FluidZero: Mastering Diverse Tasks in Fluid Systems through a Single Generative Model

Haodong Feng, Haoren Zheng, Peiyan Hu, Hongyuan Liu, Chongwen Yu, Long Wei, Rui Feng, Jinlong Duan, Dixia Fan, Tailin Wu

Abstract

Fluid dynamics are pervasive in human civilization, such as spacecraft and submersible, which are important to understand and optimize, involving four key tasks for a long time: prediction, parameter identification, design, and control. Traditionally, these tasks have been addressed by separate models, each tailored to specific scenarios, leading to redundancy and repeated development costs. To overcome this issue while utilizing the correlation between tasks to improve the effectiveness of each task, we propose FluidZero, a unified deep generative model, to tackle all these tasks using one single model. FluidZero is inspired by the success of large foundation models in several domains, making it possible to develop a unified fluid model capable of handling all these tasks. The key advantage of FluidZero is that it facilitates cross-modal and cross-task interactions on diverse data, enabling effective physical knowledge learning and significantly enhancing performance on all tasks across a wide range of fluid scenarios. We evaluate FluidZero across multiple datasets, including simulation data and real-world measured data obtained through Particle Image Velocimetry (PIV). Notably, FluidZero shows remarkable generalization capabilities, achieving superior performance even in out-of-distribution (OOD) situations and real-world applications.

Introduction

Fluid-Structure Interaction (FSI) is a cornerstone of modern engineering, essential for designing everything from aerospace vehicles to medical implants. Historically, the application of artificial intelligence to FSI has been fragmented. Specialized models are typically developed for isolated tasks: predicting flow fields, identifying system parameters (e.g., Reynolds number), optimizing geometric designs, or implementing real-time control. This "one model, one task" paradigm is not only computationally redundant but also fails to exploit the inherent physical interdependencies between these tasks.

FluidZero addresses this fragmentation by introducing a unified generative framework capable of mastering the entire engineering lifecycle within a single architecture. By learning the joint probability distribution of diverse physical modalities, FluidZero acts as a "Swiss Army Knife" for fluid systems, demonstrating that vertical integration within a physical domain can yield superior performance and robustness compared to task-specific baselines.

The FluidZero Framework

The architecture of FluidZero is built upon a diffusion-based Transformer designed to handle heterogeneous data types and complex physical relationships.

FluidZero architecture overview, illustrating the unified approach to prediction, parameter identification, design, and control.

Core Components

  1. Diffusion Backbone: FluidZero utilizes a diffusion model to learn the joint probability distribution of flow fields, geometries, parameters, and control actions. Solutions are generated by iteratively denoising Gaussian noise, conditioned on known variables.
  2. Multimodal Transformer: To process diverse inputs, the model employs modality-specific encoders. Geometries are represented as Signed Distance Functions (SDF), while sequential data (like forces) and scalar parameters are processed via Fourier-based encoders to capture periodic patterns.
  3. Task-Masked Optimization: During training, the model is exposed to incomplete data with randomly masked modalities. This forces the network to learn the underlying causal laws—for example, how a specific foil geometry dictates vortex shedding patterns.
  4. Guided Sampling: For design and control, gradient-based guidance (via DDIM sampling) steers the generative process toward specific performance objectives, such as maximizing the lift-to-drag ratio.

Experimental Evaluation

FluidZero was rigorously tested across a hierarchy of data sources to evaluate its accuracy and generalization:

  • Simulation Data: Generated using the Lily-pad CFD solver, featuring single and double cylinders/foils across various Reynolds numbers ($Re$).
  • Zero-Shot Benchmarks: Testing on square cylinder configurations never seen during training to assess geometric generalization.
  • Real-World Data: Validated using Particle Image Velocimetry (PIV) data from circulating water tanks, including complex Wake-Induced Vibration (WIV) scenarios.

Key Results and Insights

FluidZero consistently outperformed state-of-the-art specialized models across all four core tasks.

Performance comparison showing FluidZero's superior accuracy in flow prediction and parameter identification compared to specialized baselines.

Performance Metrics

  • Superior Prediction: The model reduced relative $L_2$ error in flow field prediction by 21.8% to 30.5% compared to baselines like FNO and AROMA.
  • Robust Identification: FluidZero achieved up to an 88.1% improvement in estimating Reynolds numbers from flow observations compared to traditional backpropagation methods.
  • High-Performance Design: The model generated foil shapes with median lift-to-drag ratios of ~2.0, significantly exceeding the ~0.5 achieved by standard diffusion models.
  • OOD Generalization: In control tasks, FluidZero successfully generalized to Reynolds numbers up to 100,000, despite a training limit of $Re=10,000$.

Synergy through Unification

A key finding is that multi-task training acts as a natural regularizer. The "inverse" tasks (design/identification) help the model better understand "forward" physics (prediction). Furthermore, the model's ability to handle noisy real-world PIV data demonstrates its robustness to the "sim-to-real" gap.

Conclusion

FluidZero represents a significant shift toward unified engineering agents. By integrating diverse tasks into a single generative framework, it demonstrates that AI can move beyond solving isolated equations to understanding holistic physical systems. Future work will focus on scaling these capabilities to 3D high-Reynolds-number turbulent flows and improving the interpretability of the learned physical laws.

Summary updated

Abstract

<title>Abstract</title> Fluid dynamics are pervasive in human civilization, such as spacecraft and submersible, which are important to understand and optimize, involving four key tasks for a long time: prediction, parameter identification, design, and control. Traditionally, these tasks have been addressed by separate models, each tailored to specific scenarios, leading to redundancy and repeated development costs. To overcome this issue while utilizing the correlation between tasks to improve the effectiveness of each task, we propose FluidZero, a unified deep generative model, to tackle all these tasks using one single model. FluidZero is inspired by the success of large foundation models in several domains, making it possible to develop a unified fluid model capable of handling all these tasks. The key advantage of FluidZero is that it facilitates cross-modal and cross-task interactions on diverse data, enabling effective physical knowledge learning and significantly enhancing performance on all tasks across a wide range of fluid scenarios. We evaluate FluidZero across multiple datasets, including simulation data and real-world measured data obtained through Particle Image Velocimetry (PIV). Moreover, the designed foil is directly transferred to real-world experiments through 3D printing. Notably, FluidZero shows remarkable generalization capabilities, achieving superior performance even in out-of-distribution (OOD) situations and real-world applications. By integrating diverse fluid system tasks across varied scenarios into a unified model, FluidZero demonstrates the revolutionary impact of unified AI approaches in fluid dynamics, opening new avenues for the adoption of integrated fluid system understanding and optimization throughout scientific and engineering domains.
BibTeX
@inproceedings{feng2025fluidzero,
  author = {Haodong Feng and Haoren Zheng and Peiyan Hu and Hongyuan Liu and Chongwen Yu and Long Wei and Rui Feng and Jinlong Duan and Dixia Fan and Tailin Wu},
  title = {{FluidZero: Mastering Diverse Tasks in Fluid Systems through a Single Generative Model}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-6881567/v1},
}
Figure