NeuralFluid: Neural Fluidic System Design and Control with Differentiable Simulation

Yifei Li
MIT CSAIL
        
Yuchen Sun
Georgia Institute of Technology
        
Pingchuan Ma
MIT CSAIL
        
Eftychios Sifakis
UW-Madison
        
Tao Du
Tsinghua University
        
Bo Zhu
Georgia Institute of Technology
        
Wojciech Matusik
MIT CSAIL

Conference on Neural Information Processing Systems (NeurIPS) 2024

PDF Cite DOI arXiv

Abstract

We present NeuralFluid, a novel framework to explore neural control and design of complex fluidic systems with dynamic solid boundaries. Our system features a fast differentiable Navier-Stokes solver with solid-fluid interface handling, a low-dimensional differentiable parametric geometry representation, a control-shape co-design algorithm, and gym-like simulation environments to facilitate various fluidic control design applications. Additionally, we present a benchmark of design, control, and learning tasks on high-fidelity, high-resolution dynamic fluid environments that pose challenges for existing differentiable fluid simulators. These tasks include designing the control of artificial hearts, identifying robotic end-effector shapes, and controlling a fluid gate. By seamlessly incorporating our differentiable fluid simulator into a learning framework, we demonstrate successful design, control, and learning results that surpass gradient-free solutions in these benchmark tasks.

Pipeline

Pipeline Overview. (1) Our pipeline starts with an initial parametric geometry and a neural network parameterized controller. (2) The fluid dynamics is then simulated using a dynamic Navier-Stokes solver. (3) The performance of the design and control is evaluated using a loss function, the gradients of which are then back-propagated through our end-to-end differentiable framework. (4) The gradient-based optimization iteratively improves the geometry and control to achieve the task goal. This pipeline allows for efficient geometry and control co-optimization.

Demos

Case Study: Neural Control of An Artificial Heart

Goal: Train a controller to match the outlet velocity to predefined target flow profiles.

Setup:Heart geometry includes two inlets, one outlet, and a chamber.

Controller: A two-layer MLP using time step and outflow norm as inputs.

Target profiles: Cosine wave (smooth periodic flow). ECG-like waveform. Optimization: Minimized the difference between the actual and target flow velocities. Results: The controller generated flow matching the targets, validating the scalability and effectiveness of our method.

Ablation Studies

(a) Effects of Initialization on Optimization We tested our gradient-based approach for controlling a 3D heart model with networks initialized from five different random seeds. Despite initial variations, all runs converged within a few tens of iterations, demonstrating the method’s robustness to initialization and scalability to high-dimensional settings.

(b) Gradient-Based vs Gradient-Free Optimization We compared our gradient-based method against PPO and CMA-ES on the Neural Heart task. Our approach converged rapidly, while gradient-free methods struggled. Although gradient-based optimization requires a differentiable objective, it excels in speed and efficiency when such gradients are available. In more complex, non-differentiable scenarios, gradient-free methods remain viable alternatives.

Poster

Citation

@misc{li2024neuralfluidicdesigncontrol,
  title={Neural Fluidic System Design and Control with Differentiable Simulation}, 
  author={Yifei Li and Yuchen Sun and Pingchuan Ma and Eftychios Sifakis and Tao Du and Bo Zhu and Wojciech Matusik},
  year={2024},
  eprint={2405.14903},
  archivePrefix={arXiv},
  primaryClass={physics.flu-dyn},
  url={https://arxiv.org/abs/2405.14903},

}

Keywords

fluid simulation, fluid control, navier-stokes, differentiable simulation, physical simulation, neural fluid, neural control, graphics

fluid simulation fluid control navier-stokes differentiable simulation physical simulation neural fluid neural control graphics

Related