Model performance

How accurate is
Atmosphera?

Our UV and solar radiation models are trained on NASA satellite measurements and independently validated on held-out 2023 data — never seen during training.

97.3%
UV accuracy (R²)
2023 test set
0.28
UV error (MAE)
UV Index units
96.3%
Solar accuracy (R²)
2023 test set
31.8
Solar error (MAE)
W/m² out of 1200
832k
Training observations
hourly, 5 years
19
Climate zones
global coverage

Predicted vs actual UV Index — test year 2023

Model predictions against NASA satellite measurements. Training used 2019–2022 data only — 2023 was never seen.

Solar radiation accuracy — test year 2023

Predicted vs measured solar radiation (W/m²) across diverse global conditions.

Performance across climate zones

Validated across diverse climates. The model generalises well — it has never been tuned to a specific region.

Equatorial
Tropical humid
UV MAE 0.47 · Good
Desert / arid
Hot dry
UV MAE 0.22 · Excellent
Continental
Temperate seasonal
UV MAE 0.19 · Excellent
Monsoon
Seasonal humid
UV MAE 0.27 · Good
Sub-arctic
Cold seasonal
UV MAE 0.12 · Excellent
Oceanic
Marine temperate
UV MAE 0.20 · Excellent

What powers the model

Each prediction combines satellite-derived training data with real-time local inputs.

🛰️

NASA MERRA-2 satellite data

Training data sourced from NASA's Modern-Era Retrospective Analysis — the same dataset used by climate and atmospheric researchers worldwide.

☀️

Real-time solar geometry

Solar zenith angle is computed live from coordinates and UTC time. It changes every minute — no approximation or lookup table.

🌤️

Live weather from OpenWeatherMap

Current cloud cover, temperature, humidity and wind are fetched in real time and passed as inputs to the model on every request.

🏔️

Elevation-aware predictions

UV increases significantly with altitude. The model accounts for local elevation — critical for high-altitude cities across Central Asia, Africa, and Latin America.

Model trained on 2019–2022 · Validated on independent 2023 test set
Satellite reference: NASA POWER / MERRA-2 · Weather inputs: OpenWeatherMap

🔒 Model architecture and training details are proprietary