jongwonryu/TIDE-TIDES

TIDE / TIDES: Initial Sphere Dataset Release

Dataset artifacts for Interactive 4D Volumetric Liquid Forecasting under Moving-Solid Interaction, by Jongwon Ryu, Sunjae Yoon, and Junyeong Kim.

Paper | Framework code

Release Scope

This is an initial partial release of the locally preserved sphere data, not the complete benchmark reported in the paper. It contains 200 main episodes and 40 separate sphere size-OOD episodes. Cube/cone data and the full paper-scale episode counts are not included. This subset alone does not reproduce the paper's aggregate tables. Journal-extension datasets and pretrained checkpoints are not part of this release.

See RELEASE_STATUS.json for upload completion. While its status is upload_in_progress, some indexed binary files may not yet be downloadable.

The original dense and particle .npz files and rigid-pose CSVs are preserved byte-for-byte. Split assignments are unchanged. JSON metadata has been made portable by replacing server-specific paths with repository-relative paths. release_manifest.json and SHA256SUMS record the binary file inventory and SHA-256 digests. The duplicate v1 processing version is intentionally excluded.

Available Episodes

Configuration	Train	Validation	Test	Motion OOD	Size OOD
sphere	128	16	16	40	-
sphere_size_ood	-	-	-	-	40

Original manifests use val; the Hugging Face metadata index exposes that split as validation. Both OOD configurations use test_ood, distinguished by the configuration name. Each record describes a complete simulation episode, not an individual frame or an independently generated rollout window.

Dense grids are 32 x 128 x 128 in depth/height/width order, sampled at 8 FPS. Episode lengths vary; consult num_frames. The main sphere radius is 4 voxels; size-OOD spheres use 8 voxels. Coordinates and voxel spacing are stored per episode; do not assume identical world bounds across cases.

Load the Lightweight Episode Index

from datasets import load_dataset

episodes = load_dataset("jongwonryu/TIDE-TIDES", "sphere")
record = episodes["train"][0]
print(record["episode_id"], record["file"], record["num_frames"])

size_ood = load_dataset("jongwonryu/TIDE-TIDES", "sphere_size_ood")

load_dataset downloads episode metadata, not the dense velocity tensors. The Hub viewer likewise browses the episode index rather than rendering the 3D flow. Download the selected NPZ explicitly:

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="jongwonryu/TIDE-TIDES",
    repo_type="dataset",
    filename=record["file"],
)
with np.load(path, allow_pickle=False) as episode:
    initial_velocity = episode["init_flow"]
    dt = episode["dt"]
    spacing = episode["voxel_size_xyz"]

NPZ members are decompressed when accessed; full sequences can require several GB of RAM. Opening an archive does not require loading every array. Original arrays are retained rather than silently changing precision or resampling.

Download for the Released Training Code

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="jongwonryu/TIDE-TIDES",
    repo_type="dataset",
    local_dir="TIDE-TIDES-data",
)

The complete partial release, including dense samples and particle exports, is approximately 152 GB (decimal). Download individual episodes or use allow_patterns if storage is limited. For the framework's existing loaders, set dataset.root to TIDE-TIDES-data/dualsph_liquid_v2_dense for the main subset or TIDE-TIDES-data/dualsph_liquid_v3_dense_sizeood2x_testood for size OOD. Retain the raw/ folder when particle supervision is needed; portable manifest references resolve to that folder.

Files and Array Conventions

dualsph_liquid_v2_dense/
  manifest.jsonl
  summary.json
  samples/{train,val,test,test_ood}/*.npz
dualsph_liquid_v3_dense_sizeood2x_testood/
  manifest.jsonl
  summary.json
  samples/test_ood/*.npz
raw/<episode_id>/sim/
  water_surface_state.npz
  rigid_poses.csv
index/{sphere,sphere_size_ood}/
  *.parquet
schema.json
release_manifest.json
SHA256SUMS
LICENSE

Important dense NPZ members (T is the saved-frame count):

Array	Shape	Description
target_flow	[T,3,D,H,W]	Dense liquid velocity in xyz component order
init_flow	[3,D,H,W]	Initial velocity
target_velocity_voxel	[T,D,H,W,3]	Velocity in voxels per saved frame
target_velocity_world	[T,D,H,W,3]	Velocity in world distance per second
cond_flow	[T,3,D,H,W]	Solid velocity conditioning
cond_volume	[T,1,D,H,W]	Solid conditioning volume
particle_support_mask	[T,D,H,W]	Native particle-to-grid support
particle_support_velocity_world	[T,D,H,W,3]	Particle-supported velocity
fluid_mask	[T,D,H,W]	Benchmark-defined dense liquid mask
phase_onehot	[T,D,H,W,2]	Liquid/solid phase labels; no air channel
object_occupancy	[T,D,H,W]	Solid occupancy
object_center_world	[T,3]	Realized solid centers
object_velocity_world	[T,3]	Realized solid velocities
bounds_xyz, voxel_size_xyz, dt	Per-episode metadata	World bounds, spacing, frame interval

The velocity conversion is voxel_velocity = world_velocity * dt / voxel_size_xyz. Do not apply dt a second time when advecting in voxel/frame coordinates. schema.json contains representative complete array schemas for each configuration. Particle exports and rigid poses provide the preserved simulation-derived observations; solver executables and native binary dumps are not distributed here.

Intended Use and Limitations

The task is conditional liquid forecasting under a known future solid-state sequence, not closed-loop two-way fluid-solid simulation. These simulations supply realized solid motion; the forecaster does not update solid dynamics. Partition by entire episodes and preserve the supplied splits before extracting windows. Never provide future ground-truth liquid states as inference inputs.

Dense fields are constructed from SPH particle exports through particle-to-grid conversion and dense-domain filling. They are not solver-native continuum fields. Native particle support and filled liquid occupancy are different quantities; use particle_support_mask when reporting support-restricted metrics. This release contains only sphere geometry and does not establish generalization to cube/cone, real-world liquids, or new journal scenarios.

The released framework configurations are runnable starting points, not a guarantee of reproducing every paper/rebuttal schedule on this partial subset.

License and Attribution

The dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0). See LICENSE. Attribute Jongwon Ryu, Sunjae Yoon, and Junyeong Kim, cite the paper, and identify any modifications. Framework code is distributed separately in the linked GitHub repository.

@article{ryu2026interactive,
  title={Interactive 4D Volumetric Liquid Forecasting under Moving-Solid Interaction},
  author={Ryu, Jongwon and Yoon, Sunjae and Kim, Junyeong},
  journal={arXiv preprint arXiv:2610.16167},
  year={2026}
}