New Developments in Topology Optimization Using Computer Graphics and Machine Learning Techniques
Please login to view abstract download link
In this study, we focus on the intricate task of optimizing the design of a self-driving soft body, along with the strategic placement of actuators and their controllers. We utilize a GPU-accelerated Python module Taichi, which has a differentiable programming feature that includes the functionality of DiffTaichi. This optimization challenge involves devising a way for a soft body, embedded with actuators, to gain mobility through ground or wall contact, achieved via dynamic simulations over time. Previously, addressing such sensitive optimization problems was daunting, often necessitating heuristic approaches like genetic algorithms. However, DiffTaichi introduces differentiable simulations of time evolution using the material point method, allowing us to apply gradient methods for more effective optimization. The ability to perform these computations within practical timeframes marks a significant advancement in topology optimization. We aim to explore the future of this field by showcasing several computational examples, highlighting the newfound capabilities in this domain.
