---
license: other
#User-Defined Tags
tags:
  - Lagrangian
  - CFD
  - graph neural network
language:
  - en
  - zh
---
<p align="center">
  <strong>
    <span style="font-size: 30px;"> Lagrangian </span>
  </strong>
</p>

## Dataset Overview

The Lagrangian dataset is sourced from the DeepMind team's ICML 2020 paper [Learning to Simulate Complex Physics with Graph Networks](https://proceedings.mlr.press/v119/sanchez-gonzalez20a.html). It consists of the two-dimensional Water particle-dynamics data from the paper's Graph Network-based Simulator (GNS) benchmark. The data represents particles as graph nodes and describes fluid evolution over time through particle-position sequences and particle types. It can be used for Lagrangian particle-dynamics modeling, long-horizon fluid rollout prediction, and evaluation of graph-neural-network physics simulations.


## Supported Tasks

This standardized dataset repository organizes the training, validation, and test TFRecord files for Lagrangian Water, together with data metadata, a data schema, an integrity summary, and validation scripts. It can be used as data input for training, inference, evaluation, and visualization with the `OneScience/LagrangianMGN` model. The core data is placed uniformly under `data/Water/`; the training set contains 1,000 trajectories, while the validation and test sets each contain 30 trajectories.


## Dataset Format and Structure

The dataset uses the TFRecord SequenceExample format, with each record corresponding to one particle-motion trajectory. The number of particles, `num_particles`, varies by trajectory, and the spatial dimension is `dim=2`.

| Feature | Component | shape | dtype | Source Encoding | Description |
|---|---|---:|---|---|---|
| `position` | sequence feature | `[1001, num_particles, 2]` | `float32` | `bytes` | Two-dimensional particle positions for 1,001 frames, corresponding to 1,000 evolution time steps |
| `particle_type` | context feature | `[num_particles]` | `int64` | `bytes` | Type identifier for each particle |

The data splits are as follows:

| File | split | Number of Trajectories | Description |
|---|---|---:|---|
| `data/Water/train.tfrecord` | `train` | 1,000 | Used for model training |
| `data/Water/valid.tfrecord` | `valid` | 30 | Used for model validation and hyperparameter selection |
| `data/Water/test.tfrecord` | `test` | 30 | Used for model testing and result evaluation |

`data/Water/metadata.json` also contains `bounds=[[0.1, 0.9], [0.1, 0.9]]`, `sequence_length=1000`, `default_connectivity_radius=0.015`, `dt=0.0025`, and normalization fields such as `vel_mean`, `vel_std`, `acc_mean`, and `acc_std`.

## How to Use the Dataset

This dataset is compatible with the `OneScience/LagrangianMGN` model.

- Files and Download:

```bash
hf download --dataset OneScience-Sugon/lagrangian --local-dir ./data
```


## Official OneScience Information

| Platform | OneScience Main Repository | Skills Repository |
|---|---|---|
| Gitee  | https://gitee.com/onescience-ai/onescience | https://gitee.com/onescience-ai/oneskills |
| GitHub  | https://github.com/onescience-ai/OneScience | https://github.com/onescience-ai/oneskills |

## Citation and License

- Original Lagrangian Water paper: [Learning to Simulate Complex Physics with Graph Networks](https://proceedings.mlr.press/v119/sanchez-gonzalez20a/sanchez-gonzalez20a.pdf)
- This dataset is organized from the Learning to Simulate project released by Google DeepMind. Before public distribution, confirm the licensing requirements of the upstream project.