Deep-learning-based generative design for optimal reactive silencers

Abstract

A deep-learning-based generative design method is proposed to improve the frequency-dependent characteristics of a reactive silencer, and it has been validated both numerically and experimentally. The noise attenuation performance of the reactive silencer is evaluated with its transmission loss (TL), which varies with frequency and strongly depends on the partition layout inside the reactive silencer. The artificial neural network model for the generative design of the reactive silencer consists of three subnetwork models: the generator, predictor, and converter. The generator model created numerous partition layouts, and their TL curves were estimated using the predictor model. A converter model was developed to identify the frequency-dependent characteristics of the TL curves in a low-dimensional latent space. The latent space was extensively investigated to successfully select the optimal partition layouts satisfying given design requirements, including the target shape of the TL curve and its averaged target TL value. The effectiveness of the proposed method was demonstrated by applying it to three reactive silencer design problems with different design requirements. Among the three optimal silencers, one was physically investigated, and its noise attenuation performance was validated with an acoustic experiment. Because the artificial neural network model of the proposed method was developed for a normalized silencer and requires no prior knowledge of acoustics, it can be easily applied to reduce duct noise in the industry.