---
license: apache-2.0
pipeline_tag: tabular-classification
---

# SupraWeather1.5-Small

## Model Description

SupraWeather1.5-Small is a compact weather phenomenon classification model based on an FT-Transformer architecture.

The model takes a set of atmospheric variables as input and predicts one of 25 synthetic weather phenomena. It is trained entirely on procedurally generated atmospheric data and is intended for experimentation with structured machine learning models rather than real-world weather forecasting.

---

## Architecture

* Model: FT-Transformer
* Transformer blocks: 4
* Attention heads: 8
* Token dimension: 256
* Feed-forward factor: 1.333
* Attention dropout: 0.15
* Feed-forward dropout: 0.10

---

## Input Features

### Continuous Features

* Temperature
* Humidity
* Pressure
* Pressure Trend
* Wind Speed
* Altitude
* Cloud Cover
* Visibility
* Solar Radiation
* Dew Point
* Heat Index
* Wind Chill
* Storm Index
* Rain Potential
* Instability Index

### Categorical Features

* Wind Direction (8 categories)
* Month (12 categories)
* Air Mass (6 categories)
* Climate Zone (6 categories)
* Terrain Type (5 categories)

---

## Output Classes

The model predicts one of the following 25 classes:

* Clear
* Partly Cloudy
* Cloudy
* Overcast
* Mist
* Fog
* Dense Fog
* Light Rain
* Rain
* Heavy Rain
* Torrential Rain
* Thunderstorm
* Severe Thunderstorm
* Snow
* Heavy Snow
* Blizzard
* Freezing Rain
* Ice Storm
* Soft Hail
* Sleet
* Cold Front
* Heat Wave
* Cold Wave
* Windstorm
* Dust Storm

---

## Training

The model was trained using:

* Optimizer: AdamW
* Learning rate: 1e-3
* Weight decay: 1e-4
* Scheduler: Cosine Learning Rate
* Loss: Focal Loss (γ = 2.0, class-balanced α)
* Mixed precision (FP16)

Dataset split:

| Split      | Samples |
| ---------- | ------: |
| Train      | 240,000 |
| Validation |  30,000 |
| Test       |  30,000 |

Total synthetic samples: **300,000**

---

## Dataset

The training dataset is procedurally generated.

Instead of using real meteorological observations, atmospheric variables are sampled from a synthetic weather generation pipeline that produces internally consistent combinations of atmospheric conditions and weather phenomena.

The dataset includes interactions between variables such as:

* temperature
* humidity
* pressure
* pressure trend
* wind
* altitude
* cloud cover
* solar radiation
* air masses
* terrain
* climate zones

---

## Get started
```Code being writed
```

## Intended Use

This model is intended for:

* Machine learning experiments
* Synthetic weather simulation
* Educational purposes
* Procedural simulation systems
* Games and virtual environments

---

## Limitations

This model is **not** a numerical weather prediction (NWP) model.

It does not simulate Earth's atmosphere and should not be used for operational weather forecasting or scientific meteorology.

Predictions are valid only within the synthetic atmospheric distribution used during training.


---

## License

This model is released under apache-2.0 license.