CEED

CEED: California Earthquake Event Dataset

CEED contains 2.8 million three-component seismograms of earthquakes in California, 1977–2025, each labelled with analyst-reviewed P and S arrival times. The waveforms and labels are drawn from the NCEDC (Northern California Earthquake Data Center) and the SCEDC (Southern California Earthquake Data Center). Each record also carries the network's hypocentre, magnitude, pick weights and, where reported, the P-wave first-motion polarity and focal mechanism, so that subsets can be selected by label quality.

Figure 1. Example records. Blue and orange lines mark the analyst P and S picks.

Data description

A record is one earthquake observed at one station. The waveform is a (3, 12288) float32 array sampled at 100 Hz (122.88 s, beginning 40.96 s before the origin time), in channel order E, N, Z and in µm/s where the instrument response is available. Stations with fewer than three components are retained; component lists the channels that contain data, and the others are zero.

A trace is included when both the P and the S pick are manual (evaluation_mode == "manual") and S follows P. Within each SCEDC event, traces are grouped in threes by epicentral distance and one per group is kept, preferring a trace with a P first motion; this brings the number of traces per event close to NCEDC's while retaining the full distance range. The train split covers years up to 2023 and test covers 2024–2025. Because each earthquake falls within one year, no event appears in both splits. The current, incomplete year is excluded.

Usage

from datasets import load_dataset
import numpy as np

ds = load_dataset("AI4EPS/CEED", "scedc", split="train")
r = ds[0]
w = np.asarray(r["waveform"], dtype="float32")      # (3, 12288), channels E, N, Z at 100 Hz
p, s = r["p_phase_index"], r["s_phase_index"]       # arrivals, as sample offsets
print(r["event_id"], f"M{r['event_magnitude']}", f"{r['distance_km']:.0f} km",
      "first motion", r["p_phase_polarity"])
window = w[:, p - 100 : p + 400]                    # 1 s before to 4 s after P

The release is 208 GiB; streaming avoids downloading it in full:

ds = load_dataset("AI4EPS/CEED", "scedc", split="train", streaming=True)
for r in ds.take(10):
    ...

Each year also provides index.parquet, which holds every field except the waveform, so rows can be selected before the waveform shards are read:

import pandas as pd

idx = pd.read_parquet("hf://datasets/AI4EPS/CEED/scedc/2019/index.parquet")
want = idx[(idx.event_magnitude > 4) & (idx.distance_km < 50) & (idx.component == "ENZ")]

shard = pd.read_parquet("hf://datasets/AI4EPS/CEED/scedc/2019/0000.parquet")
rows = shard.iloc[want[want.shard_id == 0].row_in_shard.values]

ncedc and scedc are separate configurations. Because traces of the same earthquake are strongly correlated, any further split should be made by event rather than by trace.

Fields

The full schema has 69 columns; the remainder are the networks' location-quality parameters (arrival residuals and weights, azimuthal gap and hypocentre uncertainties).

Statistics

Figure 2. Epicentres and stations.

Figure 3. Traces per year and region; the dashed line marks the start of the test split.

Limitations

• Regional balance. The SCEDC:NCEDC trace ratio is 1.52 overall but varies by year, from 0.51 (2014) to 11 (1992).
• First motions. P polarity is reported for 93% of NCEDC and 62% of SCEDC traces. SCSN has reported it far less often since 2013 (81% of SCEDC traces before 2013, 37% since), so recent first-motion data are mostly from NCEDC: the test split holds 61,938 NCEDC and 24,734 SCEDC traces with a first motion.
• Takeoff angles are not available for SCEDC before 2007.
• Network conventions differ. Magnitudes are mostly Md in NCEDC and Ml in SCEDC (event_magnitude_type); pick weights and focal-mechanism uncertainty columns follow each network's convention and should not be compared across regions.
• Duplicate events. Earthquakes near the regional boundary (about 35–36°N) may be catalogued by both networks under different event_ids; de-duplicate before cross-region evaluation.

Provenance

Waveforms were retrieved from the public NCEDC and SCEDC archives on AWS S3 and picks from their FDSN event services. Instrument responses were removed where station metadata allow. The dataset was built with QuakeFlow (datasets/build_ceed.py).

Citation

Please cite the dataset paper and the two data centres:

@article{zhu2025ceed,
  title   = {California Earthquake Dataset for Machine Learning and Cloud Computing},
  author  = {Zhu, Weiqiang and Wang, Haoyu and Rong, Bo and Yu, Ellen and Zuzlewski, Stephane and Tepp, Gabrielle and Taira, Taka'aki and Marty, Julien and Husker, Allen and Allen, Richard M.},
  journal = {arXiv preprint arXiv:2502.11500},
  year    = {2025}
}

NCEDC (2014), Northern California Earthquake Data Center. UC Berkeley Seismological Laboratory. Dataset. doi:10.7932/NCEDC

SCEDC (2013), Southern California Earthquake Data Center. Caltech. Dataset. doi:10.7909/C3WD3xH1