Noise control has become an important task in an engineering design. As a commonly used component in engineering, reducing the sound radiation from vibrating plates has drawn much attention. An efficient tool for this job is the topology optimization technique. Topology optimization has rapidly developed since it was introduced by [1] and has been applied to a large range of engineering problems. This technique can flexibly generate holes in the structure and achieve the design objectives, such as reducing the weight or increasing the first natural frequency. Du and Olhoff [2] reduced the radiated sound power of the vibrating plate by optimizing the distribution of bi-materials. In addition, Xu et al. [3] also optimized the material distribution of the plate for minimizing the sound radiation. In their work, damping effects are neglected. To achieve a better design, damping patches can be adopted, for example, free-layer damping and constrained-layer damping (CLD). Compared with the free damping layer, CLD provides considerably more damping effects due to the motion constraints of the damping layer. Zheng et al. [4] investigated the topology optimization of passive constrained-layer damping (PCLD) treatment patches on thin plates with respect to sound radiation at low-frequency resonances.

In addition to the optimization procedure, another alternative to reducing the sound power is to locate the most contributing components of the vibrating structure and then adopt some additional patches on these regions, for example, damping layer, stiffener, or adding mass. This approach is more direct and computationally efficient because it usually does not require an iterative optimization process. It, however, sometimes cannot give a very efficient design for radiation control. Numerical techniques to identify the surface contribution to sound power include supersonic intensity and non-negative intensity. Supersonic intensity (SSI) was first proposed by [5] based on near-field acoustic holography (NAH) and Fourier transformation in wavenumber space. Another method is the surface contribution method proposed by [6]. The surface contribution method is a computational procedure to locate the regions of a vibrating object that radiate far-field sound based on acoustic radiation modes. The surface contributions depend on the acoustic radiation modes, the effectiveness of the acoustic radiation, and either the normal structural velocity or the acoustic pressure. The surface contribution is always positive, avoiding cancellation effects and providing a more direct visualization of the surface contribution to sound power than the sound intensity, compared to the supersonic intensity of the region with positive and negative contributions. Later, [7] renamed this quantity non-negative intensity (NNI) for comparison with the sound intensity. Liu et al. [8] compared these two intensities for predicting radiated sound in detail. By using SSI and NNI, the most contributing regions of vibrating structures can be determined. This information can assist in noise control.

The process of meshing can be costly when performing large-scale calculations, which makes the transition from computer-aided design (CAD) to computer-aided engineering (CAE) very cumbersome [9]. In isogeometric analysis (IGA), conventional Lagrangian basis functions are replaced by a B-spline basis function commonly used in CAD, which alleviates the meshing burden [10-19]. As a boundary representation method, IGABEM is naturally compatible with CAD and can thus perform numerical analysis directly on CAD models without having to go through the time-consuming meshing process required by traditional FEM/BEM, and without having to deal with the challenging volume parameterization required by the isogeometric finite element method. As a result, IGABEM is excellent for solving complicated boundary issues (Chen et al. [20]). IGA handles singularities and moving boundaries effectively [21, 22]. In order to solve the weakly singular and hypersingular integrals that emerge in IGABEM, specialized integration methods have been devised [20, 23]. In this work, non-uniform rational B-spline (NURBS) basis functions are used to discretize partial differential equations.

This article focuses on the adoption of NNI for topology optimization of plate structures with the CLD design. Since NNI can be regarded as the contribution to sound power, reducing the contributions defined as NNI over some partial regions of interest could be another design objective, different from the existing optimizations. By optimizing the distributions of NNI, the contributing pattern to sound power varied. That is to say, the radiation pattern of the vibrating structure is optimized.

The remainder of this article is organized as follows: in Section 2, vibro-acoustic analysis, based on the isogeometric finite element method and the Rayleigh integral equation is presented. NNI is established using the acoustic radiation modes. Section 3 presents the complete optimization problem, including the sensitivity analysis, the definition of objective functions, and updating the scheme of design variables. The validation of the proposed optimization method is demonstrated in Section 4 , by means of a baffled plate example. Finally, this work is concluded in Section 5.

The following governing equations for the structure and fluid are derived in the frequency domain. Throughout this contribution, the time-harmonic term e^−iωt will be applied, where i = − 1 is the imaginary unit, ω = 2πf denotes the angular frequency, with f as the excitation frequency in Hz, and t denoting time.

To investigate the validity and applicability of the developed optimization approach, some numerical examples are performed in this section. All the computations are implemented in an in-house Fortran 95/2003 code. Sparse direct solver PARDISO is applied to all computations involving the global dynamic stiffness K _d . Eigenvalue problems described in Eq. 23 are solved using the ARPACK routines. The convergence tolerance τ in the optimization is set to 10^–4.

A baffled plate of dimensions 1 m × 1 m will be chosen as the design object in the following example. It is discretized into 80 × 80 four-node quadrilateral elements, as shown in Figure 3. The finite element model is verified by comparing the natural frequencies and modal loss factors from the present model with those from a closed-form solution [32]. Regarding radiation analysis, the collocation method will be only used when sound intensity is required, otherwise the direct method without integration will be applied instead, in order to save computational costs. We assume the plate to be clamped at its four edges and excited by four harmonic loads F = F ₀e^−iωt as shown in Figure 4. Air is the acoustic medium, with c = 343 m/s and ρ = 1.21 kg/m³. The base and constrained layers consist of aluminum with no damping loss, and the core layer is assumed to consist of a visco-elastic material. Detailed material properties are listed in Table 1.

As mentioned previously, the computational efforts will be reduced when truncating eigensoultions in mapping the radiation modes for evaluating NNI and sound power. Figure 5 shows the Radiated sound power and truncation number versus frequency. In Figure 5A, we compare the sound power computed by sound intensity and NNI resulting from mapping truncated radiation modes, and these two results match very well. We also investigate the truncation number to achieve 99% accuracy compared with the sound power computed by sound intensity in Figure 5B. From the result, we notice that a truncation number of 4 is acceptable for the calculation with a frequency range of 200 Hz. In fact, the sound power is dominantly contributed by the fourth radiation mode because the particle velocity corresponds to the fourth modal pattern. This can be confirmed by investigating the coupling factor of particle velocity with the radiation modes [33]. The coupling factor between the particle velocity pattern and the fourth mode is several orders of magnitude larger than the coupling factor with the first three radiation modes. In the following calculations, we set a higher truncation number of 30 to guarantee correctness and accuracy, and this still saves considerable costs compared with the computation via sound intensity. As a further investigation, we compare the optimization results, which aim at minimizing the radiated computed sound power using truncated acoustic radiation modes. The excitation frequency is assumed to be f = 100 Hz. All the initial design variables are given a uniform value of 1 to avoid possible local minimum. The optimization iteration history is illustrated in Figure 6A. Figures 6B, C show the optimized distributions of the design variables μ. Red (μ _e = 1) represents the visco-elastic damping element, and blue (μ _e = 0) indicates that there is no visco-elastic element. Intermediate values (0 < μ _e < 1) are not physical and are difficult to observe in Figures 6B, C. It can be clearly seen that the two optimizations yield almost the same iteration history and optimization solution. Slight differences between the two optimized designs are possibly caused by numerical error in the eigenvalue analysis and optimization solution. This proves that mapping truncated radiation modes could be an efficient approach to evaluate the sound power without losing accuracy in the response analysis and optimization process at low frequencies. Interestingly, we find that the optimized design yields an even lower sound power (6.6 × 10^–12 W) than that (9.7 × 10^–12 W) of the initial design with a full visco-elastic damping layer, which is similar to the findings in [29]. This indicates that the investigated optimization is a nonlinear problem and thus stresses the necessity of optimization.

Figure 7 shows the distributions of sound intensity and NNI, which are different. The result of NNI is consistent with the corner radiation introduced by [34] and confirmed by [5] using the supersonic intensity, while the sound intensity is not consistent since it could be positive or negative. This is similar to the result presented by [6] where the authors used a (1 × 1) structural modal shape as the particle velocity pattern. It has been clearly shown that NNI (computed via mapped radiation modes) could localize the most contributing areas on structural surfaces, while sound intensity fails due to its cancellation effects. But in the aforementioned optimization, NNI does not outperform sound intensity since they give completely the same objective function when the whole surface is selected, that is, the radiated sound power. NNI, however, makes sense if we only want to decrease the radiation contributed by the partial surface. In this case, integration of the sound acoustic intensity, that is, P _SI, on the prescribed region does not directly correspond to the intensity that radiates to the far-field. This is the reason why we implement the integration of NNI as the objective function instead of P _SI.

In Figure 8, we split the plate surface into several pieces and choose different pieces as the investigated area Γ _c where we want to minimize the sound radiation. When choosing the left parts as Γ _c , the optimization results are obtained and illustrated in Figure 9. The initial values of the design variables are both set to 0.7. From this figure, it can be clearly seen that these two objective functions yield notably different optimized distributions of design variables, as shown in Figures 9A, B. After the optimization, P _SI drops from 4.33 × 10^–6 W to −1.18 × 10^–3 W, while the integration of NNI, that is, P _NNI, increases from 4.33 × 10^–6 W to 1.22 × 10^–3 W. That is to say, the real contributions from the left parts are enlarged by the optimization, which indicates that P _SI cannot correspond to the energy radiated to the far-field. By contrast, the optimization minimizing P _NNI produces a reduction from 4.33 × 10^–6 W to 2.19 × 10^–6 W. The coincidence between P _SI and P _NNI at the beginning is due to the symmetry along the y axis with a uniform distribution of design variables. From Figures 9C, D, we notice that the changing tendency of P _SI is opposite to that of P _NNI, but P _NNI and sound power W show a similar tendency. While the contributions from the left parts are considerably decreased, the sound power almost doubles from 8.66 × 10^–6 W to 1.64 × 10^–5 W.

Figure 10 presents the distributions of sound intensity as well as NNI of the initial design (first row) and two optimized designs (second and third rows). To achieve a better optimization result, we conduct the optimization with different initial values of the design variable: μ = 0.7 and μ = 1 and then show the better one here. This treatment will also be adopted in the following computations. The sound intensity and NNI patterns of the initial design are very similar to those in Figure 7 since they both correspond to a uniform distribution of design variables (μ = 0.7 and μ = 1.0). Hence, homogeneous damping distributions have a very small influence on the patterns of sound intensity and NNI except for detailed values. It is observed the sound intensity and NNI are redistributed by the optimizations in a different manner, which yields inhomogeneous damping distributions. Regarding the sound intensity, there are always adjacent areas exhibiting positive and negative values before and after optimization. From NNI, we notice that two radiating sources located at the left corners become weaker compared with the sources in the right corners after two optimizations. However, compared with the unoptimized source in Figure 10B, the two left sources in Figure 10D correspond to a higher NNI due to improper optimization, with P _SI being the objective function. When using P _NNI, these two corner sources are successfully suppressed. The optimization makes the contributions from the left corners to the sound power reduce from 50% to approximately 13%. This is attributed to improved cancellation effects of the acoustic source and sink in the left region, and thus less acoustic energy radiates to the far-field. In general, the optimization finally results in an inhomogeneous distribution of structural damping and redistributes the acoustic sink and source accordingly to minimize the objective function. Although the sound intensity performs well when minimizing the sound power is of interest, it shows inferiority in radiation control of the partial or local regions.

In Table 2, we present the optimization results for different investigated surface Θ _c , that is, the lower left part. Integrations of the sound intensity and NNI are also selected as the objective function. The first row shows the optimized distributions of design variables μ, resulting from the two objective functions. The second and third rows show the distributions of sound intensity and NNI, and the last row gives the power values before and after optimization. Since the optimization problem is not symmetric along the x and y axes, the resulting designs are also not symmetric anymore. From Table 2, we can see that P _SI decreases from 2.2 × 10^–6 W to −5.1 × 10^–4 W, showing a large reduction. P _NNI, however, increases from 2.2 × 10^–6 W to 1.4 × 10^–3 W. Since NNI provides a more accurate visualization of the contribution to sound power, optimization based on P _SI becomes meaningless because it increases the radiation from the chosen regions. Furthermore, the contributions from the lower left corner are still higher than those from the upper right corner after optimization aiming to decrease P _SI. That is to say, P _SI fails to give the desirable contributing/radiation pattern. When optimizing the damping distribution based on P _NNI, the power radiated from the predefined region reduces to a very low value of 4.3 × 10^–7 W. The lower left corner becomes the least contributing compared with all other corners, which is exactly as expected. In addition, two corners adjacent to the lower left corner also become inconspicuous due to the redistributed damping. After optimization, there is only one hot spot located in the upper right corner which is opposite to the chosen region. The four-corner radiator becomes a one-corner radiator. Comparing the sound power, we find that all optimized solutions yield a high sound power because the optimization target is not minimizing the sound power radiated from the entire surface. From the sound powers, it can be seen that the increase caused by optimization using P _SI can be very large, which is unfavorable. Compared with P _SI, the increase in sound power introduced by optimization using P _NNI is much smaller.

Similar to Tables 2–Tables 4 present optimization results with different investigated surfaces Θ _c . In Table 3, the lower left and upper right parts are selected. It can be clearly seen that two objective functions generate notably different distributions of the design variables with volume fractions of 0.21 and 0.47, respectively. The four-corner radiator becomes a two-opposite-corner radiator after the two optimizations, minimizing F _SI and F _NNI, respectively. Because Γ _c in Figure 8D is symmetric along both the x and y axes, this finally leads to symmetric distributions of damping, as shown in Table 4. We can notice that the distribution of sound intensity and NNI is very similar to that seen in Figure 7, which results from a homogeneous damping distribution. It can be considered that the waves traveling toward high-damped regions (but with opposite directions in one-fourth of the plate area) cancel each other. Hence, the corner mode remains nearly the same as the homogeneous mode. Interestingly, optimization minimizing F _SI produces a lower F _NNI than optimization aiming at minimizing F _NNI. From this aspect, the optimized solution of F _NNI is not optimal, indicating that the applied optimization does not make sense. However, we want to stress that this issue is very common in topology optimization for dynamic and acoustic problems since the optimization problem for these investigated systems are highly non-convex. A local optimum is generally reached by the gradient-based algorithm, and it is very hard to get a global optimum. This is the reason why we conduct optimization with different initial values of the design variables, expecting to obtain a good solution (maybe not optimal). Those optimized solutions, however, are generally good designs, which could provide guidance for design and analysis. Thus, choosing a proper initial value for the design variables is of importance for topology optimization. NNI, which could visualize the contributions of the element/region to sound radiation, has the potential to generate a good initial design. Optimization based on this initial design might converge faster than normal initial values, or converge to a better local minimum. This will be investigated in future work.

From the results discussed previously, it can be clearly seen that the radiation pattern is optimized by using predefined objective functions. Optimized designs usually show higher sound powers than unoptimized designs. However, the increases in sound power caused by optimization using F _NNI are not usually very large and can sometimes be acceptable if the sound power is not of high importance. This is attributed to how we select regions with high contributions to sound radiation as the investigated area Γ _c , such as the corners of the plate. This is reasonable because these regions are always of greater interest compared with other less contributing areas. In contrast, optimization possibly produces a sound power several orders of magnitude larger when choosing F _SI as the objective function. Such enormous increases are usually unfavorable in engineering. Hence, it is not recommended to optimize the radiation pattern based on the integration of sound intensity.

Finally, we have to admit that the physical meanings of NNI are still unclear, and further research on NNI is required. However, it can be clearly seen that optimization using NNI has the potential to optimize the radiating pattern and outperforms the sound intensity in designing the radiation pattern.

An approach to optimizing the contributing pattern to radiated sound power from vibrating structures has been introduced. Two different predefined objective functions, namely, the integrations of NNI and sound intensity over chosen surfaces, are applied and compared in the optimization. By using these objective functions, an optimized damping layer distribution can be found which reduces the contribution to the sound power from the surfaces of interest. When the entire surface is of interest, these two objective functions are equivalent to the radiated sound power, and two optimizations give almost the same result. However, NNI can be extracted from truncated radiation modes, and usually requires less computational effort than sound intensity. This could improve the computational efficiency of optimization. When only partial surfaces are of interest, two predefined objective functions produce quite different optimized designs. However, optimization using the sound intensity possibly yields an enormous increase in the radiated sound power, which is usually unfavorable in engineering designs. Furthermore, it sometimes fails to generate a desirable radiating pattern. In contrast, optimization using NNI always leads to a desirable radiating pattern with only a slightly increased sound power. Thus, we strongly recommend using NNI to optimize the radiating/contributing pattern. The optimization procedure presented here provides a new way for radiating pattern control.