Neural Latent Arbitrary Lagrangian-Eulerian Grids for Fluid-Solid Interaction
Shilong Tao, Zhe Feng, Shaohan Chen, Weichen Zhang, Zhanxing Zhu, Yunhuai Liu
TL;DR
Fisale is a novel data-driven framework that models complex two-way fluid-solid interactions using multiscale latent ALE grids and a partitioned coupling approach, enabling scalable and geometry-aware simulations.
Contribution
It introduces a flexible, geometry-aware neural framework inspired by classical ALE methods for modeling complex two-way FSI problems with improved scalability and accuracy.
Findings
Outperforms existing models in challenging 2D and 3D FSI scenarios.
Effectively captures nonlinear, heterogeneous fluid-solid interactions.
Demonstrates scalability across multiple complex tasks.
Abstract
Fluid-solid interaction (FSI) problems are fundamental in many scientific and engineering applications, yet effectively capturing the highly nonlinear two-way interactions remains a significant challenge. Most existing deep learning methods are limited to simplified one-way FSI scenarios, often assuming rigid and static solid to reduce complexity. Even in two-way setups, prevailing approaches struggle to capture dynamic, heterogeneous interactions due to the lack of cross-domain awareness. In this paper, we introduce \textbf{Fisale}, a data-driven framework for handling complex two-way \textbf{FSI} problems. It is inspired by classical numerical methods, namely the Arbitrary Lagrangian-Eulerian (\textbf{ALE}) method and the partitioned coupling algorithm. Fisale explicitly models the coupling interface as a distinct component and leverages multiscale latent ALE grids to provide unified,…
Peer Reviews
Decision·ICLR 2026 Poster
1. The paper provides good motivation to model ALE systems for fluid-solid interactions. 2. The authors conduct thorough experiments, with a very detailed appendix section. 3. The authors have considered SOTA baselines and SOTA problem settings.
1. The writing is quite dense in section 3 and it is not clear to me exactly how this proposed architecture compares to "A Neural Material Point Method for Particle-based Emulation, O Sharabi" 2. It seems like grid update is similar to message passing. It seems like section 3.3 is describing typical neural network operations and as such can be moved to the appendix for better readability. 3. The overall writing can be significantly improved, with many parts being unclear with substantial focus
1. Multiscale modeling has proved to be an effective way in solving PDEs which involve dynamic boundaries and multiple domains, such as in fluid-structure interaction. Using more samples in interface regions for accuracy while using less samples is domain interiors for efficiency is a well-studied approach. The authors seem to leverage this well in the construction of their architecture. 2. In the experiments provided in the paper, Fisale seems to outperform other baselines, which is a good ind
1. Although the authors claim to leverage the classical numerical formulations like ALE and Partitioned Coupling, it does not seem to be the defining factor here. This is clear from the ablations performed in Appendix G. The increased accuracy in the experiments in the main paper can simply arise from the high representational capacity of the architecture itself, since there are a lot of learnable components. Also, the choice of the dimension of the latent ALE grid (i.e. D) does not seem to be e
- The paper is well and clearly written, the introduction is comprehensive, and the problem is demonstrated in an intuitive way with Figure 1. - Architecture design is well-motivated: - It suggests a novel approach to handling a gap in prior works, and the problem tackled by the method is practically relevant for many engineering fields. - The approach is inspired by established numerical techniques (ALE and partitioned coupling). - Each component has a physical justification and is ground
- The paper adopts linear attention which is empirically weaker than standard attention despite the scale of the problems being manageable for flash attention. - The method does not scale well with increasing the number of latent points, which, in my opinion, is a limitation, see Questions. - Ablation studies indicate that some of the components might not even be necessary: - For example, Table 15 demonstrates that the ordering within PCM doesn't change the performance. Since that is the case,
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsModel Reduction and Neural Networks · 3D Shape Modeling and Analysis · Lattice Boltzmann Simulation Studies