Inverse Design for Fluid-Structure Interactions using Graph Network Simulators
Authors: Kelsey Allen, Tatiana Lopez-Guevara, Kimberly L. Stachenfeld, Alvaro Sanchez Gonzalez, Peter Battaglia, Jessica B. Hamrick, Tobias Pfaff
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform a large set of experiments and find that (1) the learned simulator model is accurate and flexible enough to support a wide variety of design tasks; (2) it generalizes well enough to support design tasks far outside the training regime; (3) gradients from the learned model turn out to be very useful for inverse design, with gradient-based optimization strongly outperforming a variety of sampling-based optimizers; (4) and surprisingly, gradients remain stable even when rolling the model out for hundreds of steps. In experiments on high-dimensional fluid manipulation (2D FLUID TOOLS and 3D WATERCOURSE) we demonstrate that gradient-based optimization on learned models can find high-quality designs over hundreds of time steps, through states with thousands of particles, in tasks with up to 625 design parameters. The same model matches performance of a specialized solver on aerodynamics design (AIRFOIL) with a much simpler setup, highlighting the flexibility and precision of this approach. |
| Researcher Affiliation | Industry | Kelsey R. Allen Deep Mind, UK Tatiana Lopez Guevara Deep Mind, UK Kimberly Stachenfeld Deep Mind, UK Alvaro Sanchez-Gonzalez Deep Mind, UK Peter Battaglia Deep Mind, UK Jessica Hamrick Deep Mind, UK Tobias Pfaff Deep Mind, UK {krallen,zepolitat,stachenfeld,alvarosg,peterbattaglia,jhamrick,tpfaff}@deepmind.com |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Example code can be found at https://github.com/deepmind/deepmind-research/inverse_ design/. |
| Open Datasets | No | The paper mentions several datasets (2D Water Ramps dataset released in [63], data generated from the simulator in [8], data generated with the Open FOAM solver [55]) but does not provide concrete access information (link, DOI, repository, or specific citation with authors/year for dataset itself) for them being publicly available, beyond citing the papers where the simulators are described. |
| Dataset Splits | No | The paper mentions training, validation, and test datasets in the schema questions (3b), but it does not specify the splits (percentages or counts) within the main text or in Appendix sections mentioned for details. |
| Hardware Specification | Yes | Our approach requires only 21s (single model) to 62s (size-5 ensemble) on a single A100 GPU, compared to 1021s for DAFoam run on an 8-core workstation, despite requiring 10 more optimization steps. ... (up to 4 TPUv3/v100 GPUs, see Appendix B). |
| Software Dependencies | Yes | DAFoam [40], an open-source adjoint framework for multidisciplinary design optimization with openfoam. ... Open FOAM CFD solver [55]. |
| Experiment Setup | No | The paper mentions hyperparameter tuning (e.g., for CEM-M) and refers to Appendix B for model implementation details and Appendix C for domain-specific details, but does not provide specific hyperparameter values or detailed training configurations in the main text. |