ntk-uq-weather 0.1.0

Last-layer empirical Neural Tangent Kernel uncertainty quantification for AI weather models

NTK-UQ-Weather-Models

Last-layer empirical Neural Tangent Kernel (NTK) uncertainty quantification for pre-trained AI weather forecasting models. This is the reference implementation for:

J. M. A. Miñoza, R. G. Laylo, and S. C. Ibañez. "Scalable Uncertainty Quantification for Extreme Weather Forecasting via Empirical Neural Tangent Kernels." Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '26), 2026.

The method is post-hoc: it adds calibrated, spatially-adaptive uncertainty to any frozen deterministic weather model, with no retraining and a single matrix-vector product at inference. It treats a model's last-layer features as an empirical NTK, builds the feature kernel, decomposes it (SVD or ICA), and returns the Gaussian-process posterior variance.

What this repo gives you

1. Per-model last-layer feature extractors for four production weather models (the centerpiece — how to pull empirical-NTK features from each architecture).
2. NTKCalibrator — kernel construction, SVD/ICA decomposition, GP posterior, and per-variable post-hoc scaling.
3. The EM-DAT initialization-date list (n = 100) used in the paper, plus the extraction utility.

Install

git clone https://github.com/JomaMinoza/NTK-UQ-Weather-Models.git
cd NTK-UQ-Weather-Models
pip install -e .

The core install (kernel + calibration + GP posterior) needs only torch, numpy, scipy, scikit-learn. Each weather model needs its own backend — install only the ones you use:

Model	Architecture	Backend (extra)	Features global_avg / multi_stat
FourCastNetV2	SFNO	pip install -e ".[fourcastnet]"	256 / 1536
Pangu-Weather	3D Swin Transformer (ONNX)	pip install -e ".[pangu]"	69 / 414
Aurora	Swin backbone + Perceiver I/O	pip install -e ".[aurora]"	65 / 390
AIFS	GNN encoder + Transformer	pip install -e ".[aifs]"	1024 / 6144

EM-DAT date extraction needs pandas/openpyxl (pip install -e ".[data]"). Install everything with pip install -e ".[all]".

Quickstart (no model weights needed)

The empirical-NTK calibration runs on plain feature arrays. This demonstrates the core property — inputs dissimilar to the calibration set get higher uncertainty:

python examples/01_calibrate_synthetic.py
# Mean uncertainty, in-distribution test inputs : 0.03
# Mean uncertainty, out-of-distribution inputs  : 1.55
# Ratio (out / in)                              : 47x

import torch
from ntk_uq import NTKCalibrator

features = torch.randn(100, 256)              # last-layer calibration features
cal = NTKCalibrator(model_name="demo", rank_k=10, device="cpu")
cal.calibrate_lead_time(features, lead_time_hours=24)

sigma = cal.compute_uncertainty(torch.randn(8, 256), lead_time_hours=24)
print(sigma["uncertainty"])                   # GP posterior std per sample

Extracting empirical-NTK features from a weather model

Every extractor follows the same recipe: load the checkpoint, register a forward hook on the last layer, roll out to the lead time, and aggregate into the dual feature vectors (global_avg for spatial UQ, multi_stat for scalar UQ). See examples/02_extract_features.py for all four. For example, FourCastNetV2:

from ntk_uq.features import FourCastNetFeatureExtractor

ext = FourCastNetFeatureExtractor(device="cuda")
ext.load_model("path/to/fourcastnetv2/weights")
feats = ext.extract_features_dual(x, lead_time_hours=24)
# feats == {"global_avg": tensor(256,), "multi_stat": tensor(1536,)}

AIFS pulls ERA5 by date through the Anemoi runner, so it is date-based:

from ntk_uq.features import AIFSFeatureExtractor

ext = AIFSFeatureExtractor(device="cuda")
ext.load_model("ecmwf/aifs-single-1.1")
feats = ext.extract_features_from_date_dual("2021-12-16", lead_time_hours=24)

EM-DAT initialization dates

from ntk_uq.data import load_calibration_dates, load_precursor_pool

dates = load_calibration_dates()   # 100 init dates (paper n), bundled with the package
pool = load_precursor_pool()       # 208-date EM-DAT precursor pool (3-day lookback)

The 100 calibration dates are a 3-day lookback before each EM-DAT event onset, deduplicated and selected from the 208-date precursor pool. EM-DAT itself is not redistributed; download it from https://www.emdat.be/ and rebuild the pool with ntk_uq.data.extract_precursor_dates(path_to_xlsx).

Paper ↔ code mapping

Paper	Code
Last-layer feature kernel K(x,x') = φ(x)ᵀφ(x')	NTKCalibrator.calibrate_lead_time
SVD / ICA decomposition (Section 3.4)	NTKCalibrator._calibrate_svd / _calibrate_ica
GP posterior variance (Eq. 4)	NTKCalibrator.compute_uncertainty
Dual features: global_avg, multi_stat	*.extract_features_dual
Post-hoc scaling α (Section 3.5)	NTKCalibrator.calibrate_all_lead_times
FourCastNetV2 (SFNO)	FourCastNetFeatureExtractor
Pangu-Weather	PanguWeatherFeatureExtractor
Aurora	AuroraFeatureExtractor
AIFS	AIFSFeatureExtractor
EM-DAT events, n = 100	ntk_uq.data.load_calibration_dates()

Tests

pip install pytest
pytest tests/

The tests exercise the kernel/decomposition/GP path, the discrimination property, save/load, and the metrics — all without weather-model weights.

Citation

@inproceedings{minoza2026ntkuq,
  title     = {Scalable Uncertainty Quantification for Extreme Weather Forecasting via Empirical Neural Tangent Kernels},
  author    = {Mi{\~n}oza, Jose Marie Antonio and Laylo, Rex Gregor and Iba{\~n}ez, Sebastian C.},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '26)},
  year      = {2026},
  doi       = {10.1145/3770855.3818106}
}

License

Apache License 2.0 — see LICENSE and NOTICE. Copyright 2026 Department of Education (Center for AI Research), Philippines.