TL;DR
This paper introduces a hybrid neural-physics system that enables real-time, interactive fluid simulations with high fidelity and low latency, suitable for practical applications and user control.
Contribution
It presents a novel hybrid approach combining numerical simulation, neural physics, and generative control for real-time fluid simulation and manipulation.
Findings
Achieves real-time fluid simulation at 11-29% latency.
Supports diverse 2D/3D scenarios and obstacle interactions.
Enables fluid control guided by freehand sketches.
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…
Peer Reviews
Decision·Submitted to ICLR 2026
The blend of classical and neural approaches effectively balances their respective strengths and limitations, making this a pragmatic and promising path forward.
While the hybrid direction is practical, I find the core mechanism of monitoring the error then hard-switching to a classical solver is a generic wrapper with limited novelty. As presented, it could be applied to most neural simulators; the paper should demonstrate if any part of their design makes fallback uniquely efficient, and compare against to other neural simulation baselines by applying the same fallback onto other neural simulators. The generative controller is also weakly motivated:
- The proposed fallback mechanism is am elegant safeguard against simulation drift, enabling stability against potential issues encountered in long-horizon rollouts. - Integrating a diffusion model to learn reverse accelerations for artistic control is a creative idea, offering a novel way to steer physically based simulations through generative methods.
- The presentation in its current format is sub-par. Several plots rely on scattered points that difficult clear understanding, while Figure 7 is particularly confusing. Why does the red rollout abruptly stop? It's also suspicious that the hybrid error high before the cutoff, so it seems like the approximation has a high baseline error. Moreover the abrupt truncation of the red plot could be an attempt of potentially masking issues of the method on further roll-outs. - The related-work discuss
1. The research addresses a highly interesting and valuable problem, aiming to balance speed, accuracy, and interactivity, presenting a prototype of a potentially practical system. 2. The core idea is novel. 3. The method for generating training data for the controller is simple yet effective. 4. The paper is clearly written.
1. In Figure 7, why does the grid MSE finally decrease when transitioning from the neural physics phase to the MPM phase? Intuitively, one might expect it to increase monotonically. Could the authors provide insight into this phenomenon? 2. **Sensitivity of the Fluid Complexity Threshold (`r_c`)**: How sensitive is the hyperparameter `r_c`, used to trigger the MPM solver? If the simulation scenario changes, would the value of `r_c`require adjustment? This dependency may limit the practical utili
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
