LarNO — Latent Autoregressive Neural Operator

Paper: Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO Journal: Journal of Hydrology, 2026 DOI: 10.1016/j.jhydrol.2026.135686 GitHub: holmescao/LarNO Colab Demo:

---

Model Description

LarNO is a memory-efficient, discretization-invariant neural operator for large-scale urban flood spatiotemporal forecasting. It learns continuous-space hydrodynamic mappings to predict water depth distributions based on dynamic rainfall and static terrain.

Key capabilities:

• Zero-shot super-resolution: trained at 20 m resolution, applied at 5 m with no retraining
• ~940× faster inference than MIKE+ hydraulic solver (with TensorRT)
• Few-shot transfer to unseen catchments via fine-tuning
• Large-scale: handles ~100 km² at 5 m / 5 min resolution

---

Performance

Benchmarked on Futian district, Shenzhen (~100 km²) — 5 m resolution, zero-shot super-resolution (trained at 20 m):

Method	Params	Inference†	Speedup	MAE (m) ↓	CSI ↑
MIKE+ (hydraulic solver)	—	~8.9 h	1×	Reference	Reference
UNO	109.1 M	710 s	~570×	0.024 ± 0.007	0.343 ± 0.026
FNO	29.1 M	760 s	~530×	0.019 ± 0.004	0.620 ± 0.027
LarNO (ours)	29.1 M	34 s ‡	~940×	0.008 ± 0.003	0.722 ± 0.016

† Single 6-hour event on NVIDIA RTX 4090. ‡ LarNO inference uses TensorRT (TRT) acceleration; UNO and FNO do not support TRT. Note: The released dataset is a 20 m downsampled version for accessibility. Metrics on the released 20 m data will differ from the 5 m paper results above.

---

Model Architecture

• Backbone: TFNO2d (Factorized Fourier Neural Operator)
• Temporal memory: CGRU (Convolutional GRU) for latent autoregression
• Input channels: 13 (6 rainfall + 6 cumulative-rainfall + 1 DEM)
• Output: water depth field (H × W), one time step at a time

Checkpoint architecture (must match for loading):

Parameter	Value
hidden_channels	32
n_modes_height	100
n_modes_width	140
n_layers	4

---

Usage

Quick inference (no local setup)

Local inference

git clone https://github.com/holmescao/LarNO
cd LarNO/code/urbanflood_larfno
pip install -e .
pip install -r requirements.txt

# Download weights from this HuggingFace repo or Google Drive
# Place under: LarNO/exp/<expr_id>/weights/<checkpoint_name>/

python test.py --config urbanflood_config_2d.yaml --expr_id <expr_id>

See the GitHub README for full installation and training instructions.

---

Files

File	Description
exp_weights.zip	Pre-trained Futian (region1_20m) checkpoint, hidden_ch=32, epoch ~992

Directory structure after unzipping:

exp/
└── <expr_id>/
    └── weights/
        └── <checkpoint_name>/
            └── <checkpoint_name>_state_dict.pt

---

Dataset

The benchmark dataset is available at holmescao/LarNO-dataset.

Dataset	Resolution	Area	Events
UKEA (ukea_8m_5min)	8 m / 5 min	~0.4 km²	8 train + 12 test
Futian (region1_20m)	20 m / 5 min	~100 km²	64 train + 16 test

---

Citation

@article{cao2026large,
  title={Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO},
  author={Cao, Xiaoyan and Yao, Yao and Wang, Zhi and Zhao, Zhangxinyue and Borthwick, Alistair GL and Qin, Huapeng},
  journal={Journal of Hydrology},
  pages={135686},
  year={2026},
  doi={10.1016/j.jhydrol.2026.135686},
  publisher={Elsevier}
}