Derivative-enhanced Deep Operator Network

The deep operator networks (DeepONet), a class of neural operators that learn mappings between function spaces, have recently been developed as surrogate models for parametric partial differential equations (PDEs). In this work we propose a derivative-enhanced deep operator network (DE-DeepONet), which leverages derivative information to enhance the solution prediction accuracy and provides a more accurate approximation of solution-to-parameter derivatives, especially when training data are limited. DE-DeepONet explicitly incorporates linear dimension reduction of high dimensional parameter input into DeepONet to reduce training cost and adds derivative loss in the loss function to reduce the number of required parameter-solution pairs. We further demonstrate that the use of derivative loss can be extended to enhance other neural operators, such as the Fourier neural operator (FNO). Numerical experiments validate the effectiveness of our approach.

Paper

Similar papers

Peer review

Reviewer nyqb7/10 · confidence 5/52024-06-15

Summary

The authors propose incorporating derivative information into the training of DeepONets to improve predictive accuracy on PDE problems. Instead using an encoder-decoder neural operator architecture as in prior work, the authors use the DeepONet architecture as well as incorporate spatial derivative information on top of functional derivatives previously used. The authors provide numerical comparisons with other models using both types of derivative training as well as results for two types of input space dimensionality reduction methods.

Strengths

- The paper is very well written, well structured, and with the right level of detail and understanding to describe the setup in the preliminary section without lingering too long. The authors clearly have a deep understanding of all components of this method and convey the steps well in the space given. - The appendix is thorough, high quality, and provides relevant details to better comprehend and reproduce the results in the main text. I found the visualizations in section B.6 particularly compelling for the low data case with vanilla FNO and DeepONet failing compared to the gradient enhanced models. - The improvement in dm prediction for control is an important but overlooked metric for neural operators which primary focus on the output solution accuracy despite the gradient information being needed in real-world problems. The authors may want to consider the existing real-world example of DeepONets for optimization in [1], shown in Figure 7 to require gradient information of the objective w.r.t to the input parameters to optimize aerodynamic shapes which could benefit from this methodology which is a strength over accuracy improvement alone. [1] Shukla, Khemraj, et al. "Deep neural operators as accurate surrogates for shape optimization." Engineering Applications of Artificial Intelligence 129 (2024): 107615.

Weaknesses

- The cost of generating the dm and dx labels with which to incorporate into training was 2-3 times greater than the cost of solving the PDE itself given Table 1. Of course, as with all neural operators the question is the tradeoff between offline data generation and training versus the benefit of quick online predictions so this cost could be mitigated given the accuracy benefit it provides. Nevertheless, its quite the increase and one may wonder if its worth doing at that point compared to just using the high fidelity FEM solver for N number of problems. Unless N is high enough, the tradeoff looks less appealing here. - I would like to see the authors rephrase their findings in the context of the results in B.4 which show that the dx information provides minimal to no benefit compared to the dm training, and it some cases it even makes the accuracy worse than dm alone. This comparison of dx and dm alone should be stated in the main text as it appears to be critical information. Given that beside the solution accuracy, the dm accuracy is vastly more important that the dx accuracy to perform PDE-constrained optimization problems, it begs the question why dx regularization is done at all here. This then begs the question how novel this is compared to DINO which already trained on dm information. The authors claim Sobolev learning is novel on top of the DINO results but it does not appear impactful. The authors need to address this. - Following onto the prior point, I am on the edge regarding impact and novelty here. That being said I still think the work is informative and of high enough quality to publish, albeit without as strong of an impact due to the main benefits shown to be in line with the existing method DINO. Looking at DE-DeepONet vs. DINO in Figure 1, there is not much difference, and in Table 2 DINO is shown to be much faster. I would consider changing my rating if this was adequately addressed.

Questions

- Could the authors please include total wall-clock time to Figure 2 in addition to per epoch. This would be helpful in comparing to the data cost in Figure 1 which is a total. Additionally, the authors may want to consider including a row for end-to-end data and training time such that the DeepONet and FNO include the PDE solution generation time and DINO and DE-DeepONet include the data for all three loss terms. This way the reader can very easily compare the total cost associated with making predictions with a DeepONet versus DE-DeepONet. I think this would benefit the manuscript since it mitigates the cost of the dm dx label generation. - How can the epochs only be 1,000 in B.3? In the original DeepONet paper 50,000 – 500,000 iterations were used. Are all models converged to fairly compare them? What do the convergence plots look like? It would be helpful to see them. - Is the vanilla DeepONet and FNO also trained with the dimensionality reduction technique on the input, and if so, which one? Figure 1 shows KLE and ASM for DE-DeepONet, and DINO is stated to use ASM, but what about the baseline models. How is it a fair comparison if the inputs are not the same? - The usage of CNN for the trunk and ResNet for the branch should not be hidden in the appendix. The vanilla DeepONet uses a fully-connected NN for both the trunk and branch and that would be assumed here, please mention it in 3.1. Additionally, what justification is there for those choices? The CNN for the trunk makes sense to construct the basis but I’m curious about the ResNet for m? - The authors should consider the following papers [2,3]. In [3] the authors train DeepONets the physics-informed (dataless) way using gradient information to obey the governing equation. In [2], the authors train a PINN with gradient-enhanced information which could also be done with the previously mentioned PI-DeepONet in [3]. How might DE-DeepONet methodology be incorporated into these models and in what ways is it distinct from them? It would be nice to see this discussion in the main or appendix text. [2] Yu, Jeremy, et al. "Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems." Computer Methods in Applied Mechanics and Engineering 393 (2022): 114823. [3] Wang, Sifan, et al. "Learning the solution operator of parametric partial differential equations with physics-informed DeepONets." Science advances 7.40 (2021): eabi8605. - How were the gradients computed using automatic-differentiation (AD)? I do not see the ML package mentioned, a package like Jax has more accurate and substantially faster AD than one like PyTorch, see [4] Table 6 and Figure 15. The authors may want to consider this to improve performance and reduce the overhead cost for dm and dx label generation. [4] Jagtap, Ameya D., et al. "How important are activation functions in regression and classification? A survey, performance comparison, and future directions." Journal of Machine Learning for Modeling and Computing 4.1 (2023).

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

2

Limitations

- Adequately described in manuscript.

Reviewer WTPz6/10 · confidence 3/52024-07-11

Summary

This manuscript introduces a derivative-enhanced deep operator network that utilizes derivative information to improve prediction accuracy.

Strengths

S1) This paper presents a new method for improving the accuracy of approximating the output function for DeepONet, along with its directional derivative relative to the input function and spatial derivatives. S2) The suggested enhancements improve the performance of the basic DeepONet and perform well in scenarios with limited training samples.

Weaknesses

W1) The baselines appear limited as recent benchmarks like GNOT, LSM, ONO, Transsolver, etc., have ye t to be compared. Numerous variations of FNO have been introduced, such as FFNO, UNO, CoNO, GFNO, UFNO, etc. W2) The benchmarks dataset appears to be restricted to datasets for Navier-Stokes equations and hyperelasticity, which are not open-source. W3) ASM is costly in computation, and derivative information is necessary when the PDE form is unknown. Additionally, it needs to scale better with dimensions. W4) The proposed method builds upon the DINO paper by introducing additional informed losses. W5) The new losses were found effective only for DeepOnet and lacked experimental evidence compared to FNO.

Questions

Q1) The primary distinction between DINO and the proposed approach lies in utilizing DeepOnet and spatial derivatives. However, spatial derivatives were also introduced in [1]. Also, have you used the approximation for spatial derivative as used in [1]? [1] DEEP MULTI-SCALE VIDEO PREDICTION BEYOND MEAN SQUARE ERROR. Michael Mathieu, Camille Couprie & Yann LeCun. Q2) What is the impact of proposed losses when combined with other operators like FNO? Do they enhance the performance of different operators? Q3) Could approximating the Frobenius norm along a random direction be analogous to score-based diffusion models? Q4) Is there a specific reason why adding new loss terms relative to L2 still results in poorer performance than FNO when dealing with hyperelasticity as training samples increase? Q5) How does incorporating this noisy gradient version aid in training?

Rating

6

Confidence

3

Soundness

2

Presentation

2

Contribution

3

Limitations

Yes, the authors adequately addressed the limitations.

Reviewer WTPz2024-08-09

Thank you for the authors' addressing of my comments, weaknesses, and questions. New experimental results were conducted to help the readers understand the proposed losses more appropriately. Concerning the last question, I wanted to see how proposed losses perform when we have noisy datasets. Does it improve the performance?   Although the paper should be revised with the above information before publication, the author continues to polish the manuscript. Lastly, in the revised version, try to make it a little bit more readable for someone not from a purely mathematical background. So, I am willing to raise the score. Thank you for the responses. I have raised my score and hope to see the new changes incorporated into the revised version.

Authorsrebuttal2024-08-09

We greatly appreciate your positive feedback and support. Considering how dm loss performs with noisy datasets is indeed an interesting and meaningful direction to explore, but we believe it would be more natural to study this within the context of, say, inverse problems, which is beyond the scope of our manuscript (our work only considers enhancing surrogate models for forward operators governed by known PDEs). In terms of readlability, we notice some parts of the manuscript might be too dense or abstract for readers that are not from heavy math backgrounds. When revising the manuscript to includer the new results, we will also pay attention to adding more intuitive explanations (although sometimes at the expense of accuracy), concrete examples, or computational details to help a broader audience understand, particularly in how to derive both DE-DeepONet's (same for FNO+dm) and DINO's dm labels in practice.

Reviewer 41ML5/10 · confidence 3/52024-07-12

Summary

The paper proposes an extension of deep operator networks (DeepONets) enhanced by matching the derivatives of the output function with respect to the input function, for example, the derivatives of the PDE solution w.r.t. the input coefficient. To make the computation tractable, the dimensionality of the input function is reduced via a dimension reduction technique, active subspace method, avoiding expensive Jacobian evaluations.

Strengths

- The paper is written clearly, explaining the relevant preliminaries in a concise way and delivering the core contributions of the proposed method. - The proposed method leverages well-established classical methods (such as KLE and ASM) for building low-dimensional surrogate and incorporate them into advanced neural operator settings (DeepONet). - Although not very extensive, the paper present a set of numerical results from experimentation on important benchmark problems.

Weaknesses

- The results of the empirical evaluation do not seem to be strong enough to make the proposed method look like an effective alternative to the existing method (FNO). The gain obtained at the cost of increased computation seems to be marginal. - Although KLE or ASM is a standard method for reducing dimensions of a field, the number of required bases depends on smoothness or regularity of the field and, thus, the proposed method could be benefitted in some specific scenarios. Some discussions on where this method could be benefitted and where this method would struggle are required. - Presenting wall time per epoch is informative, but it would provide a more complete picture if the authors could provide the wall time for training to achieve a certain level of accuracy. - Regarding the statement “when the training data are limited, the increased computation cost is compensated for a significant reduction of errors”: It would be informative if the authors could provide a summarizing figure presenting a result for varying number of training instances and show that the benefits of the proposed method is more pronounced in the data-scarce regime. - Although it is not in the main body of the paper, Figure 2 seems to miss the entire information. This omission undermines the completeness of the paper and gives the impression that the paper was prepared hastily.

Questions

- Eq (4). Does the branch net take the reduced input vector as an input? Could the branch net just take the original input and approximate the Frechet derivative using the reduced representation? - Can the authors describe how the Frechet derivative is computed (the derivative with respect to the reduced input) to collect the ground truth data? - Is the reason for using the relative error for measuring the loss for preventing numerical issues caused by different numerical scales?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

- A thorough empirical investigation on in which scenarios this proposed method outperforms baseline (and also the opposite scenarios) would be needed.

Reviewer row26/10 · confidence 2/52024-07-22

Summary

The paper introduces a modified version of DeepONet, termed Derivative-Enhanced DeepONet, by incorporating derivative terms relative to the input function and spatial domain into the loss function. It outlines a practical method for calculating these derivatives, which serve as supplementary supervision terms in the training process. The authors demonstrate enhanced performance over other Neural Operator baselines with this approach, particularly in low-data scenarios, using two datasets centered on hyperelasticity and Navier-Stokes equations.

Strengths

- The paper is engaging and well-motivated. - The methodology presented appears novel and is supported by solid theoretical underpinnings. It demonstrates consistent enhancements across various metrics, particularly with the hyperelasticity dataset.

Weaknesses

- As a reviewer with limited expertise in functional analysis, I found the sections detailing the supplementary supervision terms challenging to comprehend. A more introductory overview explaining how the derivative ground truths are derived would be beneficial for clarity and accessibility.

Questions

- Could you provide a higher-level explanation of how the derivative labels are calculated? Are they derived from the underlying equation? - On line 325, you mention reducing training costs by introducing additional derivative losses in the later stages of training. Do you have any data or figures to support this claim? - Why does supervision of the derivatives become less significant as more data becomes available?

Rating

6

Confidence

2

Soundness

3

Presentation

2

Contribution

2

Limitations

Limiatations have been addressed

Area Chair W6BS2024-08-08

Dear reviewers, could you have a look at the authors response and comment on them if you have done so, yet. thanks in advance your area chair

Reviewer nyqb2024-08-08

I thank the authors for their response to my review and particularly the new results provided in the overarching rebuttal. In the context of these items, a number of my points have been addressed (removing dx from training, gaining improvement over DINO with the new FNO + dm results, including convergence plots, etc.) and therefore I am raising my score from 6 $\rightarrow$ 7 under the condition these new (necessary) results make it into the revised manuscript before publication and don't die here.

Authorsrebuttal2024-08-08

We sincerely appreciate your feedback and support. We will definitely incorporate the new results into the revised manuscript.

Reviewer 41ML2024-08-12

Thank you for providing the clarification and the experimental results. The new information on training/inference time is informative and clears my concern to some extent. I adjusted the score accordingly.

Authorsrebuttal2024-08-12

Thank you for your updated response. We're glad that the new results help address your concern to some extent. When revising the manuscript to include the new results, we will also keep your (and other reviewers') feedback in mind.

Reviewer row22024-08-13

I thamk the authors for their clarifications and additional experiments, which help me understand the paper better and answer my questions. I have raised my score from 5 to 6.

Authorsrebuttal2024-08-13

We thank the reviewer for the updated response!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC