DiffPhyCon: A Generative Approach to Control Complex Physical Systems

Controlling the evolution of complex physical systems is a fundamental task across science and engineering. Classical techniques suffer from limited applicability or huge computational costs. On the other hand, recent deep learning and reinforcement learning-based approaches often struggle to optimize long-term control sequences under the constraints of system dynamics. In this work, we introduce Diffusion Physical systems Control (DiffPhyCon), a new class of method to address the physical systems control problem. DiffPhyCon excels by simultaneously minimizing both the learned generative energy function and the predefined control objectives across the entire trajectory and control sequence. Thus, it can explore globally and plan near-optimal control sequences. Moreover, we enhance DiffPhyCon with prior reweighting, enabling the discovery of control sequences that significantly deviate from the training distribution. We test our method on three tasks: 1D Burgers' equation, 2D jellyfish movement control, and 2D high-dimensional smoke control, where our generated jellyfish dataset is released as a benchmark for complex physical system control research. Our method outperforms widely applied classical approaches and state-of-the-art deep learning and reinforcement learning methods. Notably, DiffPhyCon unveils an intriguing fast-close-slow-open pattern observed in the jellyfish, aligning with established findings in the field of fluid dynamics. The project website, jellyfish dataset, and code can be found at https://github.com/AI4Science-WestlakeU/diffphycon.

Paper

Similar papers

Peer review

Reviewer Mp9v5/10 · confidence 4/52024-06-22

Summary

This paper proposes an algorithm for controlling complex physical systems, particularly in long-term settings. The method is based on diffusion models and energy methods, utilizing data generated by traditional finite difference methods.

Strengths

* The problem studied in this paper is important and well-motivated. * The presentation is clear and easy to follow. * Detailed experimental results of the proposed algorithm are presented in both main text and appendices.

Weaknesses

* The method proposed in the paper does not demonstrate its strength in challenging tasks. I agreed with the claim that the fundamental challenge of simulating complex physical systems, which are high-dimensional and highly nonlinear. However, the current paper does not address these problems convincingly for the following two reasons: * First, both numerical examples are low-dimensional so that the curse of dimensionality does not affect much. * Second, their dynamics are not very complex. At first glance, the 2D Jellyfish task appears complex, as it involves solving the 2D Navier-Stokes equation, which might yield turbulent solutions. However, I realized that the optimization focuses on the jellyfish itself. I was wondering if there is a chance to obtain a good solution even if the simulation of fluid dynamics is not accurate. If this is the case, the results of this example may not provide strong evidence that the algorithm can handle complex dynamics. Additionally, this paper does not present the error analysis for the fluid dynamics simulation, only stats of energy comparison are provided. * This algorithm depends on the data generated by finite difference methods, hence suffering from the curse of dimensionality.

Questions

* What is the energy term $E_\theta(u, w, c)$? * In the 2D Jellyfish task, what is the value of the kinematic viscosity $\nu$? * Is it possible to make the proposed algorithm independent of classical methods (e.g., the finite difference method used in this paper) for data generation?

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

This paper has addressed its limitations in the main text.

Reviewer n6Yf8/10 · confidence 4/52024-07-03

Summary

This paper introduces Diffusion Physical Systems Control (DiffPhyCon), where diffusion models are used to generate a near-optimal controller for a system described by a partial differential equation (PDE). In this generative approach, a learned generative energy function and control objective is minimized. Additionally, a prior reweighting technique is developed to mitigate the effect of prior distribution of control sequences.

Strengths

The paper is well-written and the method is well-motivated. The performance of the developed generative approach is demonstrated using extensive simulation studies on 1D Burgers equation and 2D Jellyfish movement control. The efficiency of this method is compared with a decent number of baselines. Additional experiments are provided to study the effects of hyperparameters.

Weaknesses

The contribution of the paper is not clear compared to the following recently published result in ICLR: Wei, L., Hu, P., Feng, R., Du, Y., Zhang, T., Wang, R., Wang, Y., Ma, Z.M. and Wu, T., Generative PDE Control. In ICLR 2024 Workshop on AI4DifferentialEquations In Science. The method looks identical to the generative PDE control result in the above reference which is published and peer-reviewed. The additional contribution, if any, is too minor for NeurIPS. For this reason, I cannot recommend the paper to be accepted for NeurIPS. However, I do think this paper would have been a contribution to NeurIPS otherwise. ----- UPDATE: It appears the ICLR workshop is not considered an archival publication. With this consideration, I update my score to 7.

Questions

Besides the big question of novelty in comparison to the "Generative PDE Control" paper from ICLR, I have the following comments for the authors: 1. The authors acknowledge the limitation that DiffPhyCon presently operates in an open-loop manner. This limitation by itself is not a problematic, but the claims made in the introduction need revision in that case. Due to robustness concerns, a controller is seldom implemented open-loop even for simple systems, let alone complex physical systems. However, open-loop optimal control design can be applied to motion planning or trajectory generation objectives. In the context of this paper, the generative approach could be used to construct a nominal pre-planned open-loop control sequence term which can then be added to a feedback term (e.g. PID). Thus, I suggest to revise the introduction to claim planning as the objective. For more information about motion planning, the authors are referred to Chapter 12 (Motion Planning for PDEs) from the reference: Krstic, M. and Smyshlyaev, A., 2008. Boundary control of PDEs: A course on backstepping designs. Society for Industrial and Applied Mathematics. 2. In line 72 of page 2, the authors claim "...DiffPhyCon facilitates global optimization of long-term dynamics...". It is not clear how the globality (or even non-triviality) of near-optimal control sequence is guaranteed to be achieved. I am not sure why Algorithm 1 would not yield local optima.

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

Yes, the authors have stated the limitations.

Reviewer DFgi8/10 · confidence 5/52024-07-13

Summary

The model proposed a variant of diffusion models to control complex dynamical systems. Its contributions are threefold. First, the proposed model could optimize the trajectory and control sequence simultaneously. Second, it proposed a reweighting prior technique to generate a superior control sequence to the training dataset. Third, it contributes a benchmark dataset, jellyfish movement control, for the complex dynamical control community.

Strengths

Novelty: The novelty lies in two parts. A variant of the diffusion model is proposed to minimize the energy function and control objectives simultaneously. Moreover, it proposes a prior reweighting algorithm to allow the sample to enhance the sampling of good but low-probability trajectories. Clarity: It is clear in illustrating the methodologies and the experiments. Quality: It provides comprehensive details on the background of the model and dataset. The experiment results substantiate its claim. Significance: The proposed method can potentially control complex systems with a generative model with reduced cost and good long-term control sequences.

Weaknesses

Several questions need to be clarified; see details in the questions section.

Questions

1. How is the offline dataset generated? Does it include the optimal control trajectory? What if the offline collected training control trajectory is too far from the optimal one? I am trying to understand if your model "finds" the existing trajectory in the dataset or if it "stitches" different pieces of trajoetiers and generates one that doesn't exist in the dataset. 2. For the cases in this paper, using the online control algorithm is not too expensive. Online RL for these cases could also give global optimal control results. Could you compare the computational cost of your proposed method and an online RL method to demonstrate the benefits against the online RL algorithm? 3. For the diffusion-generated trajectory, how did you get Figure 4? Did you input the diffusion-generated trajectory to the numerical solver (Lilly-Pad) to get the physics simulation result? Or did you take the diffusion-generated flow field result as the "physics trajectory"? 4. I didn't quite get the intuition behind Figure 2. Why does reducing $\gamma$ to less than 1 shift the red point from margin to center?

Rating

8

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitations are explicitly mentioned in the paper.

Reviewer DFgi2024-08-13

Rebuttal period

Thanks for providing a detailed rebuttal to my questions and general proof. The quality and relevance of the rebuttal resolved my concerns. Together with the high-quality manuscript, which involves rich detail, it is definitely a paper beyond the acceptance level. Therefore, I would raise my score to 8. I would strongly encourage the author to publish the final version of the code to facilitate future research.

Authorsrebuttal2024-08-13

Thank you for your positive feedback and for raising your score. We are delighted to hear that our rebuttal addressed your concerns and that you found the manuscript to be of high quality. We appreciate your encouragement to publish the final version of the code. We plan to release the code upon acceptance of the paper, and we will make sure to include clear documentation and instructions for usage. Once again, thank you for your constructive comments and support.

Reviewer 8Er56/10 · confidence 4/52024-07-15

Summary

The method learns the energy function, $\epsilon_\theta$, which is used in the energy optimization target to generate the control sequences and system trajectory. A denoising network is trained to approximate the gradient of $\epsilon_\theta$. The network and optimization framework takes a global state of u and w of all times steps. The training and inference follow the process of a diffusion model. The paper also introduces prior reweighting to enable the discovery of control sequences that diverge significantly from training. Another network $\epsilon_\phi$ to learn the prior distribution of the control sequences is introduced. The proposed method is tested by two systems (1D Burgers’ equation and 2D jellyfish movement control). The experiments compared different control methods.

Strengths

- The paper develops a generative method to control complex physical systems. The method optimizes system trajectory and control sequences jointly in the entire horizon by diffusion models. A prior reweighting technique is also proposed to improve the model's generalization ability. - The results of the jellyfish systems are strong. The proposed methods generates realistic control sequences and trajectory that align with the established findings in fluid dynamics. - The paper generated a dataset for the jellyfish system, which contributes to the benchmark of this area.

Weaknesses

- The experiments are relatively limited - only include one 1D example and one 2D example. - I have minor reservations about the novelty of the paper since the primary method relies on the existing diffusion model.

Questions

- Compared with other methods, does the proposed method use less, similar, or more time to train (for the learning-based methods) and inference? - Is it easy or hard to extend this method to 3D systems? In the 2D example, a 3D U-Net is used. Does this make it hard to apply this method to 3D systems as it requires a much larger 4D network? - In the prior reweighting method, are $\epsilon_\theta$ and $\epsilon_\phi$ trained jointly, or is $\epsilon_\theta$ trained first and then $\epsilon_\phi$?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors discussed the limitations in the paper.

Authorsrebuttal2024-08-07

Theoretical analysis of the proposed prior reweighting technique and proof

Several reviewers (8Er5, DFgi, n6Yf) are concerned with the theoretical property of our method. Here, we present the **formal theoretical analysis of the prior reweighting technique**. Consider a pair $(u,w)$ sampled using the prior reweighting technique with hyperparameter $\gamma$: $(u,w)\sim p_{\gamma}(u,w) = p(u|w) p^{\gamma}(w)/C_{\gamma}$. Denote $J^*$ as the global minimum of the control objective $J$. Define $Q(\varepsilon)$ as the "$\epsilon$-optimal" solution set of $(u, w)$ such that $J(u, w) - J^* \leq (\epsilon) $, whose complement set is $Q(\varepsilon)^c$, and denote $\mathbb{I}\_{Q(\varepsilon)}(u,w)$ as its indicator function, i.e. $\mathbb{I}\_{Q(\varepsilon)}(u,w)=1$ if $(u,w)\in Q(\varepsilon)$; otherwise 0. Define $Y$ to be the random variable of "whether use $J$ as a guidance for sampling", namely, $ p(Y|u,w)=\begin{cases} e^{-J(u,w)} / Z, &Y=1 \\\\ 1 - e^{-J(u,w)} / Z, & Y=0 \end{cases} $ Consider $E(\gamma)=\mathbb{E}\_{(u,w)\sim p_{\gamma}(u,w)}[\mathbb{I}\_{Q(\varepsilon)}(u,w)|Y=1]$, which indicates the expectation of getting an $\epsilon$-optimal solution by using the prior reweighting technique with $\gamma$ under the guidance of $J$. Define $ F(\gamma) = \frac{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w) \ln(p(w))]}{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)]} - \frac{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w) \ln(p(w))]}{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)]} $. Then, we have the following theorem: **Theorem**: Assume $E(\gamma)$ is a smooth function, then: - (i): If $F(1) < 0$, there exists $\gamma_{-} < 1$, s.t., $E(\gamma_{-})> E(1)$; - (ii): If $F(1) > 0$, there exists $\gamma_{+} > 1$, s.t., $E(\gamma_{+})> E(1)$. **Proof** \begin{align} E(\gamma)= &\int \mathbb{I}\_{Q(\varepsilon)}(u,w) p\_{\gamma}(u,w|Y=1) \mathrm{d}(u,w) \\\\ = &\int \mathbb{I}\_{Q(\varepsilon)}(u,w) \frac{p\(Y=1|u,w) p\_{\gamma}(u,w)}{p\(Y=1)} \mathrm{d}(u,w) \\\\ = &\frac{\mathbb{E}\_{(u,w)}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)]}{\mathbb{E}\_{(u,w)}[p(Y=1|u,w)]} \\\\ = &\frac{\mathbb{E}\_{(u,w)\}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p\(Y=1|u,w)]}{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p\(Y=1|u,w)] + \mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)]} \\\\ = &\frac{1}{1 + \frac{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p\(Y=1|u,w)]}{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p\(Y=1|u,w)]}} \end{align} Define \begin{align} G(\gamma) = &\frac{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)]}{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)]} \\\\ = &\frac{\mathbb{E}\_w[\mathbb{E}\_u[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)|w]]}{\mathbb{E}\_w[\mathbb{E}\_u[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)|w]]} \\\\ = &\frac{\int \mathbb{E}\_u[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)|w] p^{\gamma}(w) \mathrm{d}w}{\int \mathbb{E}\_u[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)|w] p^{\gamma}(w) \mathrm{d}w} \end{align} Then $E(\gamma) = \frac{1}{1 + \frac{1}{G(\gamma)}}$. Since $G(\gamma)>0$, $E(\gamma)$ and $G(\gamma)$ have the same monotonicity. \begin{align} G'(\gamma) = &\frac{\int \mathbb{E}\_{u}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)|w] p^{\gamma}(w) \ln(p(w)) \mathrm{d}w \mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)]}{(\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)])^2 } \\\\ & -\frac{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)] \int \mathbb{E}\_{u}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)|w] p^{\gamma}(w) \ln(p(w)) \mathrm{d}w}{(\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)])^2} \\\\ = & \frac{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w) \ln(p(w))] \mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)]}{(\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)])^2} \\\\ & - \frac{\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)}(u,w) p(Y=1|u,w)] \mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w) \ln(p(w))]}{(\mathbb{E}\_{u,w}[\mathbb{I}\_{Q(\varepsilon)^c}(u,w) p(Y=1|u,w)])^2} \end{align} By definition, $F(\gamma)$ is a positive multiple of $G'(\gamma)$, which implies our conclusion: - If $F(1) < 0$, then $G'(1) < 0$ and thus $E(\gamma)$ decreases around 1. Hence, there exists $\gamma_{-} < 1$, s.t., $E(\gamma_{-})> E(1)$, thus (i) holds; - otherwise, for similar reason, (ii) holds. **Remark**: Here $F(\gamma)$ can be interpreted as some kind of difference between "entropies" in $Q(\varepsilon)^c$ and $Q(\varepsilon)$. When $F(1) < 0$, it means that $Q(\varepsilon)^c$ has higher "entropies", implying that the training trajectories are far from optimal. As a result, we may need to flatten the distribution of training trajectories, which corresponds to using the prior reweighting technique with $\gamma<1$. Since this is the most common case, we usually set $\gamma<1$.

Reviewer n6Yf2024-08-09

I thank the authors for their response. The response was convincing. I am conditionally increasing my score to 8, on the condition the authors actually revise the final version with the keyword planning instead of control.

Authorsrebuttal2024-08-10

Thanks for your feedback and support. We appreciate your willingness to increase the score based on our revisions. We will ensure that the final version of the paper reflects your suggestion by using the keyword "planning" instead of "control."

Reviewer Mp9v2024-08-12

Response

I would like to thank the authors for their rebuttal. When saying the curse of dimensionality, I was referring to the dimension of the state space, which is 2-dimensional, instead of the number of nodes. It would be more convincing to see your method performs well in high-dimensional tasks. I will maintain my score.

Authorsrebuttal2024-08-13

Thanks for your feedback. We appreciate your clarification regarding the curse of dimensionality and acknowledge the importance of higher-dimensional tasks. However, our method is a generic control method for physical system control. It is **agnostic** to specific tasks, model architectures, and data generation methods. Based on our results on the 1D task, the 2D jellyfish control task, and the new 2D smoke control task during rebuttal, the effectiveness of our method is convincingly demonstrated. For 3D tasks, we believe our method is also applicable, provided that efficient solvers, effective model architectures, and sufficient offline data are accessible. Exploration on 3D tasks is left as future work. Thank you again for your valuable comments. We will continue to refine our work based on your suggestions.

Reviewer 8Er52024-08-14

I thank the authors for their detailed response. I initially had the same concern as Mp9v that using only 2D examples might not be sufficiently convincing. If applying to 3D grids, this method requires a 4D network which could be much harder to train. However, I agree with the authors' response that the existing 2D examples are complex enough, and the proposed method could be applicable to 3D tasks. Developing a more efficient network architecture for 4D data might be beyond the scope of this paper. The authors' response regarding the new application area and the prior reweighting technique also alleviated my concerns on novelty. Therefore, I raised my score to 6.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC