همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Neural-Network Surrogate Modeling for Fast Optimization of Compliant Flapping Mechanisms

نویسندگان :

Hassan Sayyaadi¹ Ali Zouelm²

1- دانشگاه صنعتی شریف 2- دانشگاه صنعتی شریف

کلمات کلیدی :

Surrogate Model،Multilayer Perceptron (MLP)،Flapping Mechanisms،Bio-Inspired Robotics،Data-Driven Modeling،Optimization

چکیده :

Recent research on bioinspired flapping-wing robots has shown that compliant flapping mechanisms play a crucial role in achieving high aerodynamic efficiency. However, optimally designing compliant mechanisms that can reproduce the complex flapping patterns of birds using only a single actuator remains a significant challenge and constitutes a difficult optimization problem. This paper addresses the issues of computational speed and cost in optimizing such mechanisms, and aims to leverage deep learning to design a compliant mechanism capable of autonomously adjusting its flapping pattern under varying flight conditions. A planar two-segment flapping-wing mechanism with two configurations is introduced and modeled: one with two degrees of freedom and a single linear spring, and another with three degrees of freedom and four springs (including compliant hinges). Both configurations are driven by a single rotational motor. Using multibody simulations, a reference MATLAB/SimMechanics model was employed to generate a dataset of roughly 1,300 samples per configuration. A multilayer perceptron (MLP) neural network was then trained as a fast surrogate model to map design parameters to performance indicators for the target of mimicking birds. Subsequently, surrogate-based optimization was performed using a multi-objective genetic algorithm to identify design parameters capable of reproducing target bird-inspired wingbeat patterns for medium birds operating at 3 and 6 Hz. Validation simulations demonstrate that the proposed framework can learn the dynamic behavior of the mechanisms under aerodynamic loading across different frequencies and reproduce frequency-dependent patterns unattainable with purely rigid mechanisms. Moreover, the optimization time is reduced from tens of hours to less than a minute. A comparative analysis of the two configurations further shows that increasing mechanical compliance (through additional and varied springs) improves the fidelity of pattern reproduction.