Materials & Design 207 (2021) 109855 Contents lists available at ScienceDirect Materials & Design journal homepage: www.elsevier.com/locate/matdes Accelerated topological design of metaporous materials of broadband sound absorption performance by generative adversarial networks Hongjia Zhang a,⇑, Yang Wang a, Honggang Zhao a, Keyu Lu b, Dianlong Yu a, Jihong Wen a,⇑ a Vibration and Acoustics Research Group, Laboratory of Science and Technology on Integrated Logistics Support, College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China b Institute of Unmanned Systems, College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China h i g h l i g h t s g r a p h i c a l a b s t r a c t GANs newly used for topological design of metaporous materials for sound absorption. Huge acceleration (~0.04s/design) for design process enabling instantaneous designing. Successful designs of broadband sound absorption checked by simulation and experiment. Creative configurations and rich local features generated in GANs-designed patterns. AI-guided designing/optimizing as new possibility for AI-materials interdiscipline. a r t i c l e i n f o Article history: Received 10 March 2021 Revised 23 May 2021 Accepted 26 May 2021 Available online 28 May 2021 Keywords: GANs Topological design Metaporous sound absorption materials AI-materials interdiscipline a b s t r a c t The topological design and optimization of metaporous materials is one of the key challenges in the field of sound absorption. Limited by the expensive computational cost, it is particularly disadvantaged when instantaneous multiple designs are required. In recent years, an increasing number of research fields are harnessing machine learning approaches thanks to their experience-free manner and outstanding efficiency. Generative Adversarial Networks (GANs), as a type of machine learning algorithms, enjoy the special benefit of powerful generative capability, making them brilliantly suitable for designing purposes. Additionally, it can fully explore the data distribution space with enormous computational power and create brand new designs. In this work, GANs are newly employed for the topological design of metaporous materials for sound absorption. Trained with numerically prepared data, they successfully propose designs with high-standard broadband absorption performance, verified by simulation and experiment. The designing process is dramatically accelerated by hundreds of times using GANs (100 designs in 4.372 s). This allows GANs to easily provide more structures and configurations, and achieve instantaneous multiple solutions, giving designers more choices to satisfy various constraints such as mass or porosity. In addition, GANs are demonstrated remarkably capable of generating creative configurations and rich local features. This work proposes a new designing principle, illustrates the value of machine learning in guiding the designing and optimizing process in the mechanical world, and opens new possibilities for the future of AI-materials interdisciplinary research. Ó 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). ⇑ Corresponding authors. E-mail addresses: h_zhang00@163.com (H. Zhang), wenjihong@vip.sina.com (J. Wen). https://doi.org/10.1016/j.matdes.2021.109855 0264-1275/Ó 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). H. Zhang, Y. Wang, H. Zhao et al. Materials & Design 207 (2021) 109855 dimensional architectures that approach the Hashin-Shtrikman upper bounds [29]. Conditional GANs were used to design nanophotonic antennae with desired optical properties [30]. In this paper, GANs are employed to carry out topological design of metaporous materials for sound absorption in a significantly accelerated and efficient manner. Trained with numerically prepared data, GANs are demonstrated capable of generating designs with high-standard broadband absorption performance, verified by Finite Element Modeling (FEM) simulation and experiment. Meantime, brand new configurations and local features are seen in the designs by GANs, illustrating its ability of ‘creativity’ and potential to further improve the existing materials for enhanced performance. 1. Introduction Materials for high-efficient airborne-sound absorption have been extensively pursued for decades. Classical porous materials, e.g., foam, glass fiber and mineral wool, are unsuitable for low frequency sound absorption as bulky and heavy materials are required in this case [1]. To overcome the shortcoming of the classic porous materials, using metaporous materials, i.e., a porous matrix embedded with inner scatterers, provides a promising alternative. The last twenty years has seen a growing number of works devoted to the designing of metaporous materials [2–10]. However, limited by intuition, the reported designs feature mostly regular shapes, leaving the potential of the metaporous materials largely unexploited. As an inverse design strategy, topology optimization was capable of producing more complex designs that achieve far more enhanced absorption at multiple target frequencies owing to the possession of different resonance mechanisms in the same design [11]. Nevertheless, the existing topology optimization methods suffer inherent drawbacks in terms of computational time and cost due to extensive variables and numerous iterations needed. Particularly, it becomes even more expensive if a number of designs are required. Moreover, instantaneous designing is hardly possible. To this end, the emerging machine learning algorithms are considered to possess great potential to substantially boost the efficiency of designing and exploring new topologies. The last decade has witnessed the booming of machine learning technology. It has become not only popular in computer science related subjects but also other fields such as mechanics, biology and material science [12–22]. It enjoys the advantage of being an experience-free approach as well as allowing instantaneous multiple designs simultaneously with extremely high efficiency [23]. In addition, it has been favorable for its capability of discovering complex intrinsic pattern of large datasets with the prevailing big data era [13].To the best of our knowledge, machine learning methods have so far not been employed for the design of sound absorption materials. However, we are seeing a rising trend of machinelearning-based designing of two-dimension (2D) architectures [24]. He et al. achieved inverse design of topological metaplates for flexural waves through finding the non-linear mapping between the key parameters and the topology via neural networks [23]. A framework integrating topology optimization and generative models was proposed and validated on the case of 2D wheel design using Boundary Equilibrium Generative Adversarial Networks (GANs) [25]. A convolutional neural network (CNN)-based encoder and decoder network together with conditional GANs was employed as a two-stage refinement to realize near-optimal topological design [26]. Most of the existing works on topological design or optimization adopt basic convolutional neural networks. Those algorithms are usually effective at indirectly designing the structure by deciding its individual parameters (such as the length of an edge or the radius of a circle), but can hardly make a sophisticated design based on complete patterns. As a contrast, our work can directly generate topological patterns using a generative network - GANs. GANs feature a generator and a discriminator that learn simultaneously in competition with each other [27,28]. The generator can generate brand new data (e.g. the topological structure of metaporous materials in this paper) based on the captured potential distribution of the training data. Since the generator is extraordinarily powerful computationally, it can thoroughly explore the data space and produce creative designs that potentially exceed the limitation of the existing ones. Previous works have successfully employed GANs to obtain over 400 complex two- 2. Dataset The training set prepared with FEM simulation consists of 1832 images of a a = 64 64 pixels. Each image represents a pattern for metaporous material. Typical patterns are shown in Fig. 1 where the solid black domain denotes the scatterer and the white area is filled with porous media. The patterns are stored as binary images composed of 0 and 1, corresponding to porous media and scatterer, respectively. The shapes of the scatterer in the training set are composed of three types of basic primitives, i.e., slotted cube, slotted elliptic cylinder and thin partitions. For the slotted cube, the lower (upper) bounds of the outer side lengths and slot width are 0.4a (0.8a) and 0.08a (0.48a), respectively. The minimum (maximum) lengths of outer semi-axis and slot are 0.2a (0.4a) and 0.08a (0.48a) for the slotted elliptic cylinder. As for the thin partition, the lower and the upper bounds of the length are respectively 0.2a and 0.8a. The thickness for all three primitives is no larger than 0.0625a. The generation procedure for all the patterns starts from a porous media (64 64 null matrix). The scatterer is subsequently ‘embedded’ into the porous media at a random location, imitating the mutation process in the genetic algorithm [31], i.e., some pixels of porous media (0) are changed to that of scatterer (1). To increase the diversity of the patterns, a perturbation operator is applied to the pattern where the pixels inside the scatterer are reversed with 20% probability. Finally, in order to improve the geometric topology and eliminate checkerboard pattern, the abuttal filter is employed [32], i.e., some isolated porous elements are changed to scatterer elements and vice versa. Fig. 1. Typical patterns in the training set. 2 Materials & Design 207 (2021) 109855 H. Zhang, Y. Wang, H. Zhao et al. Fig. 2. Schematic diagram of the full-wave simulation model solid line at the right end is the hard acoustic boundary. The porous media in the metaporous material is melamine foam as in our previous work [9]. The acoustic parameters of the melamine foam are described using Johnson-Champoux-Allard (JCA) model [33] in Poroacoustics Domain, as listed in Table 1, where U is the porosity, r the flow resistivity, a1 the tortuosity, K the viscous length and K0 the thermal characteristic length. The sample frequencies in the simulation are from 300 Hz to 3000 Hz at the interval of 50 Hz. Since GANs generate new patterns based on what they are feed with, we prepared 1832 patterns whose average absorption coefficients are larger than 0.9 to form the training set, aiming to obtain patterns with at least equally high absorption performance. The patterns in the training data contain up to three primitives. Fig. 1(a)-(d) plots those with only one primitive. The employment of perturbation operator during the generation process leads to the non-smooth peripheries of the slotted scatterer as well as the discrete dotted black pixels inside. It is also the reason for the existence of the closed cylinder shown in Fig. 1(c). The patterns consisting two or three primitives are shown in Fig. 1(e)-(l). For instance, Fig. 1(e) and Fig. 1(g) show the patterns with two primitives of the same type. The examples for multiple primitives with different types are given in Fig. 1(k) and (l). Particularly, we shall pay attention to the correlation between primitives in the same pattern. They can be simply apart as in Fig. 1(g), half intersected in Fig. 1(f) or internally tangent in Fig. 1(h). To obtain the absorption coefficients of each pattern, full-wave simulation is performed via Pressure Acoustic and Frequency domain module with the commercial finite element software COMSOLÒ. As shown in Fig. 2, a plane wave is incident to the pattern with the amplitude of 1 Pa from the background pressure field. Perfectly matched layer (PML) is applied at the left boundary of the background pressure field to simulate anechoic termination for outgoing wave (i.e., reflective wave). The horizontal black dashed lines denote the periodic boundary while the green black 3. Methods 3.1. Generative adversarial networks GANs feature a pair of networks, a generator G and a discriminator D, competing with each other. As a two-player zero-sum game, G aims to generator ‘fake’ data that are as ‘real’ as possible to ‘fool’ G, while G learns to discriminate fake data generated by G from real ones as accurately as possible. This is essentially a minmax process where G and D can both respectively sharpen their generative and discriminative skills [34]. Importantly, G does not directly access real images and it learns only through interacting with D [28]. Consequently, G learns to generate usable data (patterns in our case) for the purpose of the specific model setup. Eventually, a Nash equilibrium is reached, meaning each player cannot obtain a better result without changing the other players’ strategy, and two players stay equally good [35]. Table 1 Acoustical parameters of the melamine foam for FEM simulation. U a1 K K0 r 0.95 1.42 180 lm 360 lm 8900N s m4 Fig. 3. Schematic illustration of the design procedures of metaporous materials with GANs. 3 H. Zhang, Y. Wang, H. Zhao et al. Materials & Design 207 (2021) 109855 accelerated tensor computation and performance at the same time. With packages from Pytorch, procedures in deep learning such as network building and gradient descent can be programed with a couple of code lines, making it far more efficient. In general, the architectures of G and D are described as follows. The input of G is a greyscale image with 64 64 size. G contains ten convolutional layers. Each convolutional layer is followed by a batch normalization layer and a ReLU layer. The output of G is still a grey image with the same size as the input one. D takes the grey image generated by G as the input. D consists of five convolutional layers and a softmax layer. Each convolutional layer is also followed by a ReLU layer. The output of D is a binary classification label which indicates if the generated image is legitimate or not. 3.2. Design procedures Fig. 3 demonstrates the procedure to design and verify the metaporous materials. Three main steps are necessary for a complete and reliable design: 1) training data preparation; 2) generating designs using trained GANs; 3) verify those designs with FEM simulation and experiments. The left part of Fig. 3 shows the design procedures of metaporous materials with GANs in this work. The model is built with Pytorch. Pytorch is a commonly used open-source deep learning framework, basically a library that provides building blocks for designing, training and validating deep neural networks via a high-level programing interface [36,37]. Being a python library, Pytorch enjoys the advantages of possessing both GPU- Fig. 4. Designs proposed by GANs and their absorption performance simulated by FEM: (a) I: designs 1–4; (b) II: designs 5–8; (c) Absorption coefficients of Design 1–4; (d) Absorption coefficients of Design 5–8. 4 Materials & Design 207 (2021) 109855 H. Zhang, Y. Wang, H. Zhao et al. Specifically, with a random noise (64 64 size) as input, G outputs a ‘fake’ pattern (64 64 size) which is subsequently input into D together with the ‘real’ patterns – the training data. D judges whether the generated pattern is legitimate by comparing it with training data. With the advancement of the minmax process, G finally learns to generate designs that are comparable with the training data. Delightfully, patterns generated this way possess new features since G is computationally extremely powerful to explore the full distribution of the data space (i.e., full possibilities). Note that the criterion of D’s judgement on whether a pattern is legitimate or not is actually learned via training, instead of adopting a handcrafted criterion. The criterion is essentially represented by the weights in the convolutional kernels that are constantly applied to the input images to extract features at each convolutional layer in the neural networks. Starting with a random D with a cross-entropy loss that calculates the difference between the output pattern of the G with the input pattern, D’s aim is to maximize the loss while G’s aim is to minimize it, which forms a minmax ‘competition’. D’s capability of judging whether the generated pattern is legitimate is basically equal to its capability to maximize the loss, which relies on the choice of weights. The same logic goes to G’s capability of generating legitimate patterns. During training, the ‘competition’ advances, where D and G continuously adjusts their own weights based on the loss until a balance is reached (i.e. the Nash equilibrium). Eventually, a capable D is obtained with a set of learned weights to represent the criterion that is good at ‘judging’. 4. Results 4.1. GANs-designed patterns and numerical verification GANs bring extreme acceleration of the designing process, enhancing the efficiency by hundreds of times compared with Table 2 Average absorption coefficient for designs. Design 1 Design2 Design3 Design4 Design5 Design6 Design7 Design8 0.925 0.929 0.930 0.926 0.925 0.931 0.927 0.927 Fig. 5. Pressure distribution inside Design 3 at the characteristic frequencies of: (a) f1 = 950 Hz; (b) f2 = 1300 Hz; (c) f3 = 1960 Hz; (d) f4 = 2600 Hz. 5 H. Zhang, Y. Wang, H. Zhao et al. Materials & Design 207 (2021) 109855 It is illustrated that GANs are skilled at generating and alternating local features. Being greatly capable at manipulating input information, it can comprehensively assemble different shapes/ primitives in creative configurations. For instance, the model has learned from Fig. 1(h) that one primitive can be inside another one. Meanwhile, it has captured several different primitives out of the training data including slotted cylinders and thin rigid partitions. Design 5 in Fig. 4(b) is a result of putting thin partitions inside a slotted cylinder as well as making some modification and variation on the cylinder’s periphery. Hence, we reckon that GANs are capable of capturing different shapes in the training data, treating them as separate features and generating brand new combinations and create rich variation on the features themselves. Particularly noteworthy is the position of the opening on the cavity. The training data only contain patterns with opening positioned towards the incident end of the sound wave, i.e., the top of the pattern. However, GANs have created new opening positions towards the backing (the bottom of the pattern) as shown in Design 3 and 4 in Fig. 4(a). This operation is of great importance for sound absorption and is often highlighted in the existing nonmachine-learning-based designing methods [39]. the existing approaches such as density-based approach and genetic algorithm [38]. The generation of 100 patterns takes approximately only 4.372 s (with commonly used Nvidia RTX3090) once the model is well trained. More details regarding the time needed for data preparation and model training are described in Appendix A. More than a hundred designs of metaporous materials are proposed by GANs out of which eight typical ones are presented in Fig. 4. The height and the width of a pattern are both 64 mm (and 64 pixels). So, one pixel inside the pattern has the size of 1 mm 1 mm. Overall, all the designs can be classified into two categories: I - multiple slotted cavities with different orientations and creative variation on the periphery (Design 1–4 in Fig. 4(a)); II - a single slotted cavity with large opening and rich local features inside (Design 5–8 in Fig. 4(b)). Note that ‘annulus’ here is a general term describing the circle-like or rectangle-like outlines in the pattern. Additionally, to differentiate GAN-generated patterns from the training data, patterns below are highlighted in blue (scatterer) and pink (foam). The FEM-calculated absorption coefficient curves of the eight designs are illustrated in Fig. 4(c) and (d). Overall, all the eight designs show broadband absorption capability. The average absorption coefficients (between 300 Hz and 3000 Hz with the interval of 50 Hz) are listed in Table 2. It is shown that the average absorption coefficients of all designs are larger than 0.925, which is a rather high standard. 4.2. Broadband absorption mechanism To explore the mechanism behind the broadband absorption performance, we choose Design 3 and Design 8 as representatives Fig. 6. Pressure distribution inside Design 8 at the characteristic frequency of (a) f1 = 930 Hz; (b) f2 = 1280 Hz; (c) f3 = 1820 Hz; (d) f4 = 2400 Hz. 6 Materials & Design 207 (2021) 109855 H. Zhang, Y. Wang, H. Zhao et al. Fig. 7. Sample preparation process. Design 3 shows a different coupled mode at the frequency of f4. Notable acoustic scattering is observed between the two slotted cylinders near the incident end and partial acoustic energy also enters the left slotted cylinder. It could be concluded that the collaboration of resonance, acoustic trapping and scattering effect contributes to the overall broadband absorption of Design 3. Fig. 6(a)-(d) further illustrates the pressure distributions of Design 8 at four representative frequencies. At low frequency f1, of the two categories of designs for deeper examination. Fig. 5(a)Fig. 6(d) plot the pressure distribution of Design 3 at four characteristic frequencies. It is noted that the two slotted cylinders are sole ‘active’ at the frequencies of f1 and f3, i.e., maximum pressure is located at the two slotted cylinders and resonance state exists near the two frequencies. At the frequency of f2, the pressure is trapped between the two slotted cylinders and the backing, while both two slotted cylinders are simultaneously ‘active’. However, Fig. 8. Experimental setup: (a) schematic illustration; (b) photo. Fig. 9. Experimental results compared with FEM simulation for (a) Design 3 and (b) Design 8. 7 H. Zhang, Y. Wang, H. Zhao et al. Materials & Design 207 (2021) 109855 approach and alternating primitives (as well as adding local features) for the assembly to form a new design. The other is that GANs have created new opening positions for the design, which is of crucial significance in practice. Overall, this work has for the first time utilized GANs to design metaporous materials for sound absorption. The accelerated designing process and the satisfying results demonstrate that AImaterials interdisciplinary approach is of excellent potential and importance in guiding the designing and optimizing process in the mechanical world. the main acoustic energy enters the full big cavity while partially trapped between the cavity and the backing. At frequencies f2 and f4, it is interesting to note that two different local resonance modes are triggered by the thin partitions inside the cavity. The two local resonance modes together with the scattering outside the cavity also contribute to the sound absorption of Design 8 at frequency of f3. Different form Design 3, Design 8 possesses more complex local modes inside the cavity, thanks to the rich local features of thin partition generated by GANs. 4.3. Experimental verification Declaration of Competing Interest Fig. 7 illustrates the process of sample preparation. Overall, the metaporous material is composed of three components, i.e., a cover layer, an inner scatterer and melamine foam. The inner scatterer and the cover layer are 3D printed with polylactic acid (PLA) with the parameters of density q = 1250 kg/m3, Young’s modulus E = 1.2GPa and Poisson’s ratio v = 0.35. The thickness of the cover layer is 1 mm. Compared with the size of the cubic sample 60 mm, the influence of the cover layer is neglectable. To facilitate the embedding of the scatterer, a pattern with the shape of the scatterer is etched into the melamine foam. The metaporous material is an assembly of all the components. Note that the sound wave is incident into the top face along the -y direction shown in Fig. 7. The experiment is carried out in a home-made impedance tube using the two-microphone method [40]. As shown in Fig. 8(a), the impedance tube features a square cross-section with the side length of d = 60 mm. The sample is attached to a rigid backing. The distances between the sample and the two microphones are L1 = 200 mm and L2 = 150 mm, respectively. Design 3 and Design 8 are chosen for experimental verification. Fig. 9(a) and (b) compare results obtained by the experiment and FEM simulation. Overall, a good agreement is reached. However, some discrepancies still exist, especially for Design 8. We consider that the mismatch is a result of two reasons: (1) the uncertainty of fabrication for both 3D-printed frame and the etched melamine foam; (2) the inaccuracies occurred during the assembling process. The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgement This work was supported by the National Natural Science Foundation of China (Grant No. 52001325, 51775549, 11991032 and 11991030). Appendix A. Computation time The 1832 patterns with >0.9 average absorption coefficients are selected from a larger database with 7636 patterns in total via examining all their average absorption coefficients. That means we use COMSOL (full-wave simulation) to calculate the average absorption coefficient of all the 7636 patterns. Each calculation takes about 3 s, making it 22,908 s (that’s roughly 382 min or 6.36 h) in total. Note that the 7636 patterns are generated following the generation procedure described in ‘section 2 Dataset’, which is rather fast. The generation of 7636 patterns takes around 234.7 s (roughly 3.9 min) in total. So, the preparation of the training dataset of 1832 patterns takes approximately 386 min (~6.43 h), including 2 parts: the generation of 7636 patterns (3.9 min) and the COMSOL FEM calculation of 7636 patterns (382 min). Model training time is related to the GPU computation power. With NVIDIA RTX3090, the whole training takes about 8 h. The training process can easily speed up via training in parallel on multiple GPUs. In total, the pre-designing stage (including training set preparation and model training) need <15 h. On the other hand, based on our experience of the existing topology optimization methods, one generation averagely takes about 3 s and it needs around 200 generations for the algorithm to converge to give out a design, i.e. 600 s for one pattern. Therefore, to design 100 well-performed patterns with genetic algorithms takes about 1000 min (~16.7 h), which is already longer than the GANs-based approach. The above comparison is only for designing 100 patterns. When more patterns are needed, the advantage of GANs-based approach becomes a lot more obvious as it only needs multiple seconds (for generation) while the existing methods need hours. Therefore, the GANs-based approach has dramatically accelerated the designing procedure especially when instantaneous multiple solutions are needed. 5. Conclusions In this work, GANs are used to design metaporous materials for sound absorption. The training set is prepared with FEM simulation, containing 1832 binary images of 64 64 pixels. With Pytorch framework, trained GANs can generate 100 patterns in roughly 4.372 s, bringing tremendous acceleration of the designing process, boosting the efficiency by hundreds of times compared with the existing approaches. Eight typical designs proposed by GANs are chosen to be verified by FEM simulation, out of which two are printed out with a 3D printer for experimental evaluation. The advantage in the computation time enables this approach to provide more structures and configurations, and achieve instantaneous multiple solutions with high absorption performance, compared to a unique solution as the case with the existing methods. The results illustrate that GANs are capable of generating metaporous material designs with satisfying broadband absorption performance. The FEM simulation of the eight designs demonstrated that their average absorption coefficients are all larger than 0.925. The experiments again confirm the quality of the two 3Dprinted designs. Additionally, the numerical and experimental results show satisfying agreement. As to the generative capability, GANs are demonstrated skilled at capturing input information and generating brand new configurations and rich local features comprehensively. Two typical creations of GANs’ are observed. One is to learn a new assembling Data availability The raw/processed data required to reproduce these findings cannot be shared at this time as the data also forms part of an ongoing study. 8 Materials & Design 207 (2021) 109855 H. Zhang, Y. Wang, H. Zhao et al. [20] J. Carrasquilla, R.G. Melko, Machine learning phases of matter, Nat. Phys. 13 (5) (2017) 431–434. [21] W. Ma, Z. Liu, Z.A. Kudyshev, A. Boltasseva, W. Cai, Y. Liu, Deep learning for the design of photonic structures, Nat. Photonics 15 (2021) 77–90. [22] K.T. Butler, D.W. Davies, H. Cartwright, O. Isayev, A. Walsh, Machine learning for molecular and materials science, Nature 559 (7715) (2018) 547–555. [23] L. He, Z. Wen, Y. Jin, D. Torrent, X. Zhuang, T. Rabczuk, Inverse design of topological metaplates for flexural waves with machine learning, Mater. Des. 199 (2021) 109390. [24] H.T. Kollmann, D.W. Abueidda, S. Koric, E. Guleryuz, N.A. Sobh, Deep learning for topology optimization of 2D metamaterials, Mater. Des. 196 (2020) 109098. [25] S. Oh, Y. Jung, S. Kim, I. Lee, N. Kang, Deep generative design: Integration of topology optimization and generative models, J. Mech. Des. 141 (11) (2019). [26] Y. Yu, T. Hur, J. Jung, I.G. Jang, Deep learning for determining a near-optimal topological design without any iteration, Struct. Multidiscip. Optim. 59 (3) (2019) 787–799. [27] K. Wang, C. Gou, Y. Duan, Y. Lin, X. Zheng, F.-Y. Wang, Generative adversarial networks: introduction and outlook, IEEE/CAA J. Autom. Sin. 4 (4) (2017) 588– 598. [28] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, A.A. Bharath, Generative adversarial networks: an overview, IEEE Signal Process Mag. 35 (1) (2018) 53–65. [29] Y. Mao, Q. He, X. Zhao, Designing complex architectured materials with generative adversarial networks, Sci. Adv. 6 (17) (2020) eaaz4169. [30] S. So, J. Rho, Designing nanophotonic structures using conditional deep convolutional generative adversarial networks, Nanophotonics 8 (7) (2019) 1255–1261. [31] J.H. Holland, Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence, MIT Press, 1992. [32] H.-W. Dong, X.-X. Su, Y.-S. Wang, Multi-objective optimization of twodimensional porous phononic crystals, J. Phys. D Appl. Phys. 47 (15) (2014) 155301. [33] J. Allard, N. Atalla, Propagation of Sound in Porous Media: Modelling Sound Absorbing Materials 2e, John Wiley & Sons, 2009. [34] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative adversarial nets, Adv. Neural Inf. Process. Syst. (2014) 2672–2680. [35] W. Fedus, M. Rosca, B. Lakshminarayanan, A.M. Dai, S. Mohamed, I. Goodfellow, Many paths to equilibrium: GANs do not need to decrease a divergence at every step, arXiv preprint arXiv:1710.08446 (2017). [36] NVIDIA, Deep Learning Frameworks. https://developer.nvidia.com/deeplearning-frameworks. [37] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, Pytorch: An imperative style, high-performance deep learning library, arXiv preprint arXiv:1912.01703 (2019). [38] M.P. Bendsoe, O. Sigmund, Topology Optimization: Theory, Methods, and Applications, Springer Science & Business Media, 2013. [39] C. Lagarrigue, J.-P. Groby, V. Tournat, O. Dazel, O. Umnova, Absorption of sound by porous layers with embedded periodic arrays of resonant inclusions, J. Acoust. Soc. Am. 134 (6) (2013) 4670–4680. [40] ISO, 10534-2: Acoustics—Determination of sound absorption coefficient and impedance in impedance tubes—Part 2: Transfer function method first edition, 1998. References [1] M. Yang, P. Sheng, Sound absorption structures: From porous media to acoustic metamaterials, Annu. Rev. Mater. Res. 47 (2017) 83–114. [2] C. Boutin, F.X. Becot, Theory and experiments on poro-acoustics with inner resonators, Wave Motion 54 (2015) 76–99. [3] J. Yang, J.S. Lee, Y.Y. Kim, Metaporous layer to overcome the thickness constraint for broadband sound absorption, J. Appl. Phys. 117 (17) (2015) 174903. [4] C. Lagarrigue, J.-P. Groby, O. Dazel, V. Tournat, Design of metaporous supercells by genetic algorithm for absorption optimization on a wide frequency band, Appl. Acoust. 102 (2016) 49–54. [5] Y. Fang, X. Zhang, J. Zhou, Acoustic porous metasurface for excellent sound absorption based on wave manipulation, J. Sound Vib. 434 (2018) 273–283. [6] X. Liu, C. Yu, F. Xin, Gradually perforated porous materials backed with Helmholtz resonant cavity for broadband low-frequency sound absorption, Compos. Struct. 113647 (2021). [7] F. Xin, X. Ma, X. Liu, C. Zhang, A multiscale theoretical approach for the sound absorption of slit-perforated double porosity materials, Compos. Struct. 223 (2019) 110919. [8] N. Gao, J. Wu, K. Lu, H. Zhong, Hybrid composite meta-porous structure for improving and broadening sound absorption, Mech. Syst. Sig. Process. 154 (2021) 107504. [9] H. Zhao, Y. Wang, D. Yu, H. Yang, J. Zhong, F. Wu, J. Wen, A double porosity material for low frequency sound absorption, Compos. Struct. 239 (2020) 111978. [10] X.-F. Zhu, S.-K. Lau, Z. Lu, W. Jeon, Broadband low-frequency sound absorption by periodic metamaterial resonators embedded in a porous layer, J. Sound Vib. 461 (2019) 114922. [11] W.U. Yoon, J.H. Park, J.S. Lee, Y.Y. Kim, Topology optimization design for total sound absorption in porous media, Comput. Methods Appl. Mech. Eng. 360 (2020) 112723. [12] Z. Pei, J. Yin, Machine learning as a contributor to physics: understanding Mg alloys, Mater. Des. 172 (2019) 107759. [13] Y.-G. Ham, J.-H. Kim, J.-J. Luo, Deep learning for multi-year ENSO forecasts, Nature 573 (7775) (2019) 568–572. [14] Q. Zhao, H. Yang, J. Liu, H. Zhou, H. Wang, W. Yang, Machine learning-assisted discovery of strong and conductive Cu alloys: data mining from discarded experiments and physical features, Mater. Des. 197 (2021) 109248. [15] M. Zhong, K. Tran, Y. Min, C. Wang, Z. Wang, C.-T. Dinh, P. De Luna, Z. Yu, A.S. Rasouli, P. Brodersen, S. Sun, O. Voznyy, C.-S. Tan, M. Askerka, F. Che, M. Liu, A. Seifitokaldani, Y. Pang, S.-C. Lo, A. Ip, Z. Ulissi, E.H. Sargent, Accelerated discovery of CO2 electrocatalysts using active machine learning, Nature 581 (7807) (2020) 178–183. [16] Z. Wei, J. Yu, Y. Lu, J. Han, C. Wang, X. Liu, Prediction of diffusion coefficients in fcc, bcc and hcp phases remained stable or metastable by the machinelearning methods, Mater. Des. 198 (2021) 109287. [17] Q. Xie, M. Suvarna, J. Li, X. Zhu, J. Cai, X. Wang, Online prediction of mechanical properties of hot rolled steel plate using machine learning, Mater. Des. 197 (2021) 109201. [18] Y. Zhang, A. Mesaros, K. Fujita, S. Edkins, M. Hamidian, K. Ch’ng, H. Eisaki, S. Uchida, J.S. Davis, E. Khatami, Machine learning in electronic-quantum-matter imaging experiments, Nature 570 (7762) (2019) 484–490. [19] J. Yu, C. Wang, Y. Chen, C. Wang, X. Liu, Accelerated design of L12-strengthened Co-base superalloys based on machine learning of experimental data, Mater. Des. 195 (2020) 108996. 9