MIGRATE/Surfdisp96-Roccastrada-10k-2025

🧠 Dataset Overview

The Surfdisp96-Roccastrada-10m dataset contains synthetic seismic velocity models and their corresponding Rayleigh-wave dispersion curves, generated using the SurfDisp96 simulator with Roccastrada priors.
It is designed for benchmarking seismic inversion algorithms and for training machine-learning models in geophysics.

Each sample includes:

1. A seismic velocity model with shear-wave velocities, depths and layer thicknesses
2. Three dispersion curves covering low, middle and high period ranges

🔢 Priors Configuration (Roccastrada)

Parameter	Value	Description
vs	[0.5, 4.0]	Range of shear-wave velocities (km/s)
z	[0.0, 5.0]	Depth range (km)
layers	[2, 20]	Minimum and maximum number of layers
vpvs	1.3	P/S velocity ratio
mohoest	null	Estimated Moho depth (unused)
mantle	null	Mantle properties (unused)
thickmin	0.1	Minimum layer thickness (km)
lvz	null	Low-velocity zone (not used)
hvz	null	High-velocity zone (not used)

⚙️ Generation Parameters

Parameter	Value	Description
seed	43	Random seed for reproducibility
random_generator	numpy.default_rng	Random number generator backend
n_samples	10 000	Total synthetic models
samples_per_shard	1 000	Number of samples per Parquet shard
n_shards	10	Total number of shards
source	sample_model + forward	Data generation process
dispersion_curve_length	108	Number of points per dispersion curve
folds	2-fold, 5-fold, 10-fold	Available cross-validation splits
fold_file	folds.json	JSON file defining the folds

🧩 Feature Schema

Feature	Type	Description
vs	list	Shear-wave velocities (km/s), one per layer
z	list	Depths of layer boundaries (km)
vpvs	float32	Vp/Vs ratio
disp_x	list	Period values (s), length = 108
disp_y	list	Corresponding velocities (km/s), length = 108
wave_type	string	Type of surface wave (Rayleigh / Love)
velocity_type	string	Velocity measurement type (group / phase)

📊 Column Descriptions

Column Name	Data Type	Description	Typical Values
Model Parameters
vs	list	Shear-wave velocities (km/s)	0.5–5.0
z	list	Layer-center depths (km)	0–15
h	list	Layer thicknesses (km)	depends on prior
z_disc	list	Discontinuity depths (km)	depends on prior
vp	list	P-wave velocities (km/s)	1.5–8.5
vpvs	float32	Vp/Vs ratio	1.3
nlayers	int	Number of layers	2–20
velmap_vs	list	Interpolated Vs profile (60 points)	0.5–5.0
velmap_z	list	Depth grid for velmap_vs	0–15 km
Low Range Dispersion Curve
L_disp_x	list	Periods (s)	0.1–1.0
L_disp_y	list	Velocities (km/s)	0.5–3.0
L_wave_type	string	Wave type	Rayleigh / Love
L_velocity_type	string	Velocity type	group / phase
Middle Range Dispersion Curve
M_disp_x	list	Periods (s)	1.0–10.0
M_disp_y	list	Velocities (km/s)	1.0–3.5
M_wave_type	string	Wave type	Rayleigh / Love
M_velocity_type	string	Velocity type	group / phase
High Range Dispersion Curve
H_disp_x	list	Periods (s)	10.0–40.0
H_disp_y	list	Velocities (km/s)	2.0–4.0
H_wave_type	string	Wave type	Rayleigh / Love
H_velocity_type	string	Velocity type	group / phase

🧭 Notes

1. All arrays are stored as float32 to reduce storage size.
2. Dispersion curves represent the relationship between period (T) and velocity (v) for seismic surface waves.
3. Low, middle and high ranges enable multi-scale analysis of the subsurface structure.
4. Default wave type = “Rayleigh”, velocity type = “group”.
5. Period ranges and grid spacing depend on generation parameters (low_range, middle_range, high_range, variable_grid).
6. Folds allow robust cross-validation (2, 5, 10 folds).

🧮 Usage Example

With 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("nils-schaetti/sd96-roccastrada-10m", split="train")
print(ds[0]["vs"])  # Access shear-wave velocity model

Via CLI

huggingface-cli download dataset nils-schaetti/sd96-roccastrada-10m --local-dir ./sd96-roccastrada-10m

🖼️ Sample Visualization

A sample model and its three dispersion curves are illustrated in sample_plot.png within the dataset directory.

⚙️ Generation Command

python3 migrate/cli/main.py generate-dataset-surfdisp96 \
  --name Surfdisp96-Roccastrada-10m \
  --pretty-name sd96-roccastrada-10m \
  --description "This dataset contains synthetic seismic models and their corresponding Rayleigh-wave dispersion curves generated using forward modeling with the Roccastrada priors. It is designed for benchmarking inversion algorithms and training machine learning models in geophysics." \
  --license-name "CC BY-SA 4.0" \
  --created-by "Nils Schaetti" \
  --prior-file conf/priors/roccastrada_prior.yaml \
  --output-dir data/seismic/Dispsurf96-Roccastrada-10k \
  --n-samples 10000 \
  --samples-per-shard 1000 \
  --length 108 \
  --test-ratio 0.2 \
  --folds 2 5 10 \
  --seed 43 \
  --low-range 1.0 5.0 \
  --middle-range 1.0 15.0 \
  --high-range 1.0 30.0

🧾 License

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

© 2025 Nils Schaetti DMML – Data Mining and Machine Learning Group Haute École de Gestion de Genève (HES-SO) 📧 nils.schaetti@hesge.ch

You are free to use, share, and adapt this dataset for any purpose, including commercial use, provided that you:

• Attribute the creator (Nils Schaetti, DMML Group, HEG Genève)
• Share-alike any derivative work under the same license (CC BY-SA 4.0)