Synthetic Gravitational Geohazard Mass Flow Runout Dataset for Deep Learning

This dataset contains approximately 120,000 synthetic landslide simulations generated using a physics-based numerical model and preprocessed for machine learning applications. The dataset is designed to support the development of deep learning models for predicting landslide runout extent and deposit thickness from terrain and initial conditions. 

Each sample represents a single simulation over a fixed grid and includes both input features (terrain and initial source) and target outputs (final deposit thickness), along with the governing physical parameters.

Content

- dataset.zip → contains ~120k samples (.npz)- split_cached.json → train/val/test split

Dataset Structure

The dataset is provided as compressed .npz files, organized in multiple .zip archives for efficient distribution.

Each .npz file contains the following variables:

dem (256 × 256)Digital Elevation Model (topography) of the simulation domain (normalised [0, 1]).

source (256 × 256) (meters)Initial thickness distribution (h₀) representing the landslide source area.

h_final (256 × 256) (meters)Final deposit thickness after simulation.

volume (scalar) (normalised [0, 1])Initial landslide volume.

density (scalar) (normalised [0, 1])Material density.

cohesion (scalar) (normalised [0, 1])Material cohesion.

Spatial and Physical Properties

Grid size: 256 × 256 pixels

Spatial resolution: ~30 m

Domain size: ~7.5 km × 7.5 km

Units:

Thickness: meters

Volume: cubic meters

Density: kg/m³

Cohesion: Pascals

Data Characteristics

The dataset is highly sparse, with many simulations producing little or no runout.

Thickness fields exhibit heavy-tailed distributions, with most pixels near zero and localized high values.

Initial sources are centrally located and vary in size according to volume.

Terrain is derived from real-world DEM patches.

Intended Use

This dataset is intended for:

Training deep learning models for landslide runout prediction

Surrogate modeling of physics-based simulations

Benchmarking segmentation and regression models on geophysical data

Studying uncertainty and probabilistic runout behavior

Limitations

The dataset is entirely synthetic and based on a specific numerical model.

No direct calibration to real-world events is included.

Many samples contain minimal or no flow, which may bias naive training approaches.File Organizationdata/  dataset_part_000.zip  dataset_part_001.zip  ...

Each archive contains thousands of .npz samples.