THULab/elastic_rod_dataset

Discrete Elastic Rods Simulation Dataset (TsFile)

Apache TsFile version of SamuelRx/Elastic-Rod-Dataset.

Overview

This dataset contains synthetic data generated from Discrete Elastic Rods (DER) simulations, a physical model used to represent deformable slender structures such as hair strands, ropes, cables, and elastic fibers. The data was generated frame-by-frame during physical simulations and stores geometric, kinematic, and dynamic properties for each rod vertex.

The primary objective of this dataset is to support machine learning research involving graph neural networks (GNNs), physics-informed learning, dynamics prediction, physical regression, and deformable-object simulation. Each sample corresponds to a vertex at a specific simulation frame.

• Scale: Train 4,265,580 samples (2 shards merged) · Validation 376,200 · Test 460,920 · 209 strands × 15 vertices = 3,135 simulated vertex trajectories.
• Coverage: Frames 0–1347 of randomized Discrete Elastic Rods simulations; source is Hair for all rows.
• Per-record meaning: one vertex of one elastic rod at one simulation frame, with position, velocity, force, curvature, torsion, and the adjacent segment direction/length.

Schema (TsFile structure)

• Time (INT64, milliseconds) — normalized from the source timestamp (timestamp[us]), e.g. 2026-05-06 00:51:35.711.
• source (TAG) — simulation source label (Hair in this dataset). Query with e.g. WHERE source='Hair'.
• strand (TAG) — rod/strand identifier (0–208).
• vertex_id (TAG) — vertex index along the rod (0–14). Together (source, strand, vertex_id) identifies one vertex trajectory (device).
• frame (FIELD, INT64) — simulation frame index.
• pos_x / pos_y / pos_z, vel_x / vel_y / vel_z, force_x / force_y / force_z (FIELD, DOUBLE) — vertex position, velocity, and applied force.
• curvature (FIELD, DOUBLE), torsion (FIELD, INT64) — local curvature and torsion.
• prev_segment_direction_0.._2, next_segment_direction_0.._2 (FIELD, DOUBLE) — the source list columns prev_segment_direction / next_segment_direction exploded into three scalar components each. The first vertex of each strand has no previous segment, so its prev_segment_direction_* values are null.
• prev_segment_length, next_segment_length (FIELD, DOUBLE) — adjacent segment lengths.
• boundary (FIELD, STRING) — boundary condition (clamped for the first vertex of each strand, free otherwise).

The source timestamp column is represented losslessly by Time (microseconds truncated to milliseconds); it is not duplicated as a separate field. No source column is dropped.

Usage

Install the Apache TsFile Python SDK (pip install tsfile) and read a converted file:

from pathlib import Path
from tsfile import TsFileReader

path = Path("elastic_rod_dataset_test.tsfile")
with TsFileReader(str(path)) as reader:
    schemas = reader.get_all_table_schemas()
    print("tables:", list(schemas))
    table_name = next(iter(schemas))
    table = schemas[table_name]
    columns = [column.get_column_name() for column in table.get_columns()]
    print("columns:", columns)
    field_names = [
        column.get_column_name()
        for column in table.get_columns()
        if column.get_column_name() not in {"Time", "time"}
    ]
    if field_names:
        with reader.query_table(table_name, field_names[:3], batch_size=1024) as result:
            batch = result.read_arrow_batch()
            if batch is not None:
                print(batch.to_pandas().head())

Source & license

• Original dataset: https://huggingface.co/datasets/SamuelRx/Elastic-Rod-Dataset
• Author / publisher: Samuel Ferreira Santos
• Paper: none declared by the original dataset
• License: apache-2.0
• Citation (from the original card):

@dataset{der_simulation_dataset,
  title={Discrete Elastic Rods Simulation Dataset},
  author={Samuel Ferreira Santos},
  year={2026}
}