HeatCast-W34 Distributional Mesh GNN
HeatCast is a 4,637,891-parameter GraphCast-style mesh graph neural
network for probabilistic warm-season CONUS week-3--4 maximum-temperature
prediction.
Outputs
For every PRISM-grid land cell and lead from day 15 through day 28, HeatCast
produces a Gaussian conditional mean and state-dependent standard deviation.
It does not produce dynamically independent ensemble members.
Files
Each folds/foldN/ directory contains the selected best-monitor model weights
in Safetensors format and the checkpoint architecture metadata. Five
year-disjoint folds are provided.
Intended use
Research on retrospective CONUS MJJAS week-3--4 temperature prediction and
heat-exceedance calibration.
Important limitations
-
These are research hindcast models, not an operational warning system.
-
Training and evaluation target PRISM daily maximum temperature over CONUS.
-
Exact inference requires the HeatCast preprocessing inputs and normalization
conventions described in
MODEL_INPUTS.md. -
PRISM, ERA5, ECMWF ENS, and other external datasets are not distributed here.
-
HeatCast weights provide mean and sigma. HeatCast-C probability calibration
and the HeatCast-plus-ENS stack require additional fold-safe calibration and
ECMWF ENS artifacts.
-
Performance outside CONUS, May through September, or the historical record
has not been established.
Architecture
The model uses grid-to-mesh encoding, an icosahedral multimesh message-passing
processor, multi-lead tube attention, mesh-to-grid decoding, and a
distributional head trained using Gaussian CRPS.
License
MIT for the distributed HeatCast code and weights. External datasets retain
their original terms and licenses.