New York, NY)
- VIRTUAL ATTENDANCE available
- Add to Calendar
Click to see past Lectures in Climate Data Science from Fall 2024, Summer 2024, Spring 2024, Fall 2023, Spring 2023, and Fall 2022.
TUESDAY || JANUARY 14, 2025
BERNARD MINOUNGOU
Center Régional AGRHYMET
“NextGen Approach to Hydrological Forecasting: Adapting PyCPT Tool for Hydrological Forecasting”
The NextGen forecasting system helps forecasters to evaluate the performance of different global climate models, which helps determine how best to correct and combine them. AGRHYMET Regional Climate Centre has been capacitating National Meteorological and Hydrological Services (NMHSs) in West Africa and the Sahel on NextGen seasonal forecasting systems. The capacity development efforts focus mainly on Python interface to the Climate Predictability Tool (CPT) or PyCPT, a tool developed by the International Research Institute for Climate and Society (IRI) to implement the NextGen approach to climate forecasting. While hydrological forecasts of water availability from watersheds in major river basins are essential to support operational planning and management, the latest version of PyCPT developed by IRI does not take into account the seasonal forecast of hydrological variables. To address this challenge and respond to the needs of NMHSs in the region in charge of hydrological monitoring, AGRHYMET has adapted and improving the PyCPT tool for seasonal hydrological forecasts. The primary hydrological data used to produce seasonal forecasts are the daily flows recorded at the main river stations in the region. These data are distributed across the main river basins. Forecasts are based on correlations between ocean surface temperature, precipitation, wind speed, and hydrological variables such as river flows for the high-water period. Principal component regression (PCR) is the most widely considered method. A variety of forecast performance scores consisting of scores based on continuous measurements, those on observed measurements, and, in some cases, forecasts are used as well. The work carried out by the AGRHYMET Regional Center, in response to the needs expressed by players in the sub-region, has resulted in the enhancement of PyCPT for hydrological applications. The current development has been tested through case studies in West Africa, and the results obtained are promising.
THURSDAY || JANUARY 23, 2025
SARA SHAMEKH
NYU
Watch on YouTube
LEAP Research Update: “Developing an Interpretable Model for Tropical Precipitation Using Machine Learning”
Precipitation remains one of the most challenging variables to model in both traditional climate models and modern data-driven approaches. Climate models often fail to accurately reproduce rainfall distributions, typically underestimating heavy precipitation events. Many uncertainties persist, including the precise dependence of precipitation on its environmental conditions—both thermodynamic and dynamic—and the role of small-scale structures like convective organization. In this study, we use observational data and machine learning to develop a transparent and interpretable model for tropical precipitation and precipitation area. Our findings indicate that precipitation can be accurately predicted when small-scale aspects of the domain are known, however, prediction accuracy significantly decreases when only large-scale conditions are considered, indicating the stochastic nature of precipitation. We then develop a stochastic model for rain area and connect large scale variables to to this model. Our model not only elucidates the dependency of precipitation on environmental conditions but can also be used to evaluate predicted precipitation from AI or convection schemes.
VIVIANA ACQUAVIVA
CUNY / Columbia University
Watch on YouTube
LEAP Research Update: “Tiny Ocean: Physically meaningful dimensionality reduction for pCO2 reconstruction”
Accurate reconstruction of the full pCO2 field is a crucial element in estimating air-sea carbon flux and ultimately understanding the magnitude and trend of the ocean carbon sink. Ocean data used to build machine learning models to reconstruct the full field of pCO2 rely on variables of different nature (spatial,
temporal, and physical features), and are distributed inhomogeneously in space and time. Additionally, the dimensionality of data is significant, with over a dozen features and hundreds of thousands of highly correlated data points. As a result, it is beneficial to seek physically meaningful ways to reduce the
dimensionality of data for easier manipulation. Clustering methods are promising, but they are often based on the Euclidean distance in the space of features, which is not a good tracer of similarity in the space of the target variable, in this case pCO2. To alleviate this problem, we build a new distance that follows the pCO2 more closely. We introduce the concept of information imbalance, which quantifies the asymmetry of information between different metrics, and use it to evaluate the information content of various candidate distances with respect to the target, the distance in pCO2 space, looking for the “sweet spot” between dimensionality reduction and preservation of information. This new distance can be used to perform physically meaningful clustering and reduce dimensionality along both axes (features and number of data points), to build more agile machine learning models, and to aid visualization and physical interpretation of data.
THURSDAY || JANUARY 30, 2025
HIKARU HAYAKAWA + JACKIE VANDERMEL
Climate Cardinals
Leveraging AI for Climate Justice: Partnerships Between Researchers and Civil Society
This presentation explores the critical role of partnerships between researchers and civil society in harnessing the power of artificial intelligence to advance climate justice. Drawing from real-world experience in science communication with the United Nations and collaborations with leading tech companies, we will highlight how AI can break down barriers and create equitable climate solutions. We will delve into the growing support for climate-justice-focused AI solutions and the challenges and barriers civil society organizations face in accessing and implementing these technologies. Using Climate Cardinals as a case study, we will showcase impactful partnerships with Google Cloud and Monday.com, illustrating how civil society can integrate AI to amplify its reach and efficiency in addressing climate issues. The session will conclude by exploring actionable opportunities for AI research institutions to collaborate with civil society organizations, paving the way for innovative, inclusive, and scalable climate justice solutions. This talk aims to inspire researchers and civil society leaders to co-create impactful strategies, driving meaningful change through technology and collaboration.
THURSDAY || FEBRUARY 6, 2025
TIAN ZHENG
Columbia University
Improving Distributional Learning for Dynamical Systems
Accurately modeling dynamic systems requires understanding the full distribution of outcomes, especially when accounting for uncertainties and biases. This talk explores the importance of distributional learning in dynamic systems, focusing on a recent project that uses Kernelized Stein Discrepancy (KSD) for post hoc calibration. Our approach leverages prior knowledge of the marginal distribution to correct prediction biases resulting from distribution shifts and sampling errors, enhancing model robustness. Through numerical experiments with climate models, we demonstrate the effectiveness of KSD-based calibration in refining predictions and improving adaptability to unseen data in complex, chaotic systems.
THURSDAY || FEBRUARY 13, 2025
AMY BRAVERMAN
NASA JPL / California Institute of Technology
Uncertainty Quantification for Remote Sensing
Remote sensing data provide a vast trove of information for studying the Earth system and climate. These data are often treated as if they are “truth”, but they are the results of complex inferential algorithms and are subject to uncertainty. In fact, the process of inferring the behavior of geophysical variables from noisy spectra observed from space, is itself an inverse problem. There is at present no universally agreed upon method for quantifying uncertainties in remote sensing data, and a wide variety of mostly ad hoc techniques are employed. Here I describe in general how these data sets are produced, the nature of the inverse problem that is embedded, and a probabilistic framework we have developed with an eye towards operational use. The latter demands that we address computational challenges we expect to face with NASA’s next generation of Earth observing satellites: the Earth System Observatory. Finally, I will describe an approach to “spatial” UQ that not only provides uncertainties on a footprint-by-footprint basis, but also preserves between-footprint uncertainty relationships.
THURSDAY || FEBRUARY 20, 2025
DION HO
Columbia University
Watch on YouTube
LEAP Research Update: “Data-Enhanced Two-Stream Methods for Computing Atmospheric Flux Transfer”
The two‐stream equations are the most parsimonious general models for radiative flux transfer with one equation to model each of upward and downward fluxes. This framework undergirds almost every weather and climate model worldwide for it enables blisteringly fast computations, and it is applicable to both longwave and shortwave radiation though our focus is on the latter. Current methods under the two-stream framework are, however, known to be inaccurate in many circumstances, for example, when modeling shortwave flux transfer through arctic aerosols. Consequently, we seek to use data to enhance two-stream methods while preserving their computational efficiency. We will present our new methods which involve spline interpolation, symbolic regression and neural networks as well as present both empirical and theoretical evidence for our improvements.
MATTHIEU BLANKE
NYU
Watch on YouTube
LEAP Research Update: “Interpretable Meta-Learning for Physical Systems”
Machine learning methods can be a valuable aid in the scientific process, but they need to face challenging settings where data come from inhomogeneous experimental conditions. Recent meta-learning methods have made significant progress in multi-task learning, but they rely on black-box neural networks, resulting in high computational costs and limited interpretability. We introduce a new meta-learning model capable of learning efficiently from multiple environments, with an affine structure with respect to the learning task. We prove that this architecture can identify the physical parameters of the system, enabling interpretable learning. We demonstrate the competitive generalization performance and the low computational cost of our method by comparing it to state-of-the-art algorithms on physical systems, ranging from toy models to complex, non-analytical systems. The interpretability of our method is illustrated with applications to parameter identification and to adaptive control and system identification.
THURSDAY || FEBRUARY 27, 2025
CHRISTINA TORRES
Teachers College – Columbia University
Watch on YouTube
LEAP Research Update: “Bridging educational divides: A research-practice partnership to empower teachers to integrate climate change across subjects”
No singular institution can address as large and complex a challenge as climate change education. Through an on-going research-practice partnership (RPP) between the Center for Sustainable Futures at Teachers College, Columbia University and New York City Public Schools (NYCPS) Office of Energy and Sustainability, we identified main barriers for teachers’ engagement with climate change education. These barriers include: misperception of climate as only a science topic, teacher knowledge gaps, curriculum resource gaps, and overall teacher preparedness. We sought to address these barriers by developing professional learning opportunities for NYCPS teachers aiming to develop teachers’ knowledge of climate change, situate climate change in a local NYC context, and build connections between the impacts of climate change to K-12 education. Climate scientists from the Learning the Earth with Artificial Intelligence and Physics (LEAP) center at Columbia University, curriculum development experts, and NGOs/CBOs were active in these programs. We offered these programs in two modalities. The summer institute was created as a week-long, intensive workshop with 40 teachers from the same school level (e.g., elementary) to develop transdisciplinary lesson plans on climate change over the school year. The mid-winter institute provided 500 teachers access to NYCPS curricular experts and local partners around climate change education during a three-day program. This paper will present findings from both cases and future directions for this RPP. We hope this will serve as a model for future partnerships around climate change education, inspiring strategic collaborations between K-12 education systems and higher education.
JOSEPH KO
Columbia University
Watch on YouTube
LEAP Research Update: “ML for Ice Microphysics: Challenges, Progress, and Opportunities”
Clouds are a leading order source of uncertainty in weather and climate models. Ice clouds are particularly hard to model due to diverse crystal habits and poorly understood microphysics. Machine learning (ML) has the potential to improve parameterizations, by incorporating diverse observations and reducing both parametric and structural errors. Here, we present an overview of recent work to improve representations of both ice microphysical properties and processes, using different ML approaches. First, we review recent work by Lamb et al. 2024, where neural ordinary differential equations (NODE) and symbolic regression (SR) were applied to laboratory measurements of single-crystal growth to develop a new functional parameterization for vapor depositional growth. Second, we revisit existing m-D relationships in ice microphysics schemes and present a bottom up framework to infer mass from in situ optical imagery, using ML models trained on synthetic data. Finally, we present a roadmap for implementation into next-generation ice microphysical schemes, as well as opportunities for novel applications of ML in ice microphysics.
THURSDAY || MARCH 6, 2025
GABRIELE ACCARINO
Columbia University
Watch on YouTube
LEAP Research Update: “New Similarity Metrics for Earth System Model Evaluation: A Multi-Scale Wavelet-based Approach”
Earth System Models (ESMs) simulate complex interactions within the Earth system, including the atmosphere, oceans, land, and cryosphere, to predict future climate conditions. These simulations generate vast, multi-dimensional data, that capture diverse physical fields such as temperature, precipitation, and radiative fluxes, across multiple spatial and temporal scales. Comparing ESM projections poses significant challenges particularly in the context of the Coupled Model Intercomparison Project Phase 6 (CMIP6), where projections are related to a wide range of scenarios and model realizations. To address this, the assessment of differences among ESM outputs (or between ESM outputs and data) can be reframed as the task of evaluating and quantifying perceptual similarities. Drawing on well-established concepts from image processing, we focus on similarities in relevant physical patterns and structural properties at the scales of interest. In this work, we introduce: (i) a set of candidate metrics that extend traditional pixel-based error measures, adapted to handle floating-point data typical of ESM data; (ii) a benchmark dataset and an evaluation pipeline to systematically assess the sensitivity of these metrics to controlled perturbations generated from a single reference map, and (iii) a prototype wavelet-based similarity metric tailored for climate data. This metric is highly flexible and tunable, designed to capture structural variations across multiple scales while remaining sensitive to both spatial patterns and magnitude differences in field values.
MAX KAGAN
Columbia Business School
LEAP Research Update: “Corporate Sociopolitical Alignment and Policymaker Relations”
A growing body of scholarship considers how firms’ sociopolitical activism or alignment may help or harm their performance. Much of this scholarship considers how firms’ association with controversial positions can polarize stakeholders’ attitudes. These prior studies have demonstrated the potential for sociopolitical alignment to attract and/or alienate consumers, employees, and investors, but relatively few studies consider the effect of stance-taking on government policymakers. This is despite a large literature suggesting that government relations are a key element of firms’ non-market strategy and that positive firm-government relations can yield myriad strategic benefits for companies. In this study, we take up this gap and ask whether firms are rewarded and/or punished by government policymakers for when they communicate sociopolitical stances.
THURSDAY || MARCH 13, 2025
MATTHIAS IHME
Stanford
Are we there yet? Towards foundational ML models through accessible community-involved big databases
Recent advances in data-driven methods and machine-learning techniques rely on large datasets and massive deep learning models to predict quantities of interest. Within the field of combustion, large amounts of data exist in the form of high-fidelity simulation data and detailed measurements that have been accumulating. Yet, this data remains distributed and can be difficult to access. To make this data available, this presentation presents a realistic and feasible framework which combines (i) community involvement, (ii) public data repositories, and (iii) lossy compression algorithms for enabling broad access to high-fidelity data via a network-of-datasets approach. This Bearable Large Accessible Scientific Training Network-of-Datasets (BLASTNet) is consolidated on a community-hosted web-platform (at https://blastnet.github.io/), and is targeted towards improving accessibility to diverse scientific data for deep learning algorithms. We discuss procedures for community to contribute and utilize this database and present illustrative examples on how this database has contributed towards the development of new ML models in the context of turbulent subgrid-scale modeling.
THURSDAY || MARCH 27, 2025
HUGH MORRISON
NSF NCAR
Toward an Improved Representation of Warm Cloud Microphysics in Earth System Models
The formation of drizzle and rain via drop collision-coalescence (“warm rain” initiation) is a key component of Earth system models. This is typically represented using microphysics schemes that only predict a few variables to describe the drop population. When developing these relatively simple “bulk” schemes, choices must be made regarding which cloud and rain variables to predict and how to calculate the micro-scale physical processes affecting drops. These structural choices need to be simple enough so that schemes can run efficiently in three-dimensional models. This talk will describe an investigation of the accuracy of bulk schemes using different sets of predicted variables and different methods, including machine learning, for representing drop collision-coalescence. The goal is to assess how accurately these schemes simulate the formation of rain drops starting from an initial population of (small) cloud droplets, relative to a detailed reference microphysical model. Using predicted variables that represent the properties of drops of all sizes – the single category approach — leads to much greater accuracy than the traditional two-category approach with separate variables for cloud (small drops) and rain (large drops), despite both approaches utilizing the same number of predicted variables (4). A sub-optimal choice of the predicted variables, as in the two-category approach, limits overall accuracy even when the coalescence rate calculation itself is highly accurate. This places an upper limit on accuracy using machine learned rates even when the rate calculation is perfectly emulated. On the other hand, using a neural network to obtain coalescence rates with the single category approach gives accurate results for a reasonable computational cost. Furthermore, only 3 predicted variables are needed for relatively accurate solutions (at least as accurate as the best-performing two-category schemes with 4 predicted variables). This result is consistent with Lamb et al. (2024, JAMES), who used a variational auto-encoder to show that the number of intrinsic dimensions for the problem of warm rain initiation is roughly 3. Prospects for utilizing the single category approach in Earth system models will be discussed, including challenges owing to the relatively coarse spatial resolution and long time step (~5 min) of these models. Finally, I will highlight recent developments of the Lagrangian “super-droplet” method for modeling microphysics, which tracks representative cloud and rain drops in the modeled flow, and how this method can be used to further improve bulk single category schemes. This includes accounting for turbulent enhancement of drop collision-coalescence, which may be critical for rain initiation in some cloud regimes.
THURSDAY || APRIL 3, 2025
OREOLUWA BADAKI
Teachers College – Columbia University
Bridging Gaps Between Knowledge and Action in Climate Communications
More knowledge about climate change does not automatically result in more (or better) action to address climate change. In fact, more knowledge about climate change can sometimes lead to increased fear, hopelessness, anxiety, lethargy, and polarization, especially for young people (Doyle, 2019; Kahan 2012, 2014, 2015; Ojala, 2012 Majeed et al., 2017; Wu et al. 2020). In order to bridge the gap between knowledge and constructive action, therefore, climate communication and education needs to do more than give people high quality information. It needs to open up pathways for embodying and activating this information. Best practices in climate communications include co-creating narratives with audiences, harnessing emotion, and engaging media (Boycoff, 2019; Peters and Salas 2022; Peters et al., 2022 ). This talk will highlight methods for engaging climate data within climate communications.
Drawing from four ongoing ethnographic projects conducted with youth and communities in Philadelphia and Rio de Janeiro, Dr. Badaki will highlight ways in which climate data can be used in creative and multimodal ways to engage and mobilize communities for climate action. The talk will focus specifically on the work of the Body-Ecology Lab (BEL), which draws from ethnographic, performance, and multimodal methods to address the question: what stories can our bodies remember, as well as reimagine, about our relationship with the land? During the talk, Dr. Badaki will share examples from the lab’s emerging work centering four communicative methods: (1) Climate storytelling (2) Performance and exhibition (3) Community mapping and (4) AI chatbots.
THURSDAY || APRIL 17, 2025
LINNIA HAWKINS
Columbia University / NCAR
Watch on YouTube
LEAP Research Update: Leveraging ML to Systematically Constrain Parametric Uncertainty in the Community Land Model
Land models simulate complex physical and biological processes, many of which are highly parameterized and thus subject to substantial parametric uncertainty. Perturbed parameter ensembles (PPEs) are valuable tools for assessing sensitivity, quantifying uncertainty, and guiding model calibration. This work demonstrates history matching–based parameter estimation methods that incorporate PFT-specific parameters and their interactions. The approach constrains parametric uncertainty while preserving plausible model variants, yielding ensembles that inform uncertainty quantification in global carbon cycling and climate projections.
JIARONG WU
NYU
Watch on YouTube
LEAP Research Update: Data-Driven Probabilistic Air-Sea Flux Parameterization
Accurately quantifying air-sea fluxes is important for understanding air-sea interactions and improving coupled weather and climate systems. This study introduces a probabilistic framework to represent the highly variable nature of air-sea fluxes, which is missing in deterministic bulk algorithms. Assuming Gaussian distributions conditioned on the input variables, we use artificial neural networks and eddy-covariance measurement data to estimate the mean and variance by minimizing negative log-likelihood loss. The trained neural networks provide alternative mean flux estimates to existing bulk algorithms, and quantify the uncertainty around the mean estimates. Stochastic parameterization of air-sea turbulent fluxes can be constructed by sampling from the predicted distributions. Tests in a single-column forced upper-ocean model suggest that changes in flux algorithms influence sea surface temperature and mixed layer depth seasonally. The ensemble spread in stochastic runs is most pronounced during spring restratification. (Paper preprint https://www.arxiv.
THURSDAY || APRIL 24, 2025
AXEL SEIFERT
Deutscher Wetterdienst
Machine Learning of Cloud Representations for Weather and Climate Models
Clouds are a major challenge for weather and climate models. The deficiencies in current parameterizations of cloud processes contribute to model uncertainty and forecasts failures. Machine learning methods hold the promise to overcomes these deficiencies and accelerate the model development in geosciences. This promise is twofold: First, machine learning methods can help us to find efficient and simplified approximations of the complicated nonlinear equations that govern the formation of clouds and precipitation. This can be done, for example, through a statistical emulations of reference simulations from detailed physically-based numerical models, which would be too expensive for weather and climate applications. Second, machine learning methods can assist us in the calibration and tuning of weather and climate models using observations. Current weather and climate models do need calibration and tuning because the representation of the physical processes on a given numerical grid is uncertain and inaccurate. Calibration and tuning are currently part of the model development process and remain tedious and to some extent subjective. Machine learning methods can in principle do this better and more objectively, but we have to overcome some obstacles to achieve this goal. In my presentation I will report on some efforts at the Deutscher Wetterdienst, which is the National Weather Service of Germany, to use machine learning methods in the development, calibration and tuning of the ICON model. The first part of the talk will focus on the parameterization of cloud microphysical processes based on detailed super-particle simulations. In the second part, I will report on an attempt to implement an online learning method to improve the cloud scheme of ICON based on satellite data.
THURSDAY || MAY 1, 2025
KAITLYN LOFTUS
Columbia University
Watch on YouTube
LEAP Research Update: “Parameterizing cloud microphysics with machine learning-enabled Bayesian parameter inference”
Warm microphysics parameterizations (simplified representations of the properties and evolution of liquid cloud and precipitation particles) are an important—but highly uncertain—component of atmospheric modeling from cloud-resolving to global scales. Here, we use machine learning (ML) to accelerate Bayesian and process-based workflows with the goal of characterizing and reducing errors associated with warm microphysics parameterizations. Specifically, we consider errors driven by parameter selection (parametric error) and errors driven by the choice of prognostic variables and process rate formulations (structural error), as these two error types require separate paths for characterization and reduction. We focus on the performance of a standard (two-moment, two-category) bulk microphysics scheme in large-eddy simulations and use as a target emergent cloud and precipitation properties simulated with a more complex bin microphysics scheme. We use ML-enabled Bayesian parameter inference to perform quantitative parameter selection, to separate parametric and structural error, and to characterize parametric error quantitatively and structural error qualitatively. We improve parameterization performance by implementing process-level structural changes driven by insights from the Bayesian analysis. This work highlights how ML (in concert with a process-level perspective) can enable more systematic parameterization development, evaluation, and uncertainty quantification.
AYA LAHLOU
Columbia University
LEAP Research Update: “Bringing Machine Learning Phenology into the CLM”
This talk will present ongoing efforts to replace empirical phenology thresholds in the Community Land Model (CLM) with a machine learning-based phenology model. In Phase 1, we are focused on integrating the Transformer-based model into CLM using the FTorch bridge framework to predict the onset of growth in deciduous vegetation based on historical climate and vegetation data (GIMMS LAI, CRUJRA). FTorch enables seamless coupling between PyTorch models and Fortran-based Earth system models, and is central to our integration strategy. I will walk through the experimental design, site-level simulation setup, how we use the FTorch bridge to test and connect ML predictions with CLM processes, and future work.
THURSDAY || MAY 8, 2025
JAEYOUNG JUNG
Columbia University
Watch on YouTube
LEAP Research Update: “A Multi-Scaling Framework of Airflow-Canopy Interaction”
Wind behaviors within vegetation canopies critically regulate the exchange of mass, momentum, and energy between the land surface and atmosphere, thereby influencing microclimatic conditions across a wide range of scales. These dynamics govern temperature and humidity control, gas exchange, nutrient transport, and numerous biological and ecological mechanisms. However, the inherent multiscale heterogeneity in natural systems poses substantial challenges to accurately modeling canopy flows. Atmospheric models, particularly those at the global climate model (GCM) scale, typically rely on one-dimensional (1D) parameterizations—such as Monin-Obukhov similarity theory and 1D analytic profiles—that assume isotropy and spatial homogeneity in canopy distributions. Similarly, even in high-resolution modeling based on the large-eddy simulation at the microclimate scale, vegetation structure is often simplified through locally homogeneous assumptions. While such simplifications may be acceptable in regions far from canopies (i.e., far-field modeling), they become unreliable in the vicinity of canopies where most mass and energy exchanges occur. Moreover, when the scale separation is weak, the deformation (stretching and compression) of flows induced by micro-scale heterogeneity can significantly affect macroscopic phenomena. Although progress has been limited by model complexity and data constraints, understanding the influence of heterogeneity on canopy flows remains a critical direction for future research. To address this challenge, the present study introduces a multiscale modeling framework capable of resolving canopy flows across heterogeneous canopy distributions spanning multiple scales. Through a systematic multi-step upscaling—from leaf scale to GCM scale—the model comprises three sub-models corresponding to micro-, meso-, and macro-scales. Upscaling is rigorously conducted using volume-averaging theory, and machine learning techniques are employed at each scale transition to incorporate sub-grid-scale information into the coarser-scale models. In particular, the mesoscale model, which bridges the micro- and macroscale processes, leverages the concept of porosity to explicitly distinguish between sub-grid and resolvable scales. This design minimizes contamination of parameterizations arising from the scale mixing. To ensure stable and accurate solutions even under highly heterogeneous or discontinuous porosity distributions, a robust discontinuity-resolving solver is developed. The numerical scheme is formulated in an explicit form without iterative procedures, enabling future extension to differentiable models for machine learning-based applications. It is hoped that this study will ultimately contribute to improved parameterizations of atmosphere–vegetation interactions in GCMs.
YONGQUAN QU
Columbia University
Watch on YouTube
LEAP Research Update: Principled Integration of Variational Data Assimilation and Generative Models
Earth system modeling presents a fundamental challenge in scientific computing: representing multiscale, nonlinear dynamics with computationally feasible models introduces systematic forecast errors. Data assimilation (DA) mitigates these errors by optimally blending noisy observations with prior forecasts, but conventional variational methods often assume Gaussian error statistics that fail to capture the true, non-Gaussian behavior of chaotic systems. We propose a principled DA algorithm that alternates a lightweight, gradient-based analysis update on incoming observations with a pretrained flow-matching generative prior conditioned on the background forecast via a conditional Wasserstein coupling, under a Plug-and-Play (PnP) framework. This strategy relaxes restrictive statistical assumptions and leverages rich historical data without requiring an explicit regularization functional, and it also avoids the need to backpropagate gradients through the complex neural network that encodes the prior during assimilation cycles. The experiments illustrate improved forecast accuracy, with consistent gains across varying observation sparsities and noise levels, compared to standard variational DA approaches.
THURSDAY || MAY 15, 2025
N’Guessan Césaire Kouadio
Universidade Técnica do Atlântico (Cape Verde)
Empowering Climate Research with LEAP Pangeo: Insights from Crop Yield Prediction and Opportunities for Improvements
In this presentation, I will share some of my experiences with using the LEAP Pangeo platform in support of climate research around promoting agricultural resilience. I will also detail how LEAP Pangeo facilitated the design of a hybrid deep learning model (Convolutional neural network_CNN-Gatted Recurrent Unit-GRU) for predicting rice yields in India, considering different climate scenarios using CMIP6 and AgMIP datasets. This deep learning model gives indications on future warming impacts, mostly under SSP585, on food security. I will highlight the benefits that suit an early-career scientist from West Africa and global south using cloud tools provided by LEAP Pangeo to access and analyse very large climate datasets. Furthermore, I will offer constructive suggestions on further improving the platform to better serve researchers from the Global South as well as multidisciplinary applications in climate science. My goal is to illustrate that platforms like LEAP Pangeo operationally empower a new breed of scientists working for climate solutions.
THURSDAY || MAY 22, 2025
THIS LECTURE HAS BEEN POSTPONED TO A LATER DATE
DHRUV BALWADA
Columbia University