Smart India Hackathon 2026 · PS ID: SIH26068

Precision Weather Intelligence Powered by Physics & Machine Learning

WeatherGPT bridges global numerical weather prediction (NOAA GFS) with high-resolution reanalysis (ECMWF ERA5) using an XGBoost bias-correction engine and conversational AI.

94.2%
Systematic Bias Reduction
< 0.3°C
Mean Absolute Error (MAE)
0.25°
Global Grid Resolution
Real-Time
IMD Event Warning Listener

Ministry of Earth Sciences (MoES) — India Meteorological Department (IMD)

Official Title: WeatherGPT: Conversational AI for Weather Forecasting, Alerts, and Climate Information (Theme: Disaster Management).

BACKGROUND & CHALLENGE

Fragmented Weather Information Channels

Weather information is often distributed through multiple portals, bulletins, satellite products, and forecast systems, making it difficult for common users, researchers, disaster managers, and government agencies to quickly obtain actionable insights.

EXPECTED SOLUTION & OBJECTIVE

Conversational Intelligence Platform

Develop WeatherGPT: an intelligent conversational platform integrating meteorological datasets, forecasting models (GFS/WRF), disaster early warnings, location-based advisories, voice interaction for rural accessibility, and multilingual Indian language support.

Organization
Ministry of Earth Sciences (MoES)
Department
India Meteorological Department (IMD)
Category & Theme
Software · Disaster Management
PS Number
SIH26068 (S.No. 68)
GitHub Repository
rajrounak21/WeatherGpt ↗

Why Physics-Based Weather Models Need Machine Learning

Traditional Numerical Weather Prediction (NWP) models like NOAA GFS calculate atmospheric fluid dynamics on global grids. However, localized micro-climates and topography create systematic temperature errors up to ±5°C.

The Raw NWP Limitation

Global numerical circulation models (GFS) smooth out hyper-local topography such as urban heat islands, river valleys, and coastal sea breezes. This leads to consistent over-prediction of daytime surface heating and under-prediction of nighttime cooling.

🧠

The WeatherGPT ML Solution

WeatherGPT trains a specialized XGBoost error-correction model on multi-decade ECMWF ERA5 reanalysis observations. It learns the non-linear residual bias function ΔT = T_ERA5 - T_GFS across cyclic temporal, seasonal, and spatial coordinates.

End-to-End AI Forecasting Architecture

From raw numerical weather model grids to hyper-local natural language intelligence.

STEP 01

GFS NWP Ingestion

Fetches 0.25° global grid 16-day numerical forecasts from NOAA GFS run cycles (00z, 06z, 12z, 18z).

STEP 02

Feature Vector Encoding

Extracts 2m temperature, surface pressure, humidity, u/v wind vectors, solar radiation, cyclic sine-cosine encodings, and forecast lead time.

STEP 03

XGBoost Bias Correction

Evaluates the high-dimensional gradient boosted decision tree to predict exact temperature residual corrections (ΔT).

STEP 04

LLM Agent Synthesis

Groq LLM agent synthesizes structured weather parameters into natural language explanations in Hinglish or English with speech audio synthesis.

Raw GFS vs. WeatherGPT ML vs. Ground Truth (ERA5)

Interactive temperature comparison over a 24-hour diurnal cycle. Observe how WeatherGPT's corrected curve tracks the ERA5 ground-truth observations.

Diurnal Temperature Benchmark (°C)

Location: Patna, Bihar (25.59° N, 85.13° E)
Raw GFS (Overpredicts daytime heat)
ERA5 Ground Truth
WeatherGPT Corrected (MAE < 0.3°C)

Engineered for Precision Weather Intelligence & Public Safety

Aligned with Smart India Hackathon SIH26068 expected outcomes and deliverables.

💬

AI Natural Language Weather Agent

Ask weather queries naturally in English or Hinglish. Powered by Groq LLM agent with real-time tool calling into the XGBoost numerical pipeline.

📈

Unified Climate History & Forecast

Seamless timeline blending multi-year ERA5 historical climate trends with 16-day ML-enhanced GFS forecasts in a responsive graph interface.

🚨

IMD Severe Early Warning System

Event-driven background service monitoring India Meteorological Department warnings with district-matched VAPID Web Push notifications.

🎓

Meteorology Study Hub

Interactive micro-courses, quizzes, and AI-assisted simplified explanations for atmospheric science topics (Thunderstorms, Heat Waves, Cyclones).

🎙

Voice STT & Audio Synthesis

Hands-free voice query processing using Groq Whisper Speech-to-Text and real-time Text-to-Speech audio response generation for rural accessibility.

🌙

Modern Adaptive UI System

Built with rich glassmorphism aesthetics, fluid HSL color spaces, responsive grid layouts, and instantaneous dark/light theme switching.

Designed for Fast, Intuitive User Experiences

Overview of the 4 key application views available in WeatherGPT.

💬

AI Chat Interface

Natural language conversational assistant with instant response cards, voice input mic, structured weather metrics, and speech audio playback controls.

📈

History & Timeline View

Interactive climate historical graph connecting historical ERA5 reanalysis data with 16-day machine learning bias-corrected forecasts.

🚨

Severe Alert Manager

District-based alert subscription manager with live IMD warning status monitoring and instant VAPID push notification delivery.

🎓

Meteorology Study Hub

Multi-column learning grid featuring interactive lessons, atmospheric science quizzes, progress tracking, and AI-assisted simplified explanations.

Built with High-Performance Open Science Stack

Engineered for low latency, maximum precision, and scalable background processing.

🐍 Python FastAPI 🌲 XGBoost ML Engine 📡 NOAA GFS 0.25° Grid 🌍 ECMWF ERA5 Reanalysis ⚡ Groq LLM Agent 🍃 MongoDB Atlas 🔔 Web Push VAPID 🎙 OpenAI Whisper STT 🎨 Modern Vanilla CSS & JS