Compositional Generative Inverse Design
Tailin Wu, Takashi Maruyama, Long Wei, Tao Zhang, Yilun Du, Gianluca, Iaccarino, Jure Leskovec
TL;DR
This paper introduces a novel inverse design approach using diffusion models that avoids adversarial modes, enabling the creation of more complex and effective designs by compositional model integration.
Contribution
It proposes a diffusion-based inverse design method that improves over traditional optimization, allowing compositional design of complex systems with enhanced performance.
Findings
Improved design performance by avoiding adversarial modes.
Ability to compose multiple diffusion models for complex system design.
Successful application to N-body and multi-airfoil design tasks.
Abstract
Inverse design, where we seek to design input variables in order to optimize an underlying objective function, is an important problem that arises across fields such as mechanical engineering to aerospace engineering. Inverse design is typically formulated as an optimization problem, with recent works leveraging optimization across learned dynamics models. However, as models are optimized they tend to fall into adversarial modes, preventing effective sampling. We illustrate that by instead optimizing over the learned energy function captured by the diffusion model, we can avoid such adversarial examples and significantly improve design performance. We further illustrate how such a design system is compositional, enabling us to combine multiple different diffusion models representing subcomponents of our desired system to design systems with every specified component. In an N-body…
Peer Reviews
Decision·ICLR 2024 spotlight
The authors approach is very interesting. The paper is straightforward and aims at directly addressing the problem it uses. It is clear and fairly well-written. The experiments provided by the authors seem to confirm the validity of the proposed method.
I personally found the experiments slightly harder to read compared to the rest of the paper. For other remarks see questions.
1. The generative optimization structure containing both the energy-based model and the design objective is quite unique and novel. It enables the optimization problem for design to be more readily approached via the joint learning procedure. 2. The experiments conducted in Section 4 are complete which explains well the questions raised at the beginning of the section. Overall, the ability shown in the work to generalize is quite impressive and seems promising with potential to be applied to mor
1. This is more of a question. On the joint optimization, it is trying to minimize the energy component which is calculated from the trajectories and the boundary, and minimizing the design objective as well. It is proposed to achieve this by optimizing the design and the trajectory at the same time. In the joint optimization formulation as in Eqn.(3), the design objective function is weighted by $\lambda$. I am curious how this hyperparameter is estimated/configured, and how sensitive the optim
Originality: - the paper adopts or re-invents various tricks I've seen across the literature (unrolling across time steps and jointly diffusing, using a diffusion model as a smoothed ODE effectively) but does so in a clever combination - novelty: I'm not aware of any similar work, although conditional policydiffusion or codefusion might come close, and adding noise to FNO etc. is standard practice - clarity: overall clear presentation, especially on hyperparameters (kudos!), some questions (see
- maybe I missed it, but page 7, I don't think $M$ is ever defined. How exactly do you train $M$ beyond the range of timesteps in training? - I would question the compositionality of the method and call it a "piecewise" or "mixture" approach? Given that you simply partition the spaces required into overlapping pieces (unless I misunderstood something) - Were the numbers of parameters matched for the different baselines? Given that you a partitioned energy functions, there might be potential for
Code & Models
Videos
Taxonomy
TopicsModel Reduction and Neural Networks · Gaussian Processes and Bayesian Inference · Adversarial Robustness in Machine Learning
MethodsDiffusion · Masked autoencoder