[CVPR 2026] FAKER-Air: Real-Time Long Horizon Air Quality Forecasting via Group-Relative Policy Optimization

FAKER-Air (GRPO)

Model Description

This repository provides the GRPO-trained checkpoint of FAKER-Air, a regional air-quality forecasting model for East Asia.

FAKER-Air (Forecast Alignment via Knowledge-guided Expected-Reward) is built on an Aurora-based 3D encoder-decoder and trained in two stages:

1. Supervised Fine-Tuning (SFT) with Temporal Accumulation Loss to reduce exposure bias during autoregressive rollout.
2. Group-Relative Policy Optimization (GRPO) with AQI-aware rewards and curriculum rollout to improve operational reliability, especially by reducing false alarms.

The model is designed for long-horizon particulate matter forecasting up to 120 hours (5 days) ahead using real-time observations and regional CMAQ reanalysis for East Asia.

Intended Use

This model is intended for:

• Research on long-horizon air-quality forecasting
• Spatiotemporal forecasting over East Asia
• Reliability-aware forecasting and AQI-oriented evaluation
• Offline experiments and prototype operational systems

Primary Outputs

The model is mainly used to forecast:

• PM2.5
• PM10

These continuous predictions can also be mapped to AQI-style discrete categories such as Good, Moderate, Bad, and VeryBad.

Out-of-Scope Use

This model is not intended for:

• Direct public warning issuance without local validation and expert review
• Deployment outside the East Asia domain used in the paper
• Use with input data that does not match the official preprocessing and directory structure
• Clinical, legal, or regulatory decision-making without independent verification

Training Data

FAKER-Air is trained on the CMAQ-OBS Regional Air Quality Dataset for East Asia.

The released dataset combines:

• Observation fields under data/obs
• CMAQ reanalysis fields under data/cmaq

The dataset is built for East Asia at 27 km spatial resolution and 6-hour temporal intervals, combining grid-aligned observation fields with region-specific CMAQ reanalysis.

Model Inputs and Outputs

Inputs

The official pipeline expects preprocessed inputs matching the repository format.

In practice, inference uses:

• Gridded OBS inputs (.npz)
• CMAQ concentration fields (*_x_conc.npy)
• CMAQ 2D meteorological fields (*_x_metcro2d.npy)
• CMAQ 3D meteorological fields (*_x_metcro3d.npy)

Please follow the preprocessing and directory structure in the official repository before running evaluation or inference.

Outputs

The model predicts long-horizon particulate matter fields over the East Asia forecasting domain, especially:

• PM2.5 concentration fields
• PM10 concentration fields

These outputs can be further converted into AQI-style classes for operational analysis.

Training Procedure

The official training recipe uses a two-stage setup:

• Stage 1: SFT with Temporal Accumulation Loss
• Stage 2: GRPO with class-wise AQI rewards and curriculum rollout

This design is intended to combine strong spatiotemporal forecasting performance with improved operational reliability.

Evaluation

The paper evaluates long-horizon forecasting up to 120 hours ahead.

Selected Reported Results for PM2.5 at 120h

• Binary F1: 56.72
• False Alarm Rate (FAR): 17.32
• 4-class Accuracy: 45.16
• 4-class F1-macro: 41.90

Compared with the SFT baseline, the GRPO model is reported to reduce PM2.5 FAR from 32.86 to 17.32, while maintaining competitive F1.
For PM10, the paper reports a reduction in FAR from 18.44 to 10.81 in the 120h setting.

For full results, please refer to the paper and supplementary material.

Limitations

• This checkpoint is region-specific and is designed for East Asia, not for global forecasting.
• Performance depends on using the same preprocessing, variables, grid definition, and rollout setup as the official codebase.
• The model is optimized for reliability-aware forecasting, so some operating points may trade raw point-wise accuracy for lower false alarm rates.
• Performance may degrade under severe domain shift, missing local inputs, or deployment in regions outside the training domain.

How to Use

1) Download the checkpoint

from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="2na-97/FAKER-Air",
    filename="FAKER-Air_GRPO",  # change this if you rename the checkpoint file
)

print(ckpt_path)

2) Prepare the dataset

Dataset:

• 2na-97/FAKER-Air

Code:

• https://github.com/kaist-cvml/FAKER-Air/tree/main

Follow the repository instructions to place OBS and CMAQ files in the expected directories.

3) Run rollout evaluation / inference

CUDA_VISIBLE_DEVICES=0 \
torchrun \
  --nproc_per_node=1 \
  --master_addr="127.0.0.1" \
  --master_port=29502 \
  test.py --batch 1 \
  --model aurora \
  --test-start-date 2023-01-01 \
  --test-end-date 2023-12-31 \
  --data-sources obs,cmaq \
  --checkpoint-path /path/to/FAKER-Air_GRPO \
  --npz-path ./data/obs_npz_27km \
  --cmaq-root ./data/cmaq_only_npy \
  --mode rollout \
  --rollout-hours 120 \
  --use_cmaq_pm_only

Repository Links

• Code: https://github.com/kaist-cvml/FAKER-Air/tree/main
• Dataset: https://huggingface.co/datasets/2na-97/FAKER-Air
• Model: https://huggingface.co/2na-97/FAKER-Air

Citation

@article{kang2026fakerair,
  title={Real-Time Long Horizon Air Quality Forecasting via Group-Relative Policy Optimization},
  author={Kang, Inha and Kim, Eunki and Ryu, Wonjeong and Shin, Jaeyo and Yu, Seungjun and Kang, Yoon-Hee and Jeong, Seongeun and Kim, Eunhye and Kim, Soontae and Shim, Hyunjung},
  journal={arXiv preprint arXiv:2511.22169},
  year={2026}
}
``