ICTS

Earth System Models (ESMs) are important tools for understanding Earth’s climate variability and change arising from the interactions among the different Earth system components (viz., atmosphere, ocean, land, cryosphere, biosphere, including human activities). The Centre for Climate Change Research (CCCR) at the Indian Institute of Tropical Meteorology (IITM), Pune, focuses on the development of the IITM Earth System Model (IITM-ESM) to address science of climate change, including detection, attribution and future projections of global and regional climate and also create climate modelling capabilities in India. Long term (multi-century) simulations of the preindustrial and present-day climate and future projections were performed using the IITM-ESM, to assess climate variability and change with special focus on the South Asian monsoon (Swapna et al., 2018) . The IITM-ESM participated in the Coupled Model Inter-comparison Project Phase-6 (CMIP6) of the World Climate Research Program (WCRP) and contributed to Intergovernmental Panel on Climate Change Sixth Assessment Report (IPCC AR6), the first time from India (Masson-Delmotte et al., 2021).  Earth System Modelling activities will enable better understanding of science of climate change and provide scientific input for detection and attribution of climate change over the Indian region and frame climate adaptation and mitigation policies.

This session will introduce the audience to the concept of climate, and motivate the reasons to develop a hierarchy of models for climate. Over two sessions, we will work with the basic principle of conservation of energy and see how it can be used to build, understand and interpret a hierarchy of models. In the first session, we will start with a point model for climate and move towards a single dimensional model.

In three lectures, we will provide an introduction to Earth system modeling and ocean modeling, primarily using the Community Earth System Modeling (CESM) framework. We will start with a discussion of why we use Earth system models; complexity, ensemble size, and resolution considerations; and their capabilities. We will then show examples of challenges and application areas; high-resolution simulations; and on-going efforts and opportunities. Ocean modeling lecture will consist of two parts. The first part will cover ocean modeling challenges and unique properties of oceans; basic governing equations and some approximations used; discretization of equations; and grid examples. The second part will focus on common parameterizations employed in ocean modeling, such as vertical mixing and mesoscale mixing parameterizations.

The talk will focus of regional climate modeling at both climate and weather scale with primary focus on understanding local processes, developing high resolution impacts model and the development of sector specific early warning system.

In this second session, we will focus more on the development of climate model components. We will see how to develop a model component for climt, and learn how climate models are structured. Based on this, we will learn how to think about ML emulation of climate model components and how to use climt to facilitate development and validation of such components.

In three lectures, we will provide an introduction to Earth system modeling and ocean modeling, primarily using the Community Earth System Modeling (CESM) framework. We will start with a discussion of why we use Earth system models; complexity, ensemble size, and resolution considerations; and their capabilities. We will then show examples of challenges and application areas; high-resolution simulations; and on-going efforts and opportunities. Ocean modeling lecture will consist of two parts. The first part will cover ocean modeling challenges and unique properties of oceans; basic governing equations and some approximations used; discretization of equations; and grid examples. The second part will focus on common parameterizations employed in ocean modeling, such as vertical mixing and mesoscale mixing parameterizations.

"On intra-seasonal, seasonal or interannual time scales, the ocean and the atmosphere interact, and an ocean-atmosphere coupled model (AOGCM) is required for these predictions. In comparison to developments in application of AI/ML models to Weather prediction, the developments on seasonal prediction have been lagging. This Talk will start with summarizing the status and gaps in developments of AI/ML models for seasonal prediction.  It will be argued that there is still scope of original contributions even in model developments for seasonal prediction. It will also be argued that a ‘hierarchical temporal aggregation’ method of prediction may be a better strategy rather than a ‘roll over’ prediction strategy for long-lead seasonal prediction. It is also argued that to maximize the skill of prediction, ‘physics guided training’ is essential. Applying these ideas, we shall demonstrate that an AI/ML model can predict the seasonal mean Indian summer monsoon rainfall (ISMR) with useful skill up to 24-months in advance, far beyond the capability of current climate models. In another application, we shall also demonstrate that the same model could be successfully used to predict the frequency and Intensity of daily rainfall extreme event over Central India during June-September.
At present, there is no ocean-atmosphere coupled model that given an ‘initial condition’ and prescribed external forcing can produce the natural variability of the system like the CMIP6 models. In a first, we shall show results of an AI/ML model to simulate the modes of global SST variability, particularly the ENSO variability better than the CMIP6 model on which the AI/ML model is trained. Thus, there is scope for making cutting-edge contribution on seasonal prediction not only of Indian monsoon rainfall but also other climate phenomena like the ENSO and NAO etc. Unfortunately, efforts are restricted by lack of adequate manpower and computational support. A superannuated scientist like me can not even write a Research Proposal as a PI."

Atmospheric convection refers to the vertical movement of buoyant air parcels, driving processes like cloud formation, thunderstorms, and large-scale circulations such as the Hadley and Walker cells. It is a rapid, non-local transport mechanism involving heat, moisture, momentum, and phase changes of water, influencing extreme weather and climate dynamics. Convection is typically not resolved explicitly in most weather and climate models but its representation is key to accurately representing convection-circulation feedbacks that control climate sensitivity and variability across global climate models. In this lecture, we will discuss a brief overview of Atmospheric convection and its representation in Climate Models.

In recent years, there has been a rapid increase in the number of scientific studies exploring the use of machine learning (ML) to enhance rainfall forecasting. This growing interest reflects the potential of ML techniques to address some of the limitations of traditional methods. In this lecture, I will provide a concise overview of several key studies that have applied ML to improve rainfall predictions. These examples highlight the diversity of approaches and the progress made in this emerging area. Towards the end of the talk, I will present a case study evaluating the performance of a fully ML-based rainfall forecast model over India.

TBA

TBA

Air quality modelling is essential for understanding how pollutants disperse and transform in the atmosphere, helping us model effect of various sources and predict exposure levels. These models enable us to engage in evidence-based policymaking, health risk assessments, and the formulation of effective pollution mitigation strategies. Nonetheless, validating models such as WRF-CAMx is essential to ensure our simulations accurately represent real-world pollution levels. Our primary data source comes from the Continuous Ambient Air Quality Monitoring Stations (CAAQMS) established by the Central Pollution Control Board (CPCB). Subject to regular maintenance, these CAAQMS stations can provide high-quality measurements of key pollutants; however, their limited spatial coverage and high setup and maintenance costs, poses a challenge when we need hyper-local model validation.

Low-cost sensors (LCS), while currently not used for regulatory purposes, are useful for research. It has started to gain traction with ULBs and State PCBs. These compact and affordable devices can measure various pollutants and be deployed in large quantities across urban, suburban, and rural areas. However, output from these sensors is often noisy and biased, drifts over time, and/or is affected by environmental factors such as temperature and humidity. These sensors are often factory-calibrated in a controlled environment, and thus, there is a serious need for these sensors to be calibrated again in the ambient real-world environment, where they are to be deployed.

To overcome these issues, we turn to machine learning. We can gather time-paired datasets of low-cost and reference-grade measurements by co-locating LCS devices with reference-grade instruments. This data is then used to train calibration models that map the raw sensor outputs to accurate pollutant concentrations from the reference grade instruments. In this talk, I’ll explain how we apply machine learning techniques—such as Multiple Linear & Quadratic Regression (MLR), Random Forest Regression, Support Vector Regression (SVR), and XGBoost—to develop these calibration models.

These machine-learning models allow us to correct for sensor drift, cross-sensitivity, and tackle environmental noise. The calibration activity helps us develop a model that corrects the sensor data and help validate WRF-CAMx outputs. This approach offers a scalable, cost-effective solution for extending our observational network—especially valuable in regions where CAAQMS coverage is sparse or non-existent.

I’ll also share insights into how different ML models perform in calibration tasks, especially in capturing non-linear relationships and which method performs the best.

This additional data from LCS helps us improve our model setup, including inputs such as meteorology, emissions, and model parameters, to bring the model output closer to observation. Once validated, we can use the improved model setup to accurately predict pollution reduction scenarios and support more informed policy interventions.

In short, this talk is about how we can bring together air quality modelling, low-cost hardware, and machine learning tools to build more robust and validated air quality models—particularly for India, where scalable solutions are not just helpful but necessary.

Indian Monsoon is a phenomenon that is not only extremely important for the livelihood and food security of over a billion people, but it is also a topic of active research in Climate Sciences for over a hundred years. Starting from forecasting at different spatial and temporal scales (daily, subseasonal, seasonal etc) to identifying its global drivers, analyzing its different phases to understanding the impacts of climate change on it - there are many open questions around it. As Artificial Intelligence and Machine Learning have become indispensable in many scientific domains, researchers are increasingly using them to answer questions related to the Indian monsoon. This talk will give a gentle introduction to how questions related to Indian Monsoon can be formulated as ML problems, and also discuss suitable ML models and algorithms to solve them.

This talk will be discussed in two parts. In the first part, we will discuss the potential of Deep Learning (DL) technique for an improved future streamflow projection from General Circulation Model (GCM) simulations, by developing a reliable association between the observed streamflow and a set of primary meteorological variables at monthly scale over a historical period. Towards this, a DL-based Long Short-Term Memory (LSTM) framework is developed to capture the hidden complex dynamics between streamflow and its two primary hydrometeorological precursors – precipitation and temperature, identified through Kendall’s partial correlation analysis. In the second part, we will focus on the development of a long‑range hydrological drought prediction framework using DL. In general, long-range (1 to 6 months in advance) prediction of droughts is challenging due to its inherent complexity. One-dimensional Convolutional Neural Networks (Conv1D) is used to develop a Hydrological Drought Prediction Framework (HDPF). We will discuss how the developed HDPF is able to extract the hidden information from the pool of eight different meteorological precursors to provide predictions with more than 70% accuracy.
 

In this talk, I will discuss the physical nature of the problem of the subseasonal variability and seasonal variability of tropical climate, with emphasis on the monsoon. In this context, I will discuss some advantages of modern machine learning tools.

We present a lightweight, easy-to-train, low-resolution, fully data-driven climate emulator, LUCIE, that can be trained on as low as 2 years of 6-hourly ERA5 data. Unlike most state-of-the-art AI weather models, LUCIE remains stable and physically consistent for 100 years of autoregressive simulation with 100 ensemble members. Long-term mean climatology from LUCIE's simulation of temperature, wind, precipitation, and humidity matches that of ERA5 data, along with the variability. We further demonstrate how well extreme weather events and their return periods can be estimated from a large ensemble of long-term simulations. We further discuss an improved training strategy with a hard-constrained first-order integrator to suppress autoregressive error growth, a novel spectral regularization strategy to better capture fine-scale dynamics, and finally an optimization algorithm that enables data-limited (as low as 2 years of 6-hourly data) training of the emulator without losing stability and physical consistency. Finally, we provide a scaling experiment to compare the long-term bias of LUCIE with respect to the number of training samples. Importantly, LUCIE is an easy to use model that can be trained in just 2.4h on a single A-100 GPU, allowing for multiple experiments that can explore important scientific questions that could be answered with large ensembles of long-term simulations, e.g., the impact of different variables on the simulation, dynamic response to external forcing, and estimation of extreme weather events, amongst others.

The advent of Artificial Intelligence marks a paradigm shift in weather and climate science, offering unprecedented potential to enhance and even surpass traditional numerical approaches in Earth System Modeling. This seminar will delve into the AI revolution that has profoundly impacted the weather domain in recent years. We will explore how AI has improved forecasting capabilities, fundamentally altered traditional modeling chains, and democratized the accessibility and application of weather models. Our discussion will specifically highlight two particularly impactful applications of advanced Machine Learning in meteorology and climate: AI-driven climate downscaling and next-generation short-term precipitation forecasting.
We will first explore AI-Driven Climate Downscaling: Bridging the Resolution Gap. While Global Climate Models (GCMs) offer crucial climate insights, their coarse resolution limits regional applicability. We address this with an innovative AI-driven approach, developing a deep generative Latent Diffusion Model (LDM) to downscale GCM outputs (approx. 1∘) to 4 km resolution for 6-hourly precipitation and 2-m temperature. Trained on a 43-year dataset derived from ERA5 and VHR-REA_IT, our LDM, leveraging a residual approach, significantly enhances the reconstruction of extreme values and preserves spatio-temporal coherence. Once trained, this model is applied to multiple CMIP6 GCMs, creating a multi-model ensemble of high-resolution climate projections to quantify future uncertainties.
Second, we will present RUSH: Next-Generation Short-Term Precipitation Forecasting. Leveraging AI capabilities is crucial for improving short-term precipitation forecasting, especially for rapid, intense events. We introduce RUSH (Rapid-Update Short-term High-resolution forecast), a fully AI-driven latent-diffusion sequence-to-sequence model designed to produce probabilistic 30-min precipitation accumulations on a 1 km grid for lead times up to 24 hours. RUSH innovatively blends real-time radar observations, satellite imagery (SEVIRI), and large-scale dynamical context from ECMWF’s AI Forecasting System (AIFS). The model adaptively selects the most relevant information from each source, enabling a seamless transition from observation-based nowcasting to AIFS-conditioned short-range forecasting without manual blending. Preliminary experiments over Belgium demonstrate RUSH’s rapid-update capability, robust skill across lead times and rain-rate thresholds—including for extreme events—and its potential as a data-driven complement to existing operational warning systems.
Both of these cutting-edge applications leverage latent diffusion models, providing an in-depth look at how these advanced deep learning techniques are being applied to solve complex challenges in Earth System Modeling. This talk will introduce the core concepts of latent diffusion models and their specific adaptations for meteorological and climatological data. This research represents core activities within Fondazione Bruno Kessler’s strategic portfolio of AI research addressing meteorological and climatological challenges.

In this talk, I will show some examples of predictions and how AIML tools are useful for predictions. These examples will include my work and some contemporary research work in this field

"The increasing frequency and intensity of extreme weather and climate events could claim over a million lives and result in annual economic losses exceeding $1.7 trillion by 2050 (Munich Reinsurance). This highlights the urgent need for more accurate and timely weather forecasting—especially in the face of cyclones, heatwaves, and other severe events.

Over 180 global weather modeling centers rely on high-performance computing (HPC) infrastructure to run traditional numerical weather prediction (NWP) models. For example, the UK Met Office’s supercomputer utilizes over 1.5 million CPU cores, consuming 2.7 megawatts of power. NVIDIA’s GPUs are now transforming this landscape—boosting the performance of widely used models developed by ECMWF, the Max Planck Institute for Meteorology, the German Meteorological Service, and NCAR. With up to 24x speedups, GPUs also significantly enhance energy efficiency and reduce infrastructure costs.

NVIDIA’s Earth-2 platform leverages AI, GPU acceleration, physical simulations, and computer graphics to build digital twins of Earth. These digital twins enable faster, scalable, and more accurate global weather and climate forecasts. The platform integrates advanced AI models like FourCastNet and CorrDiff, running on NVIDIA’s powerful HPC infrastructure to address climate change and support extreme weather preparedness.

In this session, Dr. Manish Modani will explore the role of accelerated computing and AI in advancing weather and climate forecasting. The talk will include a live demonstration of FourCastNet installation and inference workflows, showcasing its potential in operational and research forecasting applications."

"In recent years, much progress has been made in leveraging meteorological knowledge to enable accurate and fast weather forecasting for applications. The modeling approaches blend numerical weather prediction (NWP), large eddy simulations (LES), and machine-learning (ML) models. In this way, the best of our physics knowledge is blended with ML where it has a chance to speed calculations or contribute to better physics representation.

ML is integrated in three primary ways: 1. Post-processing of physical model runs to decrease bias compared to observations, 2. Replacing or emulating physics parameterizations to speed or improve computation, and 3. Full model ML that has learned the dynamics and physics of the system from prior physical model runs and/or observations. This lecture will describe examples of each performed at NSF NCAR, delve into the details of their application, and speculate on future implementations."