SIH26078: AI-Driven Spatio-Temporal Tracking of Extreme Weather Anomalies in Medium-Range — SIH 2026 | Wynit

SIH26078: AI-Driven Spatio-Temporal Tracking of Extreme Weather Anomalies in Medium-Range — SIH 2026 | Wynit

What they'll ask you

13

Generated from the brief's own wording and the competition's published rules, never from a guess about what this ministry prefers.

Exact computational resource limits for the 36-hour final demo?
The brief never answers this, so a panel will. Whatever you decide, say it the same way twice.
Specific cloud deployment infrastructure requirements?
The brief never answers this, so a panel will. Whatever you decide, say it the same way twice.
Has any part of this been shown at a previous event, hackathon or college project?
The guidelines are explicit: your solution must not have appeared in any previous event or programme, of any sort. A recycled project is what a team under time pressure reaches for.

10 morethe ones a panel gets to if there is time

The brief asks for spatio-temporal gnn module. How would you build that?
Decoded from SIH26078 itself: Processes global forecast streams on an icosahedral mesh to isolate anomalies and output 4D bounding boxes. The brief asks for it by name.
The brief asks for generative diffusion module. How would you build that?
Decoded from SIH26078 itself: Takes the macroscale anomaly slice and downscales it from a 12 km grid to a 5 km grid while preserving extreme amplitudes. The brief asks for it by name.
The brief asks for physics-informed loss function. How would you build that?
Decoded from SIH26078 itself: Embeds fluid dynamics and thermodynamic conservation laws to penalize physically impossible weather states. The brief asks for it by name.
The brief asks for visualization dashboard. How would you build that?
Decoded from SIH26078 itself: Translates mathematical centroid arrays into clean geographic visual layers for weather data. The brief asks for it by name.
The brief asks for alerting rest api. How would you build that?
Decoded from SIH26078 itself: Programmatically drops pinpoint coordinates and triggers categorized spatial alerts across a 5 km radius. The brief asks for it by name.
Where does your data come from: a published source, one you collect, or one you generate?
No dataset is attached to this problem statement, so sourcing it is part of the work and nobody told you that.
Why not use what already exists? Name the closest thing to this that is already running.
A team that has not named the alternative themselves is answering this for the first time in the room.
Which single thing will you demonstrate end to end, start to finish, with nothing skipped?
Ours, not a rule: a narrow thing that fully works survives questioning better than a broad thing that half works. If nobody on the team can name it, that is the finding.
Show me this working: the jury will check whether your GNN correctly handles the spherical mesh without geographic distortions and if your diffusion model successfully preserves high-amplitude peaks during downscaling without smoothing them out.
This is the evaluator read for your problem statement, decoded from the brief's own wording.
Show me this working: they will test your custom physics-guided loss functions and your ability to process multigigabyte NetCDF or GRIB2 data arrays.
This is the evaluator read for your problem statement, decoded from the brief's own wording.

The government's original description (791 words)

• Problem Statement Identifying and tracking the exact geographic footprints of extreme weather anomalies (such as severe cyclones, heat domes, or cold waves) within massive global Numerical Weather Prediction (NWP) outputs is computationally intensive and heavily reliant on manual interpretation. In medium-range forecasting (3 to 10 days), atmospheric chaos renders traditional deterministic models highly uncertain. Furthermore, standard deep learning models (like standard CNNs or U-Nets) suffer from spectral smoothingâ€”they tend to 'average out' spatial data, which destroys the extreme amplitudes (the high-intensity peaks of rainfall or wind speed) that forecasters actually need to track. There is a critical gap between broad, coarse 12 km global ensemble datasets and localized, high-fidelity threat tracking. • Proposed Solution We propose an automated, state-of-the-art AI tracking and downscaling pipeline that shifts the paradigm from manual weather data sorting to automated, physics-informed anomaly tracking.Instead of relying on a single deterministic forecast run, our system directly processes multivariable, 4D Ensemble Prediction Systems (EPS) data.The system uses a two-stage hybrid AI architecture to solve the spectral smoothing problem:First, it utilizes a graph neural network (GNN) to map atmospheric variables onto a spherical mesh,instantly isolating moving anomalies and calculating their trajectory over a 3- to 10-day forecast window. Second, it pipes this isolated region into a generative diffusion model to perform statistical downscaling. This physics-constrained generative model mathematically derives a hyper-local 5km subgrid impact zone without flattening or blurring the severe amplitudes of the extreme weather event. • Technical Methodology & Architecture Spherical Anomaly Tracking (Stage 1 GNN): To eliminate the geographic distortions caused by processing the spherical Earth on flat 2D pixel grids, the system maps the 12 km NCMRWF Global Ensemble (NEPS-G) grids directly onto an icosahedral mesh. The message-passing GNN calculates the Extreme Forecast Index (EFI) against a 30-year historical ERA5 baseline distribution to isolate standard deviations and draw a macro-scale temporal bounding box around the anomaly's trajectory. Amplitude-Preserving Downscaling (Stage 2 Diffusion): The system passes the cropped, macroscale bounding box into a conditional denoising diffusion probabilistic model. Rather than optimizing for mean errors (which blurs peaks), the diffusion model learns the physical relationships between synoptic-scale features and regional topography. It iteratively generates high-resolution, high-amplitude local weather scenarios, downscaling the 12 km grid into a 5 km grid. Physics-Informed Constraints: To ensure the model remains scientifically accurate, we embed fluid dynamics and thermodynamic conservation laws directly into the neural network's loss function. The model is mathematically penalized if it generates physically impossible weather states (e.g., severe downpours missing corresponding moisture convergence vectors). • Datasets and Tools ? AI Frameworks: PyTorch / JAX (engineered with custom, physics-guided loss functions),Deep Graph Library (DGL) for icosahedral mesh networks, and Hugging Face Diffusers for generative downscaling. ? Data Wrangling & Geospatial Tools: Xarray and Dask for processing parallelized, multigigabyte 4D NetCDF/GRIB2 arrays; MetPy for physical meteorological equations; Cartopy for geographical map projections. ? Training & Testing Datasets: ? Baseline: Historical IMDAA / ERA5 reanalysis data to establish the climatological norm. ? Forecast Inputs: Historical NCUM (12 km deterministic) and NEPS-G (12 km global ensemble) datasets containing documented extreme historical events (e.g., Cyclone Amphan, severe North India heatwaves). • Expected Outcome & Key Deliverables The Tracking Core: A production-ready Spatio-Temporal GNN module that continuously processes global NWP streams to output dynamic, automated 4D bounding boxes around evolving weather threats. The Downscaling Core: A generative diffusion module capable of ingesting a 12 km resolution anomaly slice and outputting a probabilistically sound, 5 km resolution sub-grid array that retains extreme value amplitudes. The Visualization & Alert Dashboard: An automated system that translates the mathematical 5 km centroid arrays into clean, geographic visual layers. The Alerting API: A lightweight, production-ready REST API that programmatically drops a pinpoint coordinate at the core of the severe anomaly and triggers categorized spatial alerts (low, moderate, and severe) across a precise 5 km geographical impact radius. • Use Cases & Societal Impact Eliminating Alert Fatigue for the NDRF: Current weather alerts are often too broad, covering entire states or districts, which leads to public complacency. This solution allows meteorologists to issue hyper-localized, highly targeted warnings. It changes a generic'heavy rain in the district' alert into a precise 'high risk of flash flooding within your specific 5 km radius in the next 12 hours' alert, empowering first responders to deploy assets perfectly. Protecting Rural Economies: Grants farming communities a highly accurate, 3- to 10- day lead time regarding localized catastrophic anomalies like sudden frost, hail, or heat domes. This structural foresight lets farmers alter harvesting schedules or apply cropprotection covers, shielding rural livelihoods from sudden climate shocks. Democratizing Supercomputing Power: Once this hybrid AI pipeline is trained, it processes live inference data on a standard cloud GPU node in seconds, making high-fidelity climate forecasting highly affordable and easily accessible.