Core Concepts
CinDM introduces a novel approach to compositional generative inverse design, enabling complex system optimization beyond training data.
Abstract
ABSTRACT:
Inverse design is crucial across engineering fields.
Optimizing over learned energy functions improves design performance.
Compositional design enables complex system creation.
INTRODUCTION:
Inverse design challenges involve simulating complex dynamics.
Deep learning shows promise in addressing inverse design problems.
DATA EXTRACTION:
"Project website and code can be found at https://github.com/AI4Science-WestlakeU/cindm."
RELATED WORK:
Classical methods rely on slow solvers; deep learning offers promising alternatives.
CinDM introduces a compositional generative perspective to inverse design.
METHOD:
CinDM optimizes both the design objective and a generative objective simultaneously.
Diffusion models are used for denoising and optimization.
EXPERIMENTS:
Compositional Generalization Across Time:
CinDM outperforms baselines in MAE and design objectives for longer time steps.
Compositional Generalization Across Objects:
CinDM excels in MAE and design objectives when generalizing to more objects than in training.
Generalization Across Airfoils:
CinDM achieves higher lift-to-drag ratios compared to baselines, showcasing effective boundary designs.
Stats
"Project website and code can be found at https://github.com/AI4Science-WestlakeU/cindm."