DeepLag: Discovering Deep Lagrangian Dynamics for Intuitive Fluid Prediction

Accurately predicting the future fluid is vital to extensive areas such as meteorology, oceanology, and aerodynamics. However, since the fluid is usually observed from the Eulerian perspective, its moving and intricate dynamics are seriously obscured and confounded in static grids, bringing thorny challenges to the prediction. This paper introduces a new Lagrangian-Eulerian combined paradigm to tackle the tanglesome fluid dynamics. Instead of solely predicting the future based on Eulerian observations, we propose DeepLag to discover hidden Lagrangian dynamics within the fluid by tracking the movements of adaptively sampled key particles. Further, DeepLag presents a new paradigm for fluid prediction, where the Lagrangian movement of the tracked particles is inferred from Eulerian observations, and their accumulated Lagrangian dynamics information is incorporated into global Eulerian evolving features to guide future prediction respectively. Tracking key particles not only provides a transparent and interpretable clue for fluid dynamics but also makes our model free from modeling complex correlations among massive grids for better efficiency. Experimentally, DeepLag excels in three challenging fluid prediction tasks covering 2D and 3D, simulated and real-world fluids. Code is available at this repository: https://github.com/thuml/DeepLag.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Kve37/10 · confidence 4/52024-07-07

Summary

Real-world processes exhibit multi-scale spatio-temporal dynamics. Not all of this dynamics is accurately modeled by Eulerian (i.e., field-based) modeling of scientific processes and sometimes the fine-grained patterns are only able to be modeled by Lagrangian paradigms. However explicit Lagrangian-only modeling is costly and hence this paper proposes a deep learning based surrogate model that is able to jointly model the Eulerian and Lagrangian perspectives by proposing a novel `EuLag' deep learning block. The authors demonstrate the performance of the proposed architecture on multiple fluid-dynamics related tasks involving multi-phase flows.

Strengths

- The authors develop a well-motivated and novel deep learning model architecture to combine the Eulerian and Lagrangian computational modeling paradigms thereby leading to improved modeling capacity owing to modeling across a larger variety of scales. - The proposed model demonstrates performance improvements compared to state-of-the-art models like the Fourier-Neural Operator and transformer based architectures (also a central feature in the proposed EuLag block) like the Galerkin Transformer. Overall, the experimental results are convincing.

Weaknesses

- More information regarding the experimental setup needs to be included in the main text of the paper. Currently, the main body of the paper (especially related to the experimental setup and the dataset descriptions, model description) cannot stand on its own without the appendix. At least the full description of the model architecture (i.e., upsampling, down sampling and other critical operations) should appear in the main text. - Comparison with a Lagrange-only model is necessary in addition to the current field-based prediction models for a more holistic experimental comparison.

Questions

1. Why have the authors not compared to a Lagrange-Only model like that of [1] for fluid flow modeling. Since the proposed paradigm is an Euler-Lagrange paradigm, baseline with models that only in the Eulerian domain seems somewhat incomplete and for completeness, comparison with one particle based deep-learning model (like [1]) should be included. ## References [1] Sanchez-Gonzalez A, Godwin J, Pfaff T, Ying R, Leskovec J, Battaglia P. Learning to simulate complex physics with graph networks. InInternational conference on machine learning 2020 Nov 21 (pp. 8459-8468). PMLR.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Authors have addressed the potential limitations of their work sufficiently well.

Reviewer PRfM6/10 · confidence 3/52024-07-12

Summary

The authors propose DeepLag as an approach to simulating Eulerian fluid dynamics, which makes use of Eulerian-Lagrangian co-design to improve performance. In particular, the idea is to transfer information back and forth between the Eulerian grid and initially randomly placed Lagrangian particles, which themselves exist on multiple but connected discretization scales. The intuition is that Lagrangian particles are better at capturing interactions with boundaries and, in general, tracking the location of material elements, e.g., smoke. Empirically, the gain in performance over existing models is consistent and around 10%. The limitations of the approach are its algorithmic complexity and hyperparameters.

Strengths

- The idea of having "attention" between the Eulerian grid and Lagrangian particles is interesting and seems to work well. - The multi-scale approach makes a lot of sense to map fine details to the coarser resolution. Not sure whether this is only a strength though, as this means higher algorithmic complexity of the model. - The approach by construction offers Lagrangian particle trajectories, which might be useful in e.g. ocean dynamics research

Weaknesses

- **Factual mistakes**: the authors make some very crude false statements here and there, and I'm questioning whether a person or LLM wrote sections of the manuscript, particularly the introduction. - Lines 24-25: "curse of dimensionality" is NOT the reason for high computational cost in fluid dynamics! Please open the Wikipedia article on the curse of dimensionality and modify the sentence in the paper. - Lines 40-42 state that "[Lagrangian approach] helps get around the CFL condition", which is absolutely wrong. - **Related work taxonomy**: in section 2.2, a categorization of neural solvers is presented; however, this grouping of methods is new to me and, as much as I know, not what the community typically uses. - The 1st group, "ODE-based Generative Models", refers to some rather old and outdated work, - the 2nd group, "Neural PDE Solvers", actually only talks about Physic-Informed Neural Networks, and - the 3rd group, "Neural Operators for PDE", basically lists all methods that I would call modern neural PDE solvers and what is typically referred to as neural operator learning (DeepONet, FNO, etc.) is mentioned along with GNNs and CNNs. See PDE Bench [1] and PDE Arena [2] for a more modern categorization of common PDE learning approaches. - **Ablations**: I'm missing an ablation on (a) the number of scales and (b) overall model size. Judging by Fig 7, in which the U-Net is 5x faster and also smaller, I wonder how a U-Net with similar (a) parameter count or (b) runtime would perform. I'm a bit worried that the claimed good performance might be just a bad hyperparameter choice of the baselines. --- [1] Takamoto et al., "PDEBench: An Extensive Benchmark for Scientific Machine Learning", NeurIPS D&B 2022 [2] Gupta and Brandstetter, "Towards Multi-spatiotemporal-scale Generalized PDE Modeling", 2022

Questions

- Fig. 7: in this figure, it looks like the U-Net has fewer parameters than DeepLag, but in lines 345-346, as well as Table 10, you state the opposite. Can you explain that?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

- Thinking of Occam's Razor, it is questionable how many people would adopt the proposed approach, as it requires (a) a significantly more involved implementation effort than a typical U-Net, (b) tuning various hyperparameters like number of scales and number of particles per scale, and (c) is ~5x slower (Fig 7), while offering on average 10% performance improvement. Could the authors add some more hints on where their approach would be of practical interest? I would appreciate extending section 5 into, for example, one paragraph dedicated to the summary/strengths/applications, and one with the limitation/future work.

Reviewer PBwE6/10 · confidence 4/52024-07-15

Summary

In this paper, the author presents a Lagrangian-Eulerian hybrid paradigm to address the complexities of fluid dynamics. Instead of relying only on Eulerian observations to predict future states, we introduce DeepLag, which uncovers hidden Lagrangian dynamics within the fluid by tracking the movements of adaptively sampled key particles. In experiments, DeepLag show better performance in three demanding fluid prediction tasks both in 2D and 3D, as well as simulated and real-world fluids.

Strengths

- The idea of integrating Lagrangian tracking into the deep model for assisting Eulerian based fluid prediction sounds novel. - The LagToEu Attention and EuToLag Attention for exchanging information between Eulerian and Lagrangian view is intuitive. - The proposed method shows better performance on all three test cases comparing to all baselines.

Weaknesses

- In the Bounded Navier Stokes, the performance of 10 Frames is significant better than 30 Frames, while for the Ocean Current case, the results of long-term rollout shows gives better result. Just wondering what is the reason between this discrepancy? Is the proposed method better than in shot-term or long-term tasks? - How does the proposed method handle the complex boundaries? e.g., how to impose different boundary conditions into the framework?

Questions

See weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes. The author discuss the limitations in the paper.

Reviewer rCcn5/10 · confidence 3/52024-07-16

Summary

The authors propose a novel neural network architecture in order to leverage the advantages of the eulerian and lagrangian formalisms for fluid prediction. The so called "EuLag Block" acts on an eulerian grid based representation of the fluid as well as on a lagrangian particle based representation and enables information flow between both representations through 2 attention blocks: 1. The "LagToEu Attention" block allows the network to pass information from the lagrangian particle based representation to the eulerian grid based representation 2. The "EuToLag Attention" block allows the network to pass information from the eulerian grid based representation to the lagrangian particle based representation On top of these 2 attention blocks, a bilinear interpolation module is used to extract information about local dynamics from the eulerian representation, which is then combined with information about global dynamics from all particles to compute the update of the particle based representation. The distribution of the initial particle positions is generated by a learnable sampling module. The authors evaluate the resulting DeepLag architecture on 3 different fluid prediction tasks and outperform several state of the art methods. Furthermore, ablation studies were performed in order to investigate the effect of different numbers of particles, both attention blocks as well as the learnable sampling module.

Strengths

- the qualitative results look convincing (especially the long term stability shown in the videos) - fairly extensive quantitative comparison to other network architectures and ablation studies - interesting idea to use LagToEu / EuToLag attention blocks in order to reduce the squared complexity of the attention mechanism wrt the domain size in the eulerian frame of reference to a linear complexity wrt the domain size and the number of lagrangian particles (especially if the number of particles is chosen to be relatively small).

Weaknesses

- Reference to "Accelerating Eulerian Fluid Simulation With Convolutional Networks" by Tompson et al is missing. They use a particle tracer to deal with the advection term in the lagrangian frame of reference and train a CNN to perform a "pressure projection step" in the eulerian frame. They showed impressive smoke simulations in 3D similar to the experiments shown in section 4.3. Thus, I'm not sure if I can fully agree with the claim in line 55-57 that this is the first deep fluid prediction model that explicitly combines Eulerian and Lagrangian frameworks. - Regarding the efficiency analysis (Figure 7) it would be interesting to see how a U-Net would compare to DeepLag when upscaled up to a similar running time.

Questions

- How would the particle movements (shown in Figure 5 on the right) look for the Bounded Navier-Stokes dataset (Figure 4)? Would the particles still follow the velocity field although the velocity field is not considered as a field variable in this experiment but only the dye concetration? - Could the particle movements be used to extract information about the velocity field in the Bounded Navier-Stokes dataset? - Often in fluid dynamics (e.g. "Stable fluids" by Jos Stam or "Accelerating Eulerian Fluid Simulation With Convolutional Networks" by Tompson et al), solving the pressure field is done in the eulerian frame and requires a global solution. However, dealing with the advection term using a particle tracer requires only the local velocity field. Isn't the global attention for the particle updates a "slightly wasteful" overkill? - How does the positional embedding look like? - How well does your method generalize to new domains?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

- How would your method scale to larger domains? I'm wondering if the fully-connected layer in the dynamic sampling module could become a bottleneck for larger domain as it seems to scale quadratically with the domain size.

Reviewer UohQ6/10 · confidence 3/52024-07-18

Summary

The paper introduces a novel approach to predicting fluid dynamics by integrating both Lagrangian and Eulerian paradigms. The model, named ‘DeepLag,’ utilizes transformer blocks to process and integrate information from Eulerian and Lagrangian perspectives. Initially, DeepLag predicts the future state of the Eulerian space. Subsequently, it uses this prediction to infer the Lagrangian movements of key tracked particles. This methodology allows the Lagrangian dynamics to inform and guide the evolution of the Eulerian predictions, enhancing the overall prediction accuracy. DeepLag is evaluated against various baselines across three diverse datasets, including simulated bounded Navier-Stokes equations, real-world ocean currents, and 3D smoke dynamics. These experiments demonstrate the model’s effectiveness in handling both 2D and 3D fluid dynamics in simulated and real-world scenarios.

Strengths

- The DeepLag model introduces a novel framework that effectively combines Eulerian and Lagrangian perspectives, allowing for dynamic propagation in both spaces. This dual approach is innovative in fluid dynamics modeling, enhancing the prediction accuracy by leveraging the strengths of both paradigms. - The model’s effectiveness is rigorously tested across three challenging datasets—bounded Navier-Stokes equations, real-world ocean currents, and 3D smoke dynamics. The experiments demonstrate superior performance not only in standard scenarios but also in both short-term and long-term prediction tasks, highlighting the model’s versatility and robustness. - One of the standout features of DeepLag is its ability to provide interpretable results by showcasing individual particle trajectories via Lagrangian dynamics. This aspect is particularly valuable as it not only enhances the understanding of fluid movements but also aids in validating the model’s predictions through visual and traceable particle paths.

Weaknesses

- The model’s reliance on extensive hyperparameter tuning for particle tracking could pose challenges in terms of replicability and efficiency. This complexity might limit the accessibility of the model for practical applications without substantial computational resources. - Additionally, the particle tracking mechanism struggles to maintain focus on particles near the domain borders. This limitation is particularly concerning in scenarios where significant dynamic changes occur near these borders, such as the presence of obstacles. The inability to track these dynamics could potentially lead to incomplete or inaccurate modeling of fluid behavior in such areas. - The paper contains a few grammatical errors that, while minor, could detract from its overall professional presentation.

Questions

- How to interpret the difference between the left and right parts of Fig. 1? - Is there supervision over the Lagrangian space, or does it only happen in the Euler space? - Would interchanging the positions of the LagToEu and EuToLag blocks affect the model’s performance or learning dynamics? If so, how? - Given that Table 5 shows only a minimal performance boost from the EuToLag attention, how should this be interpreted in the context of the model’s overall efficiency and effectiveness? - In the Bounded Navier-Stokes experiments, are the positions of the cylinders consistent across all trials within the dataset, or do they vary?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations have been addressed

Reviewer PBwE2024-08-11

Response to rebuttal

Thanks for the authors' reply to my concerns. I don't have further questions and I remain positive about this paper.

Reviewer PRfM2024-08-12

I thank the authors for the elaboration, new ablations, and improved content. My score accordingly increases by +1.

Authorsrebuttal2024-08-13

Thanks for Your Response and Raising the Score

Thank you for your valuable suggestions and for acknowledging our rebuttal. Your feedback has been instrumental in refining both the writing and experiments in our DeepLag paper. We greatly appreciate your recognition and support.

Reviewer UohQ2024-08-13

Thank you to the authors for providing additional experiments and clarifications in response to my comments. I appreciate these efforts and will maintain my score.

Reviewer Kve32024-08-13

Response to Author Rebuttal by Reviewer Kve3

Thank you to the authors for their responses. After consideration of the responses, I continue to remain positive about the paper and will maintain my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC