A comparative study of Classical ML and Quantum ML for temperature forecasting on 30 years of NASA MERRA-2 climate data across Indian states — with an interactive simulator that predicts future temperatures up to 2035.
This project compares computational regimes, not just performance. Same climate problem. Fundamentally different computation.
Qlimate trains classical and quantum regression models to predict surface temperature (°C) for any Indian state, month, and year — including future projections up to 2035. The central question: which computing paradigm predicts climate temperature more accurately, and how do their forecasts diverge?
| Dimension | Classical ML | Quantum ML |
|---|---|---|
| Target | T2M temperature (°C) | T2M temperature (°C) |
| Training samples | 8,400 (1995–2019) | 400 (simulation limit) |
| Feature count | 17 | 4 (PCA compressed) |
| Best RMSE | 0.73°C (Gradient Boosting) | 6.56°C (QSVR) |
| Best R² | 0.993 | 0.425 |
| Training time | seconds–minutes | 15 min (QSVR), 5 min (VQR) |
| Hardware | CPU/GPU workstation | Qiskit StatevectorSampler (CPU simulation) |
| Model | Type | MAE (°C) | RMSE (°C) | R² | Training samples |
|---|---|---|---|---|---|
| Gradient Boosting | Classical | 0.554 | 0.732 | 0.993 | 8,400 |
| XGBoost | Classical | 0.546 | 0.735 | 0.993 | 8,400 |
| Random Forest | Classical | 0.609 | 0.824 | 0.991 | 8,400 |
| Neural Network | Classical | 0.740 | 0.993 | 0.987 | 8,400 |
| Ridge Regression | Classical | 1.155 | 1.426 | 0.973 | 8,400 |
| QSVR | Quantum | 4.030 | 6.558 | 0.425 | 400 |
| VQR | Quantum | 21.883 | 23.239 | −6.22 | 400 |
Classical models achieve R² > 0.99 — predicting temperature within 0.7°C. Quantum models are constrained to 400 samples and 4 compressed features due to NISQ hardware limits, resulting in significantly higher error. This gap is the core finding.
graph LR
subgraph Data Collection
A[NASA Climate Satellite] -->|raw data| B[Download and Store Files]
end
subgraph Data Preparation
B -->|raw files| C[Aggregate to Indian States]
C -->|state averages| D[Engineer Features and Split Data]
end
subgraph Model Training
D -->|full 17 features| E[Train Classical Models]
D -->|4 compressed features| F[Train Quantum Models]
end
subgraph Serving Predictions
E -->|saved model files| G[Live Prediction Server]
F -->|precomputed answers| G
end
subgraph User Interface
G -->|final answer| H[Interactive Web Simulator]
end
| Step | What happens |
|---|---|
| 1 | Thirty years of monthly climate data is downloaded from NASA satellites covering all 28 Indian states |
| 2 | The raw data is cleaned and averaged from a global grid down to individual state-level summaries |
| 3 | Extra signals are calculated from the raw numbers — things like wind speed, how much water evaporated, and which time of year it is — giving each data point 17 descriptive features |
| 4 | Five classical models and two quantum models are each trained separately on this dataset to learn the relationship between those features and the actual recorded temperature |
| 5 | Because quantum models take hours to run a single prediction, all 13,776 possible combinations of state, month, and year are calculated in advance and stored in a lookup table |
| 6 | A lightweight web server handles incoming requests — using the classical model for instant live predictions and the lookup table for quantum results |
| 7 | The browser-based simulator lets you pick any Indian state, month, and year up to 2035, then shows both predictions side by side so you can see exactly where classical and quantum computing agree or diverge |
The simulator lets you select any Indian state, month, and year (1995–2035) and see classical vs quantum temperature predictions side by side, along with a 2025–2035 forecast divergence chart.
How future predictions work: For years 2025–2035, both models use the 30-year historical average climate features for that state+month, with year set to the requested future year. The year feature captures the temperature trend learned from 1995–2024 data.
Quantum inference note: Quantum predictions are precomputed offline for all 13,776 (state × month × year) combinations and served from a lookup table. Real-time quantum inference is not feasible — training QSVR on 400 samples requires ~22 hours of IBM quantum computer time, exceeding available quotas. This is standard practice in quantum ML research.
Source: NASA MERRA-2 Monthly Reanalysis (GES DISC / Earthdata)
| Property | Value |
|---|---|
| Collections | M2TMNXSLV · M2TMNXFLX · M2TMNXRAD |
| Variables | T2M, QV2M, U10M, V10M, PS, SLP, PRECTOT, EVAP, CLDTOT |
| Coverage | 1995–2024 (355 months) · 28 Indian states |
| Total rows | 9,940 |
| Split | Chronological — train 1995–2019, val 2020–2021, test 2022–2024 |
| Target | T2M_celsius = T2M − 273.15 |
17 regression features: PRECTOT, QV2M, PS, SLP, SWGDN, LWGNT, CLDTOT, EVAP, wind_speed, wind_direction, net_radiation, precip_evap_ratio, month_sin, month_cos, pressure_anomaly, year, state_encoded
The year feature is what enables temporal extrapolation to 2035 — models learn the temperature trend over 1995–2024 and project it forward.
git clone <repo-url>
cd Qlimate
pip install -r requirements.txtexport EARTHDATA_PASSWORD="your_password" # Required for data downloadUpdate config/config.yaml with your Earthdata username.
# Full pipeline (after data download)
python run.py
# Individual stages
python run.py --only engineer # Feature engineering + chronological splits
python run.py --only classical # Train 5 classical regressors (~9 min)
python run.py --only quantum # Train QSVR + VQR (~1 hour, CPU simulation)
python run.py --only precompute # Generate quantum_predictions.json (~1.5 hours, parallel)
python run.py --only evaluate # Compare all models, write model_comparison.csv
# Resume from a stage
python run.py --from evaluatepython src/data/download.py~3.5 GB, 1,080 NetCDF4 files.
python scripts/precompute_quantum.pyGenerates results/quantum_predictions.json — 13,776 predictions (28 states × 12 months × 41 years) using all available CPU cores in parallel.
Qlimate/
├── config/config.yaml # All hyperparameters and paths
├── run.py # Pipeline orchestrator
├── scripts/
│ ├── precompute_quantum.py # Parallel offline quantum prediction
│ └── export_metrics.py # Export metrics to JSON
├── src/
│ ├── data/
│ │ ├── download.py # MERRA-2 Earthdata download
│ │ └── preprocess.py # Grid-to-state aggregation
│ ├── features/
│ │ └── engineering.py # Feature engineering, chronological splits, PCA
│ ├── models/
│ │ ├── classical.py # ClassicalRegressorTrainer (XGBoost, RF, Ridge, GB, NN)
│ │ └── quantum.py # QuantumRegressorTrainer (QSVR, VQR)
│ ├── evaluation/
│ │ └── metrics.py # MAE, RMSE, R² evaluation + comparison DataFrame
│ └── visualization/
│ └── ... # Static + interactive plots
├── backend/
│ └── predict_server.py # FastAPI: classical live inference + quantum lookup
├── frontend/ # React/Vite simulator
│ └── src/
│ ├── pages/Simulator.jsx # Main simulator page (homepage)
│ ├── components/
│ │ ├── TemperatureOutputPanel.jsx
│ │ ├── ForecastDivergenceChart.jsx
│ │ ├── DifferencePanel.jsx
│ │ ├── PredictionControls.jsx
│ │ └── IndiaMap.jsx
│ └── api/predict.js # API client + mock fallback
├── results/
│ ├── figures/ # PNG figures
│ ├── models/ # Saved model artifacts + quantum_predictions.json
│ ├── metrics/ # JSON metric exports
│ └── model_comparison.csv # Full comparison table (MAE, RMSE, R²)
└── data/
├── raw/ # MERRA-2 NetCDF4 (gitignored)
└── processed/ # State CSVs + regression splits (gitignored)
pytest tests/Covers feature engineering, regression metrics, classical model interfaces, quantum model interfaces, and data pipeline integration. All tests use synthetic data — no dependency on real data files or config.
- Dataset: NASA MERRA-2 via GES DISC / Earthdata
- Quantum framework: Qiskit + qiskit-machine-learning by IBM
- India boundaries: Subhash9325/GeoJson-Data-of-Indian-States