Forward Prediction And Reverse Design Of Periodic Wave Barrier Bandgap Based On Deep Learning

Setting periodic wave barrier on the propagation path of environmental vibration is an effective way to reduce vibration.There is no need to change the protected structure,and targeted design can be carried out according to the frequency range of vibration.However,the conventional forward solution method is difficult to carry out reverse design,and it needs many trials and errors.Although genetic algorithm and other optimization methods can design the structure that meets the requirements,they often need a lot of calculation cost,and they need to be redesigned when the objectives or conditions change.The idea of deep learning is to use a data-driven way and a large amount of data training to let the neural network learn the laws in the data,which can effectively complete the forward prediction and reverse design of the periodic wave barrier.Only one data set needs to be established,the bandgap range of the wave barrier can be given at any time,and the targeted reverse design can be carried out according to the conditions and objectives,so as to make the periodic wave barrier more practical and feasible.In this paper,a two-dimensional periodic body wave barrier is designed for the main frequency range of environmental vibration of 10～100Hz.Considering the configuration of ternary materials,the effects of the elastic modulus,density,Poisson’s ratio of soil,rubber and the core as well as the geometric parameters of six different configurations on the band gap are studied.The arrangement of hexagonal lattice and square lattice are compared,and the final unit configuration and eight design parameters are determined.With the forward prediction network established,the bandgap forward prediction is carried out by randomly generating the data sets of in-plane wave and out of plane wave,which is compared with the random forest method.The reverse design network is established and the design of periodic wave barrier is realized.On this basis,the multicondition design and one-to-many design are discussed.The conclusions of the study are as follows:(1)Using neural network to predict the band gap of periodic structure has high accuracy.The regression coefficients of in-plane and out of plane prediction can reach more than 99.7%,and the error is no more than 5%.(2)Compared with random forest,although neural network needs more data sets and higher calculation cost,the prediction accuracy of neural network is higher.(3)The reverse design network can complete the reverse design of structural parameters.The coincidence degree of the designed structural parameters and bandgap with the data set is more than 99.5%,and the design can be completed in a few seconds.(4)The reverse design network can complete the multi condition design of limiting the periodic constant or specifying the rubber parameters or limiting the parameter range.Compared with the single condition,the multi-condition design can find the structure that meets the actual needs.(5)The reverse design network can characterize the non uniqueness of the reverse problem,realizing one-to-many design,and the different results of the design can meet the target requirements.