WaveCastNet

Overview

wavecastnet is the PyHazards entrypoint for dense-grid earthquake wavefield forecasting based on the ConvLEM encoder-decoder design described by Lyu et al. (2025).

At a Glance

Hazard Family

Earthquake

Public catalog grouping used for this model.

Maturity

Implemented

Catalog maturity label used on the index page.

Tasks

1

Wavefield Forecasting

Benchmark Family

Earthquake Benchmark

Primary benchmark-family link used for compatible evaluation coverage.

Description

wavecastnet is the PyHazards entrypoint for dense-grid earthquake wavefield forecasting based on the ConvLEM encoder-decoder design described by Lyu et al. (2025).

This implementation focuses on the core dense-grid forecasting path and keeps data loading outside the model so users can adapt it to their own simulation or sensor pipelines.

Benchmark Compatibility

Primary benchmark family: Earthquake Benchmark

Mapped benchmark ecosystems: AEFA, pyCSEP

External References

Paper: Rapid wavefield forecasting for earthquake early warning via deep sequence to sequence learning

Registry Name

Primary entrypoint: wavecastnet

Supported Tasks

• Wavefield Forecasting

Programmatic Use

import torch
from pyhazards.models import build_model

model = build_model(
    name="wavecastnet",
    task="regression",
    in_channels=3,
    height=32,
    width=24,
    temporal_in=6,
    temporal_out=4,
    hidden_dim=32,
    num_layers=1,
    dropout=0.0,
)

x = torch.randn(2, 3, 6, 32, 24)
y = model(x)
print(y.shape)

Notes

• The PyHazards version currently targets dense-grid forecasting rather than the paper’s sparse-sensor variants.

• The smoke test uses reduced spatial and temporal sizes so it stays CPU-safe in CI.