2024 Fall Lectures in Climate Data Science
September 5, 2024 - December 19, 2024
- 12:00 – 1:30 pm (EDT)
- IN-PERSON at the Tang Family Hall (Rm 202) at the Innovation Hub (2276 12th Ave, New York, NY)
- VIRTUAL ATTENDANCE available
- Add to Calendar
Click to see past Lectures in Climate Data Science from Summer 2024, Spring 2024, Fall 2023, Spring 2023, and Fall 2022.
THURSDAY || SEPTEMBER 5, 2024
JULIA SIMPSON
Columbia University
LEAP Research Update: “Improving parameterizations of heat & momentum air-sea fluxes”
The ocean and atmosphere influence each other in elaborate exchanges at the air-sea nexus. These fluxes are transfers of properties such as heat, moisture, mass, and radiation. Atmospheric winds also exert stress–momentum flux–on the ocean surface. Momentum and heat fluxes from the atmosphere drive ocean circulation; on longer timescales, the ocean acts as a reservoir of heat and greenhouse gases, holding tremendous influence over weather and climate patterns. However, the limited scope of observations leads to inaccuracies in global estimates of air-sea fluxes, which then affect the prediction of ocean circulation and the subsequent storage of heat and mass in the deeper ocean. Anthropogenic emissions have caused imbalances in this natural exchange and storage of properties, highlighting the need for accurate representation of air-sea fluxes in climate models. While advances in remote sensing and further deployment of in-situ measurements are enhancing observational capabilities, the closure problem persists, limiting the ability to calculate fluxes purely from physics-based equations. Therefore, empirical parameterizations are conventionally used to calculate air-sea fluxes. These “bulk” parameterizations are based on extremely limited in-situ measurements from relatively homogeneous conditions; thus, though they are applied virtually universally, conventional bulk parameterizations are less accurate under varying conditions. Machine learning (ML)-based solutions to this problem are at an exciting but early stage, with current ML-based parameterizations trained solely on bulk parameterizations of fluxes. This work aims to utilize existing parameterizations and global satellite data for pre-training to achieve robust representation, with transfer learning on in-situ observations to further refine estimates. This approach thereby leverages existing empirical knowledge, growing remote sensing capabilities, and novel ML techniques to expand the applicability and accuracy of air-sea flux models.
THURSDAY || SEPTEMBER 12, 2024
ANDRE PERKINS
Allen Institute for AI
ACEv2 and the new frontier of AI-based emulators for climate simulation
Climate simulation has been an invaluable tool for projecting future climate changes due to human activity, while also serving as our best laboratory for studying the interconnected Earth System. These large and complex models are primarily limited to direct use by organizations with access to large computing facilities due to their demanding computational requirements. However, as in many disciplines, recent advances in AI-based predictive technology are rapidly changing this landscape. The weather forecasting community is one early leader in this transition for atmospheric science, where multiple recent AI-based models now offer competitive or even improved 1-10 day atmospheric forecasts compared to their physics-based predecessors, while requiring only a fraction of the compute power (e.g., a desktop workstation). A similar shift is already on the horizon for climate simulation, with analogous techniques showing promise for large speedups and improved performance at coarse horizontal grid resolutions.
In this talk, I will discuss the latest research from the Ai2 Climate Modeling team in partnership with GFDL, LLNL, and NVIDIA on the AI2 Climate Emulator (ACE), a machine-learning-based emulator of a physics-based global atmosphere model. As an emulator, ACE maintains realistic weather variability throughout decades-long rollouts with prescribed sea surface temperatures and sea ice, provides an accurate representation of the source model’s climate and the atmospheric response to forcing in AMIP scenarios, and is over 60x faster than the traditional model it emulates. We’ve also made progress toward coupling ACE to a more realistic ocean, with initial results from a slab ocean model. We believe this technology provides a viable pathway toward full model emulation of an Earth System, serving as a valuable part of the toolkit for climate science. However, since we are in the early stages, I will also touch upon some of the many remaining technical and ideological challenges to overcome.
THURSDAY || SEPTEMBER 19, 2024
BRIAN MEDEIROS
NCAR
Hops, Skips, and Jumps: development activities around CAM7
The development cycle of the Community Earth System Model (CESM) and its atmospheric component, the Community Atmosphere Model (CAM) is long and tortuous. This lecture presents changes being introduced for CAM version 7 and describes the processes involved in developing CAM, including some of the challenges and compromises. Supporting model development, as well as users and their applications, creates a small ecosystem of tools, documentation, and communication channels. Some of these will be highlighted, including development of updated diagnostics software and tools that facilitate novel numerical experiments. Ongoing projects that push the boundaries of CESM and CAM, including those from LEAP, will be described to start sketching what future versions might look like.
THURSDAY || SEPTEMBER 26, 2024
** This event is part of Climate Week at Columbia Engineering. **
CLIMATE CHANGE AND FLOOD RISK
LEAP presents this interactive discussion to explore the flood risks incurred by climate change, how modeling can support communities’ efforts to adapt to climate change, and how flood risk models are being improved.
During the workshop, participants will:
1. Learn about climate data and modeling, with a focus on flood risks;
2. Have the opportunity to ask questions to Columbia University flood risk experts;
3. Contribute ideas for how climate models can address their needs.
Panelists:
- George Deodatis (Vice Dean of Research for Columbia Engineering, Santiago and Robertina Calatrava Family Professor of Civil Engineering and Professor of Earth and Environmental Engineering, Columbia University)
- Candace Agonafir (Postdoctoral Research Scientist, Columbia Water Center)
- Fatoumata Camara (SP^2 Science Pathways Scholar, Barnard College)
- Adam Nayak (PhD Student, Earth and Environmental Engineering, Columbia University)
Facilitator:
- Savannah Thais (Associate Research Scientist, Data Science Institute, Columbia University)
THURSDAY || OCTOBER 3, 2024
MAGGIE POWELL
Columbia University
LEAP Research Update: “Leveraging Subgrid-Scale Spatial Organization and Variability for Improved Cloud Fraction Parametrization Using PINACLES Simulations”
Given the coarse resolution of Earth System Models (on the order of 100 km), cloud processes must be parametrized. As a consequence, cloud processes are a major source of uncertainty in climate projections, with particular intermodel disagreement in the representation of low marine boundary layer clouds. Cloud fraction parametrizations have long relied upon subgrid-scale information of the total water distribution using various probability density functional forms. However, the importance of subgrid-scale spatial organization, including coherent updraft and downdraft, for cloud fraction has not been fully examined. In this study, we use machine learning to implicitly learn the subgrid-scale organization of shallow clouds and assess the information gained from including these subgrid features in a neural network-based parametrization. For this work, we use data from the Predicting INteractions of Aerosol and Clouds in Large Eddy Simulations (PINACLES) model for domains in the Eastern North Atlantic, Northeastern Pacific, Southern Great Plains, and Southern Ocean. Particular focus is given to the varying importance of subgrid-scale variability across shallow cloud regimes.
THURSDAY || OCTOBER 10, 2024
BIN YU
UC Berkeley
Interpreting Deep Learning Models
Wavelet-like filters arise when sparse dictionary learning optimization is fed with natural image patches, and they correspond to the V1 primary visual cortex of the human brain. Moreover, wavelets as local Fourier Transforms are interpretable in physical sciences and beyond. In this talk, we will first describe adaptive wavelet distillation (AWD) to turn black-box deep learning models interpretable in cosmology, cellular biology, and climate science problems while improving predictive performance. Time permitting, we will briefly describe deep learning results about feature learning in simple settings and improving LoRA for fine-tuning deep learning models through LoRA+.
THURSDAY || OCTOBER 24, 2024
SHAWN LI
Columbia University / Data Science Institute
LEAP Research Update: “A Distribution-Driven Probabilistic Data Assimilation Framework for Climate Science”
Traditional data assimilation techniques in weather forecasting focus on matching model parameters to observed state trajectories, such as daily temperature series. However, for climate modeling, capturing the long-term distributions of state variables—like the frequency and range of temperatures—is crucial. This talk introduces a new probabilistic framework designed to address this need by emphasizing probability density functions (PDFs) of state variables, using the Lorenz’96 system as a toy model. We have developed a PDF emulator utilizing conditional normalizing flow models, which are adept at transforming simple initial distributions into complex data distributions through a series of invertible mappings. This emulator efficiently replicates the state distributions of climate variables without simulating every possible state trajectory, offering a significant computational advantage. Building on this foundation, our framework includes a sophisticated, distribution-driven approach to model parameter inference. This method aligns the emulated PDFs with observed data distributions and incorporates uncertainty quantification, enhancing model reliability and applicability in various climatic conditions.
The framework’s potential extends to applications such as extreme value estimation and joint quantile analysis, which are pivotal for understanding and predicting climate extremes. By focusing on the tails and joint distributions of state variables, our approach promises improvements over traditional data assimilation methods, presenting a new tool for advanced climate modeling and analysis.
THURSDAY || OCTOBER 31, 2024
ROBERT PINCUS
Columbia University
Shaping the Future of Data, Computation, and Collaboration at LEAP
This week’s Lecture in Data Science isn’t a lecture at all but a discussion facilitated by Tian Zheng (LEAP Deputy Director) and Robert Pincus about the evolving needs of LEAP’s broad community. We would especially like to identify possible areas where we can share technical expertise – for example, around the training of ML models or practices for reproducibility. Based on your input, the LEAP Center will try to nucleate short- or long-term communities of practice around particular tasks and, following on the examples of ClimSim and ClimSim-Online, focusing effort on specific broadly-useful tasks. We hope for lots of input from the research community but especially from those involved in education and knowledge transfer (writ broadly).
THURSDAY || NOVEMBER 7, 2024
THEA HEIMDAL
Columbia University
ABBY SHAUM
Columbia University
LEAP Research Update: “Using CMIP6 models to better understand the ocean carbon sink”
In order to understand climate impacts from rising carbon emissions, it is critical to accurately quantify the air-sea CO 2 exchange and the ocean carbon uptake (i.e., ocean sink) in time and space. Measurements of ocean carbon are sparse because of the vast size of the ocean and limited resources. It is possible to reconstruct ocean carbon values where measurements do not exist by using Machine Learning (ML). However, current estimates of ocean carbon uptake are uncertain and show a large spread.
Earth system models (ESMs) can be used as ‘testbeds’ to better understand the uncertainties in the ocean sink estimates. The target (ocean carbon) and driver variables (e.g., ocean temperature) needed for the ML reconstruction are sampled from the testbed in a time and space pattern identical to real-world measurements. The benefit of a testbed comes from the fact that the full reconstruction can be compared against a correct solution – the full testbed model field – and not just a test set (typically 20% of available data). Furthermore, a testbed can be used to explore various ocean carbon sampling strategies without actually doing the measurements, which would save time and money.
We have used the Pangeo-ESGF CMIP6 Zarr Data 2 (Busecke and Stern, 2024) as a testbed to reconstruct ocean carbon globally over the period 1982-2022. Our work represents a case study on how to utilize the CMIP6 catalogue available on the LEAP Pangeo platform. We will show how current and future sampling, as well as algorithmic design choices, can impact ML reconstructions and ocean sink estimates.
THURSDAY || NOVEMBER 14, 2024
JANNI YUVAL
Google Research
Neural general circulation models for weather and climate
Recent advancements in machine learning (ML) have led to various data-driven approaches aimed at enhancing weather prediction and climate modeling. Typically, an ML-only method is used to improve weather prediction. While current state-of-the-art ML approaches achieve lower errors at medium-range lead times, physics-based models demonstrate superior physical consistency and better forecast accuracy
at longer lead times. In the realm of climate modeling, ML is often used in a hybrid approach, where ML
components replace uncertain parameterizations while still adhering to the governing equations for the
scales they resolve. Despite advancements in atmospheric hybrid models, current attempts face challenges
such as instability over extended periods, limited applicability to idealized scenarios, and only modest
improvements on longer time scales in realistic scenarios.
In this talk, I will explore the effectiveness of a novel approach to atmospheric modeling suitable for both
weather forecasting and climate modeling. This approach involves the development of a differentiable
atmospheric model, named NeuralGCM, which for the first time combines ML techniques with governing
equations in an end-to-end training using ERA5 data. NeuralGCM competes with ML models for 1-10
day forecasts and matches the European Centre for Medium-Range Weather Forecasts ensemble
prediction for 1-15 day forecasts. Over longer time scales, NeuralGCM displays emergent phenomena
such as seasonal cycles, monsoon patterns, and tropical cyclone formation, while achieving comparable
spatial bias to a global cloud-resolving model. Additionally, we present the first successful simulation of
an AMIP-like experiment using a hybrid atmospheric model.
THURSDAY || NOVEMBER 21, 2024
ARVIND RENGANATHAN
University of Minnesota
LEAP Research Update: “Advancing Environmental Modeling with Knowledge-Guided Encoder-Decoder Frameworks”
Recent advancements in machine learning (ML) have led to various data-driven approaches for environmental system modeling. These approaches typically implement direct mapping from drivers to response instead of leveraging more suitable encoder-decoder architectures.
In this talk, I will explore how the encoder-decoder paradigm provides additional key capabilities for environmental modeling while providing avenues for additional knowledge guidance. The paradigm naturally handles context while enabling the processing of varied information types, from entity characteristics to historical observations and meteorological drivers. The sequence-to-sequence design separates context processing from query handling, enabling flexible query integration. Adding to these capabilities, the encoder allows the modeling of multi-scale processes and data assimilation more naturally while supporting uncertainty quantification for prediction outputs.
Building on these advantages, the framework enables the transfer of knowledge across entities through a systematic approach. The encoder processes contextual information and learns transferable representations from data-rich entities, while the decoder adapts these representations for predictions in data-sparse scenarios. This design enables few-shot learning for entities with limited observations and zero-shot prediction for unobserved entities. Further, the ability to handle varied context information makes this architecture suitable for foundation model development in environmental sciences.
To demonstrate these capabilities in practice, we implement this paradigm across multiple environmental applications: GPP prediction in sparsely observed locations, operational streamflow forecasting, uncertainty quantification for hydrological modeling, and time series forecasting. Our results validate the encoder-decoder architecture’s effectiveness for environmental ML systems, showing the advantages of context integration through an encoder-decoder approach.
THURSDAY || DECEMBER 5, 2024
TIM HERMANS
Utrecht University
Investigating the Use of Neural Networks to Project Changes in Extreme Storm Surges in Europe
Projections of atmospherically driven changes in extreme storm surges are often based on small ensembles of climate model simulations because they are computed with computationally expensive hydrodynamic models. Consequently, these projections are sensitive to internal climate variability and inter-model differences. Data-driven storm surge models are emerging as a promising, computationally cheaper alternative to translate simulated changes in atmospheric variables to changes in storm surges. However, the ability of data-driven models to predict extreme storm surges in Europe is unclear because previous studies did not address the underrepresentation of extremes in the training data. Therefore, we investigated how well neural networks can reconstruct extreme storm surges when addressing the imbalance in the training data using the cost-sensitive learning approach in which samples with a lower density contribute to the training loss more.
We find that density-based weighting improves both the RMSE and F1 score of predictions v.s. observed extremes, although the optimal weights depend on the location. Our neural networks typically outperform an existing multi-linear regression model, but perform less well than a state-of-the-art hydrodynamic model (GTSM). Using a ConvLSTM- instead of LSTM layer reduces both the underestimation of the highest extremes and the number of false positive predictions, improving the error metrics of the neural networks at almost all selected locations. In light of the potential application of neural networks to global climate model simulations to project changes in extreme storm surges, I will also discuss preliminary results of the application of our neural networks to atmospheric forcing from a global climate model participating in HighResMIP.
THURSDAY || DECEMBER 5, 2024
VERONIKA EYRING
DLR / University of Bremen
Understanding and Modeling the Earth System with Machine Learning
Earth System Models (ESMs) are fundamental to understanding and projecting climate change. The models
have continued to improve over the years, but systematic errors and large uncertainties in their projections
remain. A large contribution to this uncertainty stems from the representation of processes such as clouds and convection that occur at scales smaller than the resolved model grid. This impacts the models’ ability to accurately project global and regional climate change, climate variability, and extremes. New approaches are required with breakthroughs expected in particular from the combination of high-resolution simulations that can resolve small-scale and fast processes, the wealth of Earth observations, and machine learning (ML) techniques. High-resolution, cloud resolving models with horizontal grid resolution of a few kilometers
alleviate many biases of coarse-resolution models for deep clouds and convection, wave propagation and
precipitation, but they cannot be run at climate timescales for multiple decades or longer due to high
computational costs. Yet short simulations from high-resolution models can serve as information to develop
ML-based parametrisations that are then incorporated into hybrid ESMs that promise to have significantly reduced systematic errors and enhanced projection capability compared to current ESMs. In contrast to km-scale climate models, ESMs incorporate important Earth system processes and feedbacks while still being fast enough to provide large ensembles important to simulate internal variability and extremes and to improve attribution and understanding. This combination can drive a paradigm shift in current Earth system modelling and analyses towards a new data-driven, yet still physics-aware, science. The key goal is a hybrid modelling approach that maintains physical consistency and realistically extrapolates to unseen climate regimes while reducing climate projection uncertainties and improving Earth system understanding. The talk presents progress in hybrid modelling work with the ICOsahedral Non-hydrostatic (ICON) atmospheric model from the European Research Council (ERC) Synergy Grant on “Understanding and Modelling the Earth System with Machine Learning (USMILE)” as well as key challenges and visions on AI-empowered next-generation multiscale climate modeling for mitigation and adaptation.
THURSDAY || DECEMBER 19, 2024
JOSEPH LOCKWOOD
Columbia University / Data Science Institute
LEAP Research Update: “Modeling Tropical Cyclone Wind Fields and Deep Convection Drivers with Generative AI and Explainable Models”
This presentation focuses on two projects: (1) generative super-resolution for downscaling tropical cyclone (TC) wind fields, and (2) explainable AI to refine integrated moisture convergence indices for deep convection cloud regimes and identify key environmental drivers.
The first project introduces a deep learning framework that downscales TC wind fields from ERA5 data (0.25°) to high-resolution observations (0.05°). The framework includes a debiasing neural network and a conditional denoising diffusion probabilistic model (DDPM) for generative super-resolution. This approach captures fine-scale spatial details, enabling accurate high-resolution wind field prediction from coarse data.
In the second project, we refine the environmental indices governing integrated moisture convergence for deep convection cloud regimes. Existing indices such as saturation fraction and moist convective instability are often arbitrarily defined. Using a variational autoencoder (VAE), we uncover latent representations of deep convection cloud regimes, and symbolic regression identifies key environmental variables controlling moisture convergence. Preliminary results demonstrate the value of explainable AI in improving predictions of deep convection cloud processes.
MOHAMMADREZA MOUSAVI KALAN
Columbia University
LEAP Research Update: “Transfer Neyman-Pearson Algorithm for Outlier Detection”
We consider the problem of transfer learning in outlier detection where target abnormal data is rare. While transfer learning has been considered extensively in traditional balanced classification, the problem of transfer in outlier detection and more generally in imbalanced classification settings has received less attention.
We propose a general algorithmic approach which is shown theoretically to yield strong guarantees with respect to a range of changes in abnormal distribution, and at the same time amenable to practical implementation. To validate our approach, we apply the proposed algorithmic procedure to climate data for heavy rain detection, demonstrating that it outperforms several baseline methods.