×

Deep Learning–Based Forecasting of Agricultural Droughts over India

Professor: Prof. Karthikeyan Lanka

Description:

Agricultural droughts, driven by deficits in soil moisture, pose a serious risk to crop productivity and food security in India. Reliable sub-seasonal to seasonal (S2S) drought forecasts can support better crop planning and water management decisions.

This internship focuses on developing deep learning–based time-series forecasting models, including probabilistic forecasting approaches, to predict agricultural drought conditions over India. The project leverages more than 40 years of satellite and climate reanalysis data, processed to 25 km spatial resolution at monthly time steps.

Predictors include historical rainfall, temperature, vapor pressure deficit, soil moisture, winds, evapotranspiration, vegetation indices, and large-scale atmospheric teleconnections. The predictand is future soil moisture state at specified lead times. All datasets are pre-processed and will be provided.

Objectives:

1. Implement and compare state-of-the-art deterministic and probabilistic time-series deep learning models (e.g., LSTM/GRU, Temporal CNNs, Transformers, quantile regression, Bayesian or ensemble-based models).


2. Evaluate forecast skill and uncertainty using metrics such as correlation, RMSE, CRPS, reliability, hit/false alarm rates, and assess feature importance / interpretability.


3. Deliver a well-documented Python codebase for the best-performing model.

Expected Skill Gain:

1. Deep learning for time-series and probabilistic forecasting for agriculture stress early warning systems


2. Handling large-scale real-world geospatial datasets


3. Model evaluation and uncertainty quantification


4. Writing clean, production-quality Python code

Pre-requisites:

• BTech/BS students in Engineering, CS, AI/ML, Data Science, or related fields
• Prior experience with deep learning–based time-series forecasting models
• Proficiency in Python (PyTorch/TensorFlow preferred)
• Strong interest in AI applications for climate and sustainability

×

Chasing Storms: Does Longevity Drive India’s Extreme Rain?

Professor: Prof. Vishal Dixit

Description:

Globally, 40% of land-based extreme rainfall days (>250mm) are driven by Mesoscale Convective Systems (MCSs) lasting over 24 hours. However, the lifespan and geographical variance of these marathon storms across the Indian subcontinent remain poorly understood. This project investigates whether longer-lived systems are the primary architects of India’s flood events.

Using a comparative framework, we will evaluate two state-of-the-art tracking algorithms: TOOCAN, which uses 3D space-time segmentation to isolate convective cores, and MOAAP, which employs 2D water-shedding to track atmospheric features.

Key Objectives:

1. Methodological Benchmarking: Contrast TOOCAN’s 3D life-cycle mapping against MOAAP’s 2D overlap approach to redefine MCS distribution over India.


2. Rainfall Attribution: Quantify the link between system duration and precipitation volume, testing the hypothesis that long-lived MCSs contribute disproportionately to the region’s hydrological extremes.

Reference:

Roca, Rémy, and Thomas Fiolleau. "Extreme precipitation in the tropics is closely associated with long-lived convective systems." Communications Earth & Environment 1.1 (2020): 18.

Pre-requisites:

• Comfort with Python and computing, Past experience of handling NetCDF data is a plus.

×

Tracking Algorithm for Precursors of Extreme Weather Events

Professors: Prof. Akshaya Nikumbh, Prof. Vishal Dixit

Description:

Accurately forecasting the location and intensity of extreme weather events remains a major challenge for state-of-the-art weather models. Such extremes often arise from complex interactions among weather systems across a range of spatial and temporal scales. Tracking weather systems therefore offers a promising pathway toward improving process-based understanding and forecasting of extreme events. In this project, the student will develop a multi-object tracking algorithm using open-source observational datasets to detect and characterize different weather systems across various spatio-temporal scales. The student will be mentored by Prof. Akshaya Nikumbh, Prof. Vishal Dixit, and graduate students from their research groups, who are actively working on tracking weather systems across scales.

Pre-requisites:

• Strong Python programming skills

×

What leads to wintertime Fog over Northern India?

Professor: Prof. Angshuman Modak

Description:

Wintertime fog over northern India has become a frequent and severe problem in recent years, affecting transportation, air quality, public health, and everyday life across the Indo-Gangetic Plain. Fog is generally associated with strong cooling at night, high humidity near the surface, weak winds, and a stable lower atmosphere. Although fog events are natural feature of winter season but in the past few years it has become a major issue in the Northern India. This raises a basic scientific question: what conditions lead to such persistent wintertime fog over northern India?

This project aims to build a strong understanding of fog formation by studying the atmospheric conditions that favor fog development. Instead of focusing on detailed fog modeling, the emphasis will be on identifying the key factors that control fog formation and persistence.

The project will involve reading and synthesizing classic and recent research on fog, aerosols, and winter boundary-layer processes, followed by systematic analysis of atmospheric datasets. By the end of the internship, the student should be able to clearly explain why winter fog forms and persists over northern India.

Pre-requisites:

• Comfort with data analysis and visualization
• Prior experience with Python, MATLAB, or similar tools is desirable
• Motivation to read scientific literature and synthesize physical ideas.

×

Inffluence of Aerosol chemical charateristics on radiative forcing

Professor: Prof. Abhishek Chakraborty

Description:

Aerosols interactions with solar radiation influence the climate by altering the energy balance. However, understanding the influence of aerosol chemical characteristics on this remains scarce. Objective: To evaluate the influence of aerosol chemical characteristics and evolution on aerosol induced radiative forcing. Deliverables: Quantifying the changes in aerosol radiative forcing due to constantly evolving aerosol chemical characteristics.

Pre-requisites:

• Basics of aerosol/air pollution, willing to work on big data and some coding skills for data analysis.

×

Using Autoencoders for Ocean Data-assimilation

Professor: Prof. Suyash Bire

Description:

Oceanic observation datasets come in distinct formats. For example, moored buoys provide time series of various oceanic quantities but at fixed locations. Argo floats also provide time series of various properties but at different locations as they drift with the oceanic currents. Satellites provide observations of the surface oceans at regular intervals. The process that merges these distinct sources into consistent gridded products is known as data assimilation. So far, all data assimilation techniques use numerical ocean models that provide background states into which these observations are embedded. However, some recent studies have generated atmospheric gridded products using autoencoders, a type of convolutional neural network. They show that autoencoders allow them to eliminate the expensive background numerical models while also producing more consistent gridded products. In this study, we will use autoencoders to merge oceanic observations into gridded products. We will further validate these products with preexisting gridded products.

Pre-requisites:

• Previous experience of using convolutional neural networks for vision tasks would be ideal

×

Predicting the Future of Forests: Numerical and Analytical Solvers for an Eco-Evolutionary Vegetation Model

Professor: Prof. Jaideep Joshi

Description:

Understanding how global vegetation responds to climate change is critical for predicting the global carbon cycle and informing climate-change mitigation strategies. This project contributes to the development of Plant-FATE (Plant Functional Acclimation and Trait Evolution), a next-generation eco-evolutionary vegetation model that simulates plant demographic and evolutionary dynamics at multiple timescales. While the model has been successful in numerically predicting forest dynamics over time, it is difficult to scale up globally due to computational constraints. This project aims to develop approximate equilibrium solutions of Plant-FATE using both analytical and numerical techniques, thus bypassing computationally intensive transient simulations. The developed solutions will contribute a step improvement in our capacity to predict carbon stocks, structure and function, and climate responses of global forests. Consequently, it can inform forest-based climate-change mitigation actions.

Pre-requisites:

• Calculus, differential equations, basic numerical methods