models/translocation.md

Gray-swan translocation: forecast diagnostics

Model pages 路 Three-paper integration

Sun et al., Predicting regional gray swans via translocation: AI weather models and Dubai's unprecedented 2024 rainfall, Science Advances, 7 October 2026. Paper.

Classification: extreme-event generalization and interpretability. This paper does not introduce a sixth/new forecasting architecture. It evaluates GraphCast, FuXi, AIFS, Pangu-Weather and FourCastNet-v2, and investigates GraphCast/AIFS effective receptive fields (ERFs) and fine-tuning experiments.

Official resources from the paper:

Resource	Code / model
GraphCast	DeepMind code and weight links
FuXi	Original cascade code (pinned), author ONNX models / sample data
AIFS	ecmwf/aifs-single-1.0
Pangu	Original author code and four ONNX weight links, 6h checkpoint
FourCastNet-v2	NVIDIA inference code, official NGC v0.2 weights + statistics
GraphCast fine-tuning	csubich/graphcast
Dynamically similar event tracking	cyclonetracking
Paper ERF notebook and data	Zenodo 18111128

Zenodo lists the 19,427-byte notebook graphcast-ERF-cleaned.ipynb, README, and three ERF data ZIP files (~3.14, 2.03 and 2.41 GB). These are diagnostics, not newly released model checkpoints. Model licenses follow each backbone. The current FuXi repository advertises FuXi 2.1; that newer release must not be silently substituted for the original cascade used in this study. The table therefore links the original code snapshot and its author-hosted archive. Only the GraphCast route was executed for this integration; the other official backbone artifacts are discovery references, not newly validated implementations.

from weatherai import Translocation
model = Translocation.from_pretrained()  # official 0.25-degree, 13-level GraphCast
inputs, targets, forcings = model.prepare(era5_dataset, lead_time="6h")
forecast = model(inputs, targets * float("nan"), forcings)
erf = model.effective_receptive_field(
    inputs, targets_template=targets * float("nan"), forcings=forcings,
    variable="total_precipitation_6hr",
    location={"batch": 0, "time": -1, "lat": 260, "lon": 220})

The main entry downloads the official GraphCast checkpoint and normalization statistics from dm_graphcast and imports unchanged author code at 08cf73625c9d12bd9aaa038868bcb2fe488f2a22. This historical revision retains the graphcast package before the upstream repository's WeatherNext reorganization. Variants are operational (0.25°, 13 levels), 37levels (0.25°, 37 levels), and small (official pretrained 1°, 13 levels; not the paper's main resolution). Code is Apache-2.0; GraphCast weights retain CC-BY-NC-SA-4.0.

The wrapper order follows the study's released ERF notebook: GraphCast, Bfloat16Cast, InputsAndResiduals and autoregressive.Predictor. Signed input sensitivities use JAX reverse-mode AD on a selected physical output. Coordinates, variable, lead and normalization must match the intended event. ERF alone does not establish causality. Full 0.25° operational forecasting and the actual Small-model ERF adapter passed ZeroGPU execution. Current test evidence.

Translocation_lite() retains the independent random-weight PyTorch toy model. For custom PyTorch diagnostic experiments, import TranslocationLite from weatherai.models.translocation. Neither is the pretrained main entry. The complete Dubai reconstruction, event tracking and fine-tuning ablations have not been reproduced. No study-specific fine-tuned weights were found in the cited ERF archive; its large ZIP files contain diagnostic data.