We present a gradient-based optimizer, 2D-GLONets, based on generative neural networks, which can generates two-dimensional (2D) broadband acoustic cloaking devices that exhibit a suppressed total scattering cross section (TSCS) at the range of wavenumbers. We generalized the GLO-Nets [1] model developed originally for one-dimensional photonic metagratings to design 2D broadband acoustic cloaking devices. Such optimized devices can sufficiently reduce the TSCS to the global minimum and render the structure invisible to the incident waves.Our 2D-GLONets model consists of multiple fully-connected layers with LeakyRuLu as the activation function for each layer, except for the output layer, which uses Tanh function. At each iteration, we train our model to generate and continuously move a distribution of devices that widely examine the design space toward a favorable regions by perturbing the positions of each scatterer in planar configuration of cylindrical scatterers across a range of parameters, such as coordinates of each scatterer. The TSCS and its gradients with respect to the scatterer position are computed using in-house built multiple scattering solvers [2]. The gradients are used to maximize the probability of generating optimized devices in the design space. Unlike the gradient-based optimization algorithms and topology optimization which heavily rely on the initial configurations and can easily get trapped in local optima, the 2D-GLONets model reduces the effect of initial distribution by outputting multiple devices to explore the feasible design regions and allocate the computation resources to refine and search for the global optimum instead of local optima.In our inverse design problem, we introduce two geometric constraints: the scatterers do not overlap and stay within a prescribed region. These constraints are implemented by a reparameterization that transforms any infeasible devices into constrained devices. It always guarantees the enforcement of the constraints and allows devices to move toward the optimal regions without adding any penalty term in the loss function.2D-GLOnets was implemented on PyTorch for uniform and non-uniform configurations of rigid cylinders and elastic thin shells. Our results demonstrate that the 2D-GLONets can output devices with a&nbsp;lower TSCS than those yielded by a gradient-based optimization solver fmincon with MultiStarts. For example, 2D-GLOnets outperform fmincon’s state-ofthe-art sequential quadratic algorithms (SQP), SQP legacy, Interior point, and Active set for a uniform configuration of 6 rigid cylinders at 1.0 ≤ ka ≤ 1.2 (ka is a wavenumber). For 2 to 6 cylinder configurations at 0.35 ≤ ka ≤ 0.45, 2D-GLOnets and fmincon yielded the same results. Our proposed model can be also applied in elastodynamics, electromagnets, optics, and nanophotonics to solve various optimization problems with geometric constraints where gradients can be supplied analytically or by using automatic differentiation.References[1] Jiaqi Jiang and Jonathan A. Fan. Simulator-based training of generative neural networks for the inverse design of metasurfaces. Nanophotonics, 9(5):1059–1069, nov 2019.[2] Feruza A. Amirkulova and Andrew N. Norris. The gradient of total multiple scattering cross-section and its application to acoustic cloaking. Journal of Theoretical and Computational Acoustics, page 1950016, mar 2020.

Deep Learning Empowered Design of Acoustic Cloak

Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2021/volume/85543

ASME 2021 International Mechanical Engineering Congress and Exposition (IMECE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.