Hybrid Neural-MPM for Interactive Fluid Simulations in Real-Time

TL;DR

This work tackles real-time interactive fluid simulation by introducing a hybrid neural-physics framework that blends a fast, graph-based neural simulator with a fallback to the Material Point Method (MPM) for fidelity. It further equips the system with a diffusion-based controller trained via reversed simulation to produce external force fields from simple user sketches, enabling intuitive manipulation of fluids. Across diverse 2D/3D scenarios and materials, the approach achieves real-time performance with substantial latency reductions while maintaining physically plausible outcomes and enabling user-guided control. The combination of low-spatiotemporal-resolution learning, robust safeguard mechanisms, and controllable generation promises practical applicability in graphics, design, and immersive environments.

Abstract

We propose a neural physics system for real-time, interactive fluid simulations. Traditional physics-based methods, while accurate, are computationally intensive and suffer from latency issues. Recent machine-learning methods reduce computational costs while preserving fidelity; yet most still fail to satisfy the latency constraints for real-time use and lack support for interactive applications. To bridge this gap, we introduce a novel hybrid method that integrates numerical simulation, neural physics, and generative control. Our neural physics jointly pursues low-latency simulation and high physical fidelity by employing a fallback safeguard to classical numerical solvers. Furthermore, we develop a diffusion-based controller that is trained using a reverse modeling strategy to generate external dynamic force fields for fluid manipulation. Our system demonstrates robust performance across diverse 2D/3D scenarios, material types, and obstacle interactions, achieving real-time simulations at high frame rates (11~29% latency) while enabling fluid control guided by user-friendly freehand sketches. We present a significant step towards practical, controllable, and physically plausible fluid simulations for real-time interactive applications. We promise to release both models and data upon acceptance.

Figures (15)

• Figure 1: We target real-time, interactive fluid simulations. Our hybrid solver integrates a numerical simulator and neural physics (Section \ref{['sec:realtime_hybrid']}), enabling real-time simulation (Section \ref{['sec:exp_acceleration']}). In addition, we generate external force fields (Section \ref{['sec:interactive_control']}) to support users to control fluids interactively via freehand sketches (Section \ref{['sec:exp_control']}).
• Figure 2: GNN as our neural physics simulator.
• Figure 3: Method Overview. To achieve real‑time simulations, we cut latency by learning neural physics at a coarse spatiotemporal resolution, while safeguarding fidelity by automatically falling back to an MPM solver when complex fluid phenomena arise (Section \ref{['sec:realtime_hybrid']}). For interactive control, we train a diffusion‑based generative model that infers external force fields directly from user sketches (Section \ref{['sec:interactive_control']}).
• Figure 4: Our neural physics accelerates simulations by learning and inferring at low spatial ($N_l$ num. particles) and temporal ($\Delta t$ time steps) resolutions, with downsampling ratios as $r_p$, $r_t$.
• Figure 5: Negative correlation between "cosine similarity of particle accelerations over frames" vs. "simulation errors of neural physics". Scenario: Water 2D. Spearman correlation: -0.3902.