Machine learning-based sound power topology optimization for shell structures

Abstract

Design optimization for acoustics is a challenging problem for aerospace engineers. Broadband radiated sound power is a useful performance measure in aircraft design, but is computationally expensive with existing sensitivity analysis methods. Machine learning is a promising approach for learning and exploiting complex behaviour in acoustic response data. This article proposes using a reinforcement learning framework to generate designs with minimal sound power. First, a residual neural network is trained to estimate the sound power response of a given design. Then, the residual neural network is used to train a convolutional neural network to perform topology optimization. The methodology was applied in the design of unstiffened and stiffened panels. The reinforcement learning agent successfully generated designs with lower sound power than all designs in the dataset used to train the residual neural network.