Compositional Generative Inverse Design
Authors: Tailin Wu, Takashi Maruyama, Long Wei, Tao Zhang, Yilun Du, Gianluca Iaccarino, Jure Leskovec
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiments, we aim to answer the following questions: (1) Can Cin DM generalize to more complex designs in the test time using its composition capability? (2) Comparing backpropagation with surrogate models and other strong baselines, can Cin DM improve on the design objective or prediction accuracy? (3) Can Cin DM address high-dimensional design space? To answer these questions, we perform our experiments in three different scenarios: compositional inverse design in time dimension (Sec. 4.1), compositional inverse design generalizing to more objects (Sec. 4.2), and 2D compositional design for multiple airfoils with Navier-Stokes flow (Sec. 4.3). |
| Researcher Affiliation | Collaboration | Tailin Wu1 , Takashi Maruyama2 , Long Wei1 , Tao Zhang1 , Yilun Du3 , Gianluca Iaccarino4, Jure Leskovec5 1Dept. of Engineering, Westlake University, 2NEC Laboratories Europe, 3Dept. of Computer Science, MIT, 4Dept. of Mechanical Engineering, Stanford University, 5Dept. of Computer Science, Stanford University |
| Pseudocode | Yes | Algorithm 1 Algorithm for Compositional Inverse Design with Diffusion Models (Cin DM) |
| Open Source Code | Yes | Project website and code can be found at https://github.com/AI4Science-Westlake U/cindm. |
| Open Datasets | No | We use two Python packages Pymunk (Blomqvist, 2007) and Pygame (Shinners, 2000) to generate the trajectories for this N-body dataset. [...] We use Lily-Pad (Weymouth, 2015) as our data generator (Fig. 5). We generate 30,000 ellipse bodies and NACA airfoil boundary bodies and perform fluid simulations around each body. The paper does not explicitly state that these *generated* datasets are publicly available. |
| Dataset Splits | No | For the 2D airfoil dataset, it states 'We generate 10,000 trajectories for the training dataset and 1,000 trajectories for the test dataset.' For the N-body dataset, it mentions total simulations and data input shape, but does not specify explicit train/validation/test splits. No explicit validation split information is provided for either dataset. |
| Hardware Specification | Yes | Our model is trained for approximately 60 hours on a single Tesla V100 GPU |
| Software Dependencies | No | The paper mentions 'Python packages Pymunk' and 'Pygame' but does not provide specific version numbers for these or other software dependencies like PyTorch, Python, or Adam optimizer. |
| Experiment Setup | Yes | Our model is trained for approximately 60 hours on a single Tesla V100 GPU, with a batch size of 32, employing the Adam optimizer for 1 million iterations. For the first 600,000 steps, the learning rate is set to 1e-4. After that, the learning rate is decayed by 0.5 every 40,000 steps for the remaining 400,000 iterations. |