Learning Efficient Surrogate Dynamic Models with Graph Spline Networks

While complex simulations of physical systems have been widely used in engineering and scientific computing, lowering their often prohibitive computational requirements has only recently been tackled by deep learning approaches. In this paper, we present GraphSplineNets, a novel deep-learning method to speed up the forecasting of physical systems by reducing the grid size and number of iteration steps of deep surrogate models. Our method uses two differentiable orthogonal spline collocation methods to efficiently predict response at any location in time and space. Additionally, we introduce an adaptive collocation strategy in space to prioritize sampling from the most important regions. GraphSplineNets improve the accuracy-speedup tradeoff in forecasting various dynamical systems with increasing complexity, including the heat equation, damped wave propagation, Navier-Stokes equations, and real-world ocean currents in both regular and irregular domains.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Rs2G7/10 · confidence 3/52023-07-05

Summary

The paper builds upon a recent advances in deep-learning based collocation methods for physical simulation to improve prediction accuracy in continuous space and time. The authors proposed a hybrid model based on graph neural network and differentiable orthogonal spline collocation method. Interpolation in time as well as in space is adopted to enable employing coarser temporal grids to speed up prediction. The paper also introduces a new adaptive collocation sampling strategy and techniques for training such as not needing to learn interpolators or collocation weights to decrease the computational overhead. Experiments in various PDE settings demonstrate that the proposed model, GRAPHSPLINENETS outperforms state-of-the-art baselines in both accuracy and runtime.

Strengths

The paper introduces some techniques, a novel loss function comprising OSC function, COLROW algorithm to efficiently solve linear equations associated with OSC methods, and adaptive collocation sampling to allow for prioritized sampling of important location regions. Wide variety of experiments are conducted, and the proposed method outperforms the baselines.

Weaknesses

Although the authors claim that GraphSplineNets forecast continuous responses in time and space, comparison of GraphSplineNets are limited to space continuous baselines. Also, as in the main text, only a fewer points are sampled as collocation points in time — It is still difficult to assess the advantage of incorporating time-oriented OSC as well as space-oriented into the architecture. It is also unclear how is dataset for interpolation points error in the equation (2) generated. The explanation on Figure 5.1 is too brief for me to understand. Does the size of input of MGN is larger than that of MGN + OSC (12x12)? In appearance it looks strange to me that the running time of MGN w/o OSC higher than MGN w/ OSC.

Questions

* Is OSC inference time in Figure 5.1 associated with both spacial and time oriented collocation? * Does adaptive collocation sampling also perform for important time interval?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, the limitations are discussed in the main text.

Reviewer jrMD7/10 · confidence 3/52023-07-06

Summary

The authors introduce GraphSplineNets, providing a framework for combining MPNNs like MGNs with interpolation/collocation methods. Their contributions also include two methods to surpass the performance of a naive combination: end-to-end training and collocation points adaptation. Temporal and spatial downsampling allow the MPNN to run faster, while the collocation method makes possible *increased* inference-time resolution despite the NN's downsampled domain.

Strengths

Significance: 1. This is an important area, and the paper's core idea is interesting and well presented (as a proof of concept, at least). The paper offers a lot of infrastructure (code and ideas) important to future steps at the intersection of GNN surrogates and collocation. 2. The inference speed benefit of GSNs is intriguing because applying GNNs at scale can be difficult. Clarity: 3. The motivation for this work is communicated well. This work's connection to other approaches that deal with GNN scalability is clearly established. 4. The explanation of the GNN-based OSC method is very well done and nicely illustrated. Quality: 5. OSC seems like a potentially promising approach for dealing with the efficiency challenges of GNN simulations. 6. The end-to-end optimization approach (differentiating through the OSC) is well justified and intuitive. Originality: 7. Plugging GNNs into an OSC method, allowing temporal coarsening of MPNN (e.g. MGN) activity to increase efficiency, and optimizations of this joint GNN+OSC method (end-to-end training, adaptive points). 8. Giving automatic differentiation support to an ABD matrix equation solving routine.

Weaknesses

My main concern is the quality of the evaluation of the method. I think this paper would gain significance if it showed that GSN's performance benefits hold in more challenging settings (by demonstrating why the benefits of GSN will generally be present, by evaluating on harder problems and against better tuned baselines, etc.). In other words, addressing some of the following weaknesses might help other researchers better understand the potential relevance of this method. 1. The paper leaves unclear why its methods beat baseline methods' errors (i.e., is it because downsampled mesh nodes have access to information from farther away?), if the errors it achieves are practically relevant (i.e., sufficiently low for the model to be used), and if it beats stronger baselines (i.e., MGN with more message passing layers or MultiScale MGN). 2. Moreover, no large-scale or 3D problems were considered, and this is at odds with the description of the proposed method as one that addresses efficiency challenges of existing approaches. Minor concerns: 3. Figure 4.1 is nice, but I think at $t=0$, you should have a $y_3$ instead of a $y_0$.

Questions

1. Why use only 3 MP layers in baseline MGNs? This gives the GraphSplineNets method an unfair advantage in that it can communicate over a larger spatial extent (because it is operating on a coarser mesh). Relatedly, the baseline MGN results (e.g., those in table 5.1) are rather poor compared to results shown in the MGN paper, which recommends using 15 MP layers. 2. If you keep on increasing the collocation points, will GSNs eventually provide the same solution as MGNs? Assuming this is the case, Figure 5.3 suggests that more collocation points is better, but it might be nice to see when the relationship flips and the methods converge. 3. How do GSNs perform on larger scale problems? In particular, can they compete with approaches designed to be more scalable like MultiScale MeshGraphNets (Fortunato et al., 2022)? 4. Line 27: It's true that traditional solvers can struggle, but it might make sense to support this point by citing a paper that analyzes traditional solvers rather than a paper that analyzes ML solvers. 5. In the Appendix, there seem to be some minor typos/errors in the equations. E.g., equations 7 and 11 seem to assume 2 partitions when the example has 3 partitions. This did not affect my score, but cleaning it up would be great!

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors addressed some key limitations. I might add that no large-scale or 3D problems were considered, which hinders understanding of the benefits of the proposed method at the scales where its contribution (efficiency) might matter most.

Reviewer 3ehZ7/10 · confidence 3/52023-07-07

Summary

The authors propose a method based on deep learning that will speed up the prediction of physical systems by reducing the size of the grid and the number of tests to be carried out. The method consists of combining a graph neural network to predict some values with two orthogonal spline collocation, one spatial and one temporal, to predict values continuously.

Strengths

The paper is very clear and well written. The graphics are a great complement to the text and help to make it easier to understand. The choice of graph neural networks seems judicious in the context of a physical system that admits links and dynamic interactions between variables. As for orthogonal spline collocation, the theoretical guarantees and computational complexity justify their use. The results show improved predictions and increased speed on various physical problems. The authors are careful to point out the limitation of their approach.

Weaknesses

See limitations and questions.

Questions

1) Lines 176-177: I think there is a typo, 2-D domain is written twice 2) Line 203 and equation (2): you write $L$ in the equation and then you write a rounded "L", I think you mix the notation 3) To what extent is the availability of data for physical systems not too restrictive? 4) Could we imagine replacing GNNs with physical neural networks?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The impact of GNN on the overall method needs to be quantified a little more, the level of performance required by the neural network needs to be defined. This would make it easier to adapt the method, for example by changing the GNN model for another one. We can make the same point with OSC, trying to combine GNN with an alternative to show what the use of OSC brings specifically, in terms of runtime and performance. This would help justify the choice of these two methods over others.

Reviewer B9pj6/10 · confidence 3/52023-07-19

Summary

The presented manuscript proposes a novel architecture to predict the evolution of the dynamical system. The proposed method has two key ingredients: orthogonal spline collocation for both time and spatial dimensions and adaptive selection of the collocation points. Extensive experiments demonstrate the performance of the proposed method over alternatives in both prediction quality and runtime.

Strengths

The presented method is based on the natural encoder-decoder architecture and intermediate processor via message-passing layers. The main strength of the paper is the extensive experimental evaluation and sensitivity analysis of the results to the adjusting hyperparameters. The authors consider 3 datasets generated from different types of PDEs and 2 datasets with empirical data from monitoring Ocean and black sea. Their method shows the best results in all considered tests. Moreover, the proposed approach is faster than the competitors based on graph neural networks.

Weaknesses

The main weakness of the paper is its structure and presentation of the proposed method. In particular, 1) Too much attention is paid to the standard method for solving structured linear systems for computing interpolation coefficients. I suggest compressing such detailed descriptions since they are mostly textbook facts. At least a comparison with a general-purpose LU solver from PyTorch is strange since the result of such a comparison is obvious. Thus, section 3.4. can be eliminated or significantly compressed. 2) section 3.2 introduce the notation for the graph, but later there are no details on how this graph is constructed from snapshots 3) in the paragraph in lines 133-139 no information about how to select collocation points only the number of collocation points is estimated. So, it is unclear how to extract collocation points from the available samples 4) Figure 4.1 is entirely confusing. I do not find where is input, output, or ingredients of eq (1) and how they should be processed together to get the prediction. Please, specify what operations are done and in what order starting from the beginning and revise section 4.1 respectively. 5) also, please move to the main text architecture details of the encoder, decoder, and processor blocks and training setup from the supplementary materials I will increase the score if the proper revision of the text will be made in the rebuttal phase.

Questions

1) how to select the parameter $\beta$ in eq (3) ? Is it constant or adaptive too? 2) do you observe any relationship between the number of processor blocks and the resulting error? 3) why do you not consider the Chebyshev grid for collocation points and the corresponding barycentric Lagrange interpolation approach? Cheb grid should be more accurate than your static approach (Fig. 4.3 left) but less computationally intensive than the adaptive approach. 4) what item in the loss function $L_s$ or $L_i$ dominates and what are the possible explanations?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

1 poor

Contribution

3 good

Limitations

The authors provide the limitations of the proposed approach.

Reviewer Rs2G2023-08-15

Follow-up questions

Thank you for the reply. Can you expand the detail of MSE reported in Table 5.1 a bit more? Some of my questions on the MSE are: * For each of the experiment, how many time steps do the ground truth trajectories consist of and what is the cardinality of the in-time collocation points $\\{t_{0}, t_{1}, · · · , t_{K}\\}$? * On which data points is the MSE defined? Does it take the mean for all spatial points and all steps in each trajectory, and all the ground-truth trajectories in the test dataset? * Is the MSE error reported in Table 5.1 comparable to the MSE error defined on only in-space collocation points over $\\{t_{0}, t_{1}, · · · , t_{K}\\}$? I’m curious if the relatively high number of data points in the ground truth dataset would have an impact to reduce the MSE defined on only in-time and in-space collocation points. Also, for experiment of Navier-Stokes equations, the time horizon of the dataset is t=10 (in Appendix B.3), as opposed to t=5 reported in Table 5.1. Why did you choose t=5 as the rollout length? Are there any difficulties that prevent the proposed model from predicting 10 steps? Is the inference time reported in Figure 5.1 evaluated by the same models that produce the results for one of the experiments in Table 5.1?

Authorsrebuttal2023-08-16

Authors response for follow-up questions

Thanks for raising these fair concerns. Let us clarify! The 5-second rollout is, in fact, our mistake and only refers to the heat equation. We use different time stepping depending on each specific dataset - for instance, the Black Sea data has a time resolution of one full day. We summarize in the following table a more detailed overview of the experiments, such as total rollout time, space, and time resolution for both OSC space and time. | Experiment | Total Rollout Time | Ground Truth Step $\Delta t$ | Collocation Step $\Delta t$ | Ground Truth Resolution | Collocation Points Resolution | | --- | --- | --- | --- | --- | --- | | Heat Equation | 5 Seconds | 0.1 Seconds | 0.2 Seconds | 64 * 64 | 12 * 12 | | Damped Wave | 10 Seconds | 1 Seconds | 2 Seconds | 64 * 64 | 32 * 32 | | 2D Navier Stokes | 10 Seconds | 1 Seconds | 2 Seconds | 64 * 64 | 32 * 32 | | Ocean Currents | 10 Hours | 1 Hour | 2 Hours | 73 * 73 (3 features) | 36 * 36 (3 features) | | Black Sea | 10 Days | 1 Day | 2 Days | 5000 nodes (3 features) | 1000 Nodes (3 features) | | 3D Navier Stokes | 0.5 Seconds | 0.05 Seconds | 0.1 Seconds | 262,144 nodes (5 features) | 32,768 nodes (5 features) | > For each of the experiment, how many time steps do the ground truth trajectories consist of and what is the cardinality of the in-time collocation points $\{t_0, t_1, \cdots, t_K\}$? […] For experiment of Navier-Stokes equations, the time horizon of the dataset is t=10 (in Appendix B.3), as opposed to t=5 reported in Table 5.1. Why did you choose t=5 as the rollout length? Are there any difficulties that prevent the proposed model from predicting 10 steps? > Let us take the previous table. For example, in the 2D Navier-Stokes experiment, the ground truth trajectories consist of 10-time steps (10 seconds), and step $\Delta t$ is 1 second; while in our model, the neural network autoregressively performs the rollout 5 times with $\Delta t = 2$ seconds and the OSC will interpolate through the skipped steps. In other words, the ground truth consists of $\mathbf{Y}^{t}, t\in\lbrace 1, 2, 3, 4, 5, 6, 7, 8, 9, 10\rbrace$ and the neural networks in our model gives $\mathbf{\tilde{Y}}^{t_{\mathrm{interpolate}}}, t_{\mathrm{interpolate}}\in\lbrace2, 4, 6, 8, 10\rbrace$ and interpolate to $\mathbf{\hat{Y}}^{t} = \mathrm{TimeOrientedOSC}(\mathbf{\tilde{Y}}^{t_{\mathrm{interpolate}}})$. This is similar to the space resolution, in which GraphSplineNets predict only a subset of the domain and interpolate through the rest. We will add this information to our manuscript. > On which data points is the MSE defined? Does it take the mean for all spatial points and all steps in each trajectory, and all the ground-truth trajectories in the test dataset? > The MSE is defined based on the ground truth resolution and time steps, i.e., all spacial points and all time steps in each trajectory. > Is the MSE error reported in Table 5.1 comparable to the MSE error defined on only in-space collocation points over  $\{t_0, t_1, \cdots, t_K\}$? I’m curious if the relatively high number of data points in the ground truth dataset would have an impact to reduce the MSE defined on only in-time and in-space collocation points. > As we mentioned in the previous question, the MSE reported in Table 5.1 is defined on all spacial points and all time steps in each trajectory. We report this small experiment decoupling the effect of space and time collocation separately. | | MGN | MGN+SpaceOSC only | MGN+TimeOSC only | MGN+SpaceOSC+TimeOSC (GraphSplineNets) | | --- | --- | --- | --- | --- | | MSE(x10^{-3}) | 3.01±0.38 | 2.87±0.54 | 1.27±0.08 | 1.14±0.11 | | Runtime [s] | 6.99±0.12 | 1.88±0.11 | 4.63±0.15 | 1.38±0.10 | The insight is that space collocation is mostly responsible for efficiency, while time collocation yields comparatively more accuracy. > Is the inference time reported in Figure 5.1 evaluated by the same models that produce the results for one of the experiments in Table 5.1? > Correct, we use the heat equation experiment to measure the times reported in Figure 5.1, while the model structure is the same across experiments. We noticed that these trends hold in general for the experiments. Please let us know if you have further questions or concerns; we will be happy to address them!

Reviewer Rs2G2023-08-17

Good paper with solid experiments

Thank you for providing detailed overview of the experiments and the results on ablation study. Given the rebuttals and reviews, I raised my score. The additional reports made my understanding much clearer. It’s also very interesting to see that incorporating just one of SpaceOSC and TimeOSC into MGN improves performance of the baseline MGN in both of the metrics. These results support the idea of using both space and time-oriented OSC. I think this paper is above the acceptance threshold.

Authorsrebuttal2023-08-17

Author response

Thank you for your constructive feedback and the time you have taken to understand our work. We truly appreciate your positive remarks. We noticed that while you mentioned raising the score in your comments, it may not have been updated on the system. Could you please check and confirm the score update on OpenReview? It can be done by clicking on "Edit -> Official Review" on the main review you submitted. We remain available for any further support. Thank you in advance! --- Edit: we checked it, thanks a lot!

Reviewer jrMD2023-08-16

Acknowledgement of rebuttals and reviews

I have read the reviews and rebuttals, and I have updated my score. Based on the rebuttal, I have some additional comments/questions that may help me advocate for this paper in the discussion phase. Regardless, given the rebuttal and original draft, I believe this paper will have high impact on the ML-for-simulation area of AI. Follow-up questions: 1. Given that you agree that the spatial coarsening allows for communication across greater spatial extents, could you please briefly compare and contrast your method with multiscale MGN (Fortunato et al., 2022)? Such a commentary would seemingly improve your manuscript's related work section and clarify your innovation. For example, you both apply MGNs on downsampled domains, but you use different methods for coarsening, etc. 2. I like the Pareto frontier plots in your new figure (Figure R2)! It’s great to see that you can claim a Pareto improvement. Regarding this figure, you say, “For the sake of a fair comparison, the baseline processor remained consistent across all models.” Does this mean that all models have the same weight values, or that all models have the same processor architecture but each model has unique weights? 3. Thanks for adding the 3D analysis! It led me to realize that I am unsure about if and when an MGN model used the spatially and temporally downsampled domain that the GSN used. Could you please let me know if the following is correct? - MGN in Table R2 uses 3 MP steps on the original domain, and GSN uses 3 steps on the coarsened domain (leading to a large speed benefit). - The MGN method that you call "MGN+OSC(Post)" in Table 5.1 is operating on the same (spatially and temporally downsampled) domain as GSN (MGN+OSC).

Authorsrebuttal2023-08-16

Authors response for the acknowledgement and follow-up questions

Thank you for your feedback. We are thrilled to receive such a positive evaluation! We will answer your rollow-up concerns in the following sections: > 1. Given that you agree that the spatial coarsening allows for communication across greater spatial extents, could you please briefly compare and contrast your method with multiscale MGN (Fortunato et al., 2022)? Such a commentary would seemingly improve your manuscript's related work section and clarify your innovation. For example, you both apply MGNs on downsampled domains, but you use different methods for coarsening, etc. Thanks for your question! The MultiScale MeshGraphNets (MS-MGN) is an interesting work to discuss and compare against our model. Let us divide the differences into space and time, as our OSC does. ### Space Message passing is performed differently - both methods can increase efficiency but in a different way: - MS-MGNs: they perform message passing twice, both at low and high resolution - GraphSplineNets: we perform message passing at a low resolution only and use OSC in space to obtain high resolution ### Time We also have differences in how autoregressive rollouts are done: - MS-MGNs: they use a fixed-step integrator only - GraphSplineNets: we also use a fixed-step integrator but with a coarse grid, then employ the OSC to obtain higher time resolution MS-MGNs do not employ coarser time resolution, allowing GraphSplineNets for even greater speedups. Finally, we would like to mention that we plan to extend future works GraphSplineNets with MS-MGN. We believe that GraphSplineNets can enable MS-MGN to achieve even better Pareto-fronts with multi-level hierarchical prediction! > 2. I like the Pareto frontier plots in your new figure (Figure R2)! It’s great to see that you can claim a Pareto improvement. Regarding this figure, you say, “For the sake of a fair comparison, the baseline processor remained consistent across all models.” Does this mean that all models have the same weight values, or that all models have the same processor architecture but each model has unique weights? Thank you for your positive feedback on the Pareto frontier plots in Figure R2; we're pleased that you found value in them! Regarding your question, the phrase "For the sake of a fair comparison, the baseline processor remained consistent across all models" indicates that while all models share the same processor architecture - each model is optimized separately. > 3. Thanks for adding the 3D analysis! It led me to realize that I am unsure about if and when an MGN model used the spatially and temporally downsampled domain that the GSN used. Could you please let me know if the following is correct? > - MGN in Table R2 uses 3 MP steps on the original domain, and GSN uses 3 steps on the coarsened domain (leading to a large speed benefit). > - The MGN method that you call "MGN+OSC(Post)" in Table 5.1 is operating on the same (spatially and temporally downsampled) domain as GSN (MGN+OSC). Certainly! We appreciate your questions and observations and will further clarify the experimental setup. Indeed, your understanding is correct: - In Table R2, the MeshGraphNet (MGN) employs 3 Message-Passing (MP) steps on the original resolution data points. On the other hand, our proposed GraphSplineNets (GSN) utilizes 3 MP steps on the collocation points, which are derived from a coarsened version of the original data. This approach contributes to the observed substantial speed improvement. - The MGN+OSC(Post) in Table 5.1 operates on the same spatially and temporally downsampled domain as GSN (MGN+OSC). Notably, the key distinction lies in the loss consideration: MGN+OSC(Post) focuses solely on the collocation points' loss since the OSC approaches are applied post-processing, whereas GSN incorporates the loss of all original resolution data points. Specific to your first statement, to make it more clear, we prepared in the following table a more detailed overview of the experiments, such as total rollout time, space, and time resolution for both OSC space and time. We will add this table to the appendix in the revised version. | Experiment | Total Rollout Time | Ground Truth Step $\Delta t$ | Collocation Step $\Delta t$ | Ground Truth Resolution | Collocation Points Resolution | | --- | --- | --- | --- | --- | --- | | Heat Equation | 5 Seconds | 0.1 Seconds | 0.2 Seconds | 64 * 64 | 12 * 12 | | Damped Wave | 10 Seconds | 1 Seconds | 2 Seconds | 64 * 64 | 32 * 32 | | 2D Navier Stokes | 10 Seconds | 1 Seconds | 2 Seconds | 64 * 64 | 32 * 32 | | Ocean Currents | 10 Hours | 1 Hour | 2 Hours | 73 * 73 (3 features) | 36 * 36 (3 features) | | Black Sea | 10 Days | 1 Day | 2 Days | 5000 nodes (3 features) | 1000 Nodes (3 features) | | 3D Navier Stokes | 0.5 Seconds | 0.05 Seconds | 0.1 Seconds | 262,144 nodes (5 features) | 32,768 nodes (5 features) | We will be happy to clarify additional questions and concerns if there are any!

Authorsrebuttal2023-08-19

Request for Feedback

As the author-reviewer discussion phase is coming to an end, please let us know if there are further questions or concerns to address. Thanks in advance for your time and consideration!

Reviewer B9pj2023-08-20

Dear authors, Thanks for the thorough and detailed response! I wait for the revised version of the text upon acceptance and have changed the evaluation score.

Authorsrebuttal2023-08-19

Request for Feedback

As the author-reviewer discussion phase is coming to an end, please let us know if there are further questions or concerns to address. Thanks in advance for your time and consideration!

Reviewer 3ehZ2023-08-20

I'd like to thank the authors for their clear answer. I'm modifying my rating, taking into account the other reviews as well. I have no further questions.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC