Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Generation of Multimodal Fluid Simulation Data

Daniele Baieri, Donato Crisostomi, Stefano Esposito, Filippo Maggioli, Emanuele Rodolà

TL;DR

An efficient generation procedure to produce synthetic multi-modal datasets of fluid simulations that can reproduce the dynamics of fluid flows and allows for exploring and learning various properties of their complex behavior, from distinct perspectives and modalities is introduced.

Abstract

In this work, we introduce an efficient generation procedure to produce synthetic multi-modal datasets of fluid simulations. The procedure can reproduce the dynamics of fluid flows and allows for exploring and learning various properties of their complex behavior, from distinct perspectives and modalities. We employ our framework to generate a set of thoughtfully designed training datasets, which attempt to span specific fluid simulation scenarios in a meaningful way. The properties of our contributions are demonstrated by evaluating recently published algorithms for the neural fluid simulation and fluid inverse rendering tasks using our benchmark datasets. Our contribution aims to fulfill the community's need for standardized training data, fostering more reproducibile and robust research.

Efficient Generation of Multimodal Fluid Simulation Data

TL;DR

An efficient generation procedure to produce synthetic multi-modal datasets of fluid simulations that can reproduce the dynamics of fluid flows and allows for exploring and learning various properties of their complex behavior, from distinct perspectives and modalities is introduced.

Abstract

In this work, we introduce an efficient generation procedure to produce synthetic multi-modal datasets of fluid simulations. The procedure can reproduce the dynamics of fluid flows and allows for exploring and learning various properties of their complex behavior, from distinct perspectives and modalities. We employ our framework to generate a set of thoughtfully designed training datasets, which attempt to span specific fluid simulation scenarios in a meaningful way. The properties of our contributions are demonstrated by evaluating recently published algorithms for the neural fluid simulation and fluid inverse rendering tasks using our benchmark datasets. Our contribution aims to fulfill the community's need for standardized training data, fostering more reproducibile and robust research.
Paper Structure (16 sections, 11 equations, 5 figures, 4 tables)

This paper contains 16 sections, 11 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Motivation and applications
  3. Related work
  4. Dataset
  5. Data generation
  6. Datasets
  7. Coherent subsets of simulation space
  8. Hand-crafted multi-view scenes
  9. Copyright information
  10. Evaluation
  11. Examples of applications
  12. Data-driven Fluid Simulation
  13. Inverse Rendering Fluid Simulations
  14. Generation performance
  15. Conclusions
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Our generation tool performs Lattice Boltzmann simulation steps, exporting the fluid geometry and rendering output frames from multiple viewpoints. Camera positions are sampled with Fibonacci sphere sampling keinert:2015:fibonaccisphere, and exported as metadata.
  • Figure 2: Multiple rendered views for three timesteps of a smoke plume simulation from the ScalarFlow scalarflow dataset. The rendered density field is reconstructed from real smoke captures.
  • Figure 3: Samples extracted from a 50 frames rollout of a DLF Ummenhofer2020Lagrangian model trained on our obstacles dataset, with unseen initial conditions, compared to the ground truth LBM simulation.
  • Figure 4: Top: a subset of the views in our duck scene training data. Bottom: PAC-NeRF li2023pacnerf reconstruction and rendering for the same (unseen) views (the background is not learned by the model, so it was manually composited for this visualization).
  • Figure 5: PAC-NeRF evaluation on duck. Both rendering quality and IoU over occupancy grids degrade as the simulation progresses because the neural radiance field reconstruction gets increasingly constrained by the physics-based losses. Note that the image metrics are computed only on the masked foreground object.