Aardvark Weather is an end-to-end multimodal weather forecasting model that generates global gridded forecasts and station-level predictions through an observation encoder, a global forecast processor, and a station decoder.
If the environment lacks Aardvark dependencies, refer to official-src/environment.yml for additional installation. The current adaptation is compatible with the Block parameter of the newer timm package.
Data & Weights
The model package already includes the resources required for 1-day temperature inference:
The training entry point provides a complete pipeline with epochs, validation, early stopping, learning rate scheduling, best/latest checkpointing, and resumption. By default, the Encoder and Day-1 Processor are frozen and only the TAS Decoder is trained:
python scripts/train.py
Paper-style end-to-end joint fine-tuning:
python scripts/train.py --train-modules all --epochs 10 --train-steps 100
The default configuration resides in conf/config.yaml. --data can point to an official-schema pickle file or a directory containing multiple .pkl files; multiple files are partitioned into train and validation sets in a deterministic manner. --batch-size concatenates multiple tasks along the existing batch dimension of the official task. When only the single official sample is included, training and validation reuse the same task — this allows end-to-end validation of the training software pipeline but does not constitute an independent validation set, nor can it provide the data diversity required to reproduce paper-level accuracy.
Current results constitute a connectivity/end-to-end verification and do not reproduce the paper's RMSE/MAE metrics.
The result script additionally produces result/metrics.json and result/comparison.png. The normalized_mae and normalized_rmse are computed in the normalized space of the official sample and should not be directly compared to the paper's physical-unit metrics.
Paper vs. Current Implementation I/O
Item
Paper
Current Package
Input
Multimodal satellite, station, ship, and radiosonde observations
Bundled official sample pickle with a field structure consistent with the official Encoder
Global State
24 × 121 × 240, 1.5°
Day-1 supported; output [1,121,240,24]
Station Output
2 m temperature and 10 m wind, up to Day-10
Day-1 TAS only; [1,8719]
Training
Staged pre-training followed by ~25,000 steps of end-to-end fine-tuning
Configurable full training loop; supports Decoder-only or full-model joint fine-tuning
Evaluation
Grid-point weighted RMSE and station MAE in physical units
MAE/RMSE in normalized sample space
All commands should be run from the project root; scripts/inference.py --root converts to an absolute path. The official model internally depends on CUDA, so CPU is currently unavailable for inference. Training data must adhere to the official multimodal task dictionary schema; the current directory does not synthesize satellite or station observations — the bundled official sample serves as the default training-pipeline validator. The model package retains only official resources under weights/ and does not carry local training artifacts under weight/ or generated outputs under result/. Paper-level training still requires preparing observation data spanning the full date range and converting it into the same .pkl task contract.
Real Data
Using real dates requires preparing ASCAT, AMSU-A/B, HIRS, IASI, GridSat, HadISD, ICOADS, IGRA, ERA5, topography, climatology, and their corresponding normalization statistics.
