Improvement of the acoustic black hole e ect by using energy transfer due to geometric nonlinearity

Acoustic Black Hole e ect (ABH) is a passive vibration damping technique without added mass based on exural waves properties in thin structures with variable thickness. A common implementation is a plate edge where the thickness is locally reduced with a power law pro le and covered with a viscoelastic layer. The plate displacement in the small thickness region is large and easily exceeds the plate thickness. This is the origin of geometric nonlinearity which can generate couplings between linear eigenmodes of the structure and induce energy transfer between low and high frequency regimes. This phenomenon may be used to increase the e ciency of the ABH treatment in the low frequency regime where it is usually ine cient. An experimental investigation evidenced that usual ABH implementation gives rise to measurable geometric nonlinearity and typical nonlinear phenomena. In particular, strongly nonlinear regime and wave turbulence are reported. The nonlinear ABH beam is then modeled as a von Kármán plate with variable thickness. The model is solved numerically by using a modal method combined with an energy-conserving time integration scheme. The e ects of both the thickness pro le and the damping layer are then investigated in order to improve the damping properties of an ABH beam. It is found that a compromise between the two e ects can lead to an important gain of e ciency in the low frequency range.


Introduction
Controlling vibration is a major concern in many industrial applications today. Classical methods for passive vibration damping include the use of heavy viscoelastic layer [1] or tuned mass dampers [2]. These techniques are ecient and largely studied but their implementation is often limited by an important added mass or a narrow frequency range. The Acoustic Black Hole (ABH) eect is a particular surface damping method adapted 5 to mitigate vibrations on a wide frequency range without added mass. The name ABH in this context has been rst introduced by Mironov [3] and Krylov [4] and is now used in an increasing number of references in the literature [5,6,7]. Although the name may not be completely adapted as referring to an astrophysical phenomenon it is kept in this paper in order to be consistent with the current developments around this particular vibration absorber. 10 As described by Mironov [8], the ABH eect takes advantage of the exural wave properties in a plate with variable thickness. It can be shown that waves propagating in a beam or plate with thickness h(x) slow down without being reected and that their travel time in the prole tends to innity if thickness tends to zero. Due to the small thicknesses involved, manufacturing this "Acoustic Black Hole" prole is dicult and the residual thickness at the edge may be too large to obtain the desired eect. The addition of a thin viscoelastic 15 layer [4] can compensate the eect of the prole truncation and the reection coecient of the extremity is then * vivien.denis@univ-lemans.fr considerably reduced. Further theoretical works propose a more rened model for the reection coecient [5] of an ABH termination and demonstrate the resulting increase of the modal overlap [9]. Experimental evidences of the eect are numerous [10,11,12]. While it is very ecient at high frequencies, several works point out the ineciency of the ABH technique in the low frequency regime [9,13], which may be a drawback in its potential 20 applications.
The minimum thickness of the ABH extremity needed to obtain an important damping is very small (about hundred microns) and the displacement amplitudes at the extremity become large with respect to the thickness. Structures presenting this behavior are usually studied in the framework of nonlinear vibrations [14]. To the authors' knowledge, all numerical and theoretical models of ABH are based on linear assumptions such that the 25 eects of the large amplitudes are neglected. Nonlinear eects include couplings between linear eigenmodes of the structure [14,15]. Such couplings could be used for transferring energy from low frequency regime, where the ABH is inactive, to high frequency regime where vibrations can be eciently dissipated, thus improving the ABH performance.
This paper presents an investigation of the nonlinear behavior of ABH beams and evaluates the interest of 30 these nonlinear eects to improve the ABH damping eciency at low frequency. Sec. 2 reports an experimental highlight of nonlinear response of an ABH beam. Sec. 3 develops a von Kármán plate model with variable thickness for the ABH beam and its numerical resolution from a modal projection and an energy-conserving time-marching scheme. The results of the model are detailed in Sec. 4. Conclusions and perspectives are given in Sec. 5. 35 2 Experiment This section is devoted to an experimental investigation of the nonlinear behavior of a typical ABH beam. Dedicated forced vibration experiments are conducted in order to reveal the nonlinearities which can induce a variety of typical nonlinear phenomena including wave turbulence, characterized by a broadband Fourier spectrum and an energy transfer from low to high frequency range. 40

Experimental setup
The experimental study concerns two aluminium beams : the rst is uniform with dimensions 1.3 m × 20 mm × 5 mm. Given its thickness, this beam is awaited to behave linearly. The second selected beam, with dimensions 1.5 m × 20 mm × 5 mm, has an Acoustic Black Hole termination of length 0.35 m; the thickness of this extremity decreases quadratically down to 50 µm approximately (see the geometry of the sample in Fig. 3). 45 A thin viscoelastic tape (for example an electric tape) is applied on the ABH extremity. It is approximately 50 µm thick and its length is 0.2 m.
The experimental setup consists in a shaker (LDS V201), an amplier (LDS PA25E), a force sensor (PCB 208C03), an accelerometer (PCB 352C29) and an acquisition card (NI-USB4431). It is driven by a Labview interface and the results are exported and processed with MATLAB. The beam to be tested is vertically 50 suspended and excited at its midpoint with the shaker. The accelerometer is placed at an arbitrary location on the beam in order to record the vibrations at a given point.

Nonlinear regimes
The reference and ABH beams are excited at 83 Hz with a forcing amplitude increasing from 0 to 17 N in 40 s. By increasing gradually the forcing amplitude, the transition from periodic to chaotic regime can be 55 observed [16], and compared to the transition scenario given in [15]. Figs. 1(a-b) show spectrograms of output acceleration for the two congurations. The spectrum of the response for the uniform beam shows that most of the energy is concentrated in the excitation frequency at 83 Hz, which is an indication that the beam behaves linearly (see Fig. 1(a)). Additional components in the spectrum may be observed. Their amplitudes are small compared to the peak at the excitation frequency. They can be attributed to other elements in the experimental 60 setup, or residual nonlinearities in the shaker. The behavior is signicantly dierent for the ABH beam, see Fig. 1(b), the spectrum of which rapidly enriches with a very large number of harmonics present in the response. The response of the beam is periodic but nonlinear, with a spread of energy in a large harmonic series which, in this reported case, goes up to 3 kHz. Similar results have been obtained with numerous dierent forcing frequencies. Assuming that the beam's material constitutive behavior is linearly elastic, the nonlinearity present in the system comes from large amplitude oscillations of the beam, with respect to the thickness, and is denoted as geometric nonlinearity in the literature [14]. It signicantly inuences the system's response. Fig. 1(c) has been obtained for a dierent forcing frequency, namely 113 Hz. It reveals the case of a complete transition to the turbulent, chaotic behavior. A direct transition from the periodic regime, composed of numerous harmonics, to the turbulent regime, characterized by a broadband Fourier spectrum and a cascade of 70 energy to the high frequencies [17,18,19], is observed after 30 seconds of increasing the forcing amplitude. This observation is in the line of the transition to turbulence scenario for thin vibrating plates, already commented and analyzed in details with experiments and numerical simulations in [15,16]. A more peculiar feature is shown in Fig. 1(d), where the ABH beam excited at 102 Hz rst transits to the turbulent regime, clearly observed between 10 and 20 seconds, and then, while the forcing amplitude still increasing, recovers a periodic motion, 75 characterized by a Fourier spectrum composed of discrete peaks. Note that the regime is still highly nonlinear as the number of harmonics is very large with energy up to 8 kHz. This peculiar regime has been observed a signicant number of times during all the experiments realized, so that it appeared meaningful to report this behavior even if it is not analyzed specically. During these large amplitude periodic nonlinear oscillations, it has been observed that the ABH beam vibrates around a deected position, dierent from the position of the beam 80 at rest. Hence these oscillations can be interpreted as strongly nonlinear normal modes (NNM), i.e. periodic orbits, existing at large energy levels, and far from the rest (straight) position. They can be brought closer to earlier experimental observations shown on vibrating beams in [20]. In these strongly nonlinear regimes, the system may have a number of large-amplitude NNMs that are isolas in the phase space, and for some particular conditions, nonlinear oscillations can be trapped in one of these regimes, as observed in Fig. 1

Energy transfer between low and high frequencies
In order to better understand the nonlinearity and the benet that can be gained from it, the beams are now excited with a ltered white noise in the 100-500 Hz range, and for several increasing amplitudes. This frequency band has been selected as being the band in which the ABH is generally not ecient enough, and where the nonlinearity can bring an improvement by transferring the energy to higher frequencies. The parameter σ 90 denotes the standard deviation of the exciting force for each case. For a white noise, σ is proportional to the square root of the power. The Power Spectral Density (PSD) of the acceleration a(t), normalized by σ 2 , is dened as: It is displayed in Fig. 2 for uniform and ABH beams. Fig. 2(a) shows that the uniform beam exhibits a linear behavior since the power is concentrated in the excitation range, except for few artefacts that may come from 95 the setup and happen however at a much lower scale. Most importantly, the spectrum does not depend on the amplitude. On the contrary, in the ABH case (see Fig. 2(b)), the power largely leaks outside the excitation range for high levels of excitation. On the zoom of Fig. 2(c) a small reduction of the vibratory levels with the increase of amplitude can actually be observed for the ABH beam, more specically on the two peaks around 360 and 450 Hz. This is interpreted as an energy transfer from low to high frequency domains, the latter being 100 inherently more damped. This is particularly interesting as it could be used in order to increase the eciency of the ABH treatment at low frequencies. Note that here the transfer is not important enough and the gain in the damping of vibrations is very limited. The aim of the present study is thus to develop a model of the ABH including geometric nonlinearity, in order to analyze how this transfer can be optimized for a better eciency of the ABH in the low frequency range. 105 3 Model of nonlinear acoustic black hole

A von Kármán plate model with variable thickness
The ABH beam is modelled as a nonlinear von Kármán plate [14,21] with variable thickness. It is emphasized that the plate prole is assumed to vary along the x direction only in this study, hence plate characteristics such as mass density ρ and thickness h only depend on x. The equation of motion for the transverse displacement 110 w(x, y, t) can then be written, following the developments proposed in [22,23], as: where D(x), ρ(x), h(x) and ν are respectively bending stiness, mass density, thickness and Poisson ratio of the plate. The term p(x, y, t) represents the exciting force per unit surface. The Airy stress function F (x, y, t) respects the following compatibility equation: where E(x) is the Young modulus of the plate. Finally L(f, g) is the Monge-Ampère bilinear operator which 115 can be written, in cartesian coordinates, as: where f xx denotes the second derivative of f with respect to x.
The geometry of the plate is shown in Fig. 3. It is composed of three dierent regions. The rst one on the right is a uniform region (primary structure) with constant thickness h 0 . The second region is the ABH where the thickness decreases from h 0 to h t , which represents the minimum thickness of the plate. An extension of 120 the small thickness region is considered in the left part of the beam, which is parameterized by its length l add that can be tuned in order to increase the nonlinear behavior of the structure. The thickness h p (x) of the plate can thus be written as: where m is the exponent of the ABH prole and l ABH is the length of the region of variable thickness. The plate has a width b and a total length L. In all the cases tested in this study, the exponent m is selected as m=2. 125 The eect of the damping layer is represented with equivalent mechanical properties that modify the bending stiness, the mass density and the thickness, following the model of Ross-Ungar-Kerwin [1]. As shown in Fig. 3, the added viscoelastic layer is assumed to be of thickness h l , and located in the interval [x l1 , x l2 ]. Its eect on the exural rigidity is taken into account more easily in the frequency domain, where the damping is represented as an imaginary part of the complex bending stiness D * (x), which can be written as: where j is the imaginary unit, D p (x), E p , η p and h p (x) are respectively the bending stiness, the Young modulus, the loss factor and the thickness of the plate without the layer, while E l , η l and h l are the Young modulus, the loss factor and the thickness of the damping layer, respectively. The real bending stiness D(x) is dened as the real part of D * (x): which is nally inserted into the equations of motion (3)-(4) in order to take into account all these eects in 135 the nal model. The eect of the mass added by the presence of the viscoelastic layer is taken into account by introducing the following equivalent mass density ρ(x): where ρ p and ρ l are the mass densities of the structure and the layer respectively. The total thickness h(x) of the sandwich plate can nally be written as: The equivalent Young modulus of the sandwich plate that have to be inserted in (4) for the compatibility 140 equation on the Airy stress function, is deduced from the real bending stiness D(x) and the total thickness as: x z More concisely, the equations of motion can be written as: with A(x) = 1/E(x)h(x), and and ♦ the spatial operators applied to w and F : In the case of a plate with free edges, the boundary conditions write, on an edge with normal line n and tangent one τ [9,23]: One can note that Eq. (15c) are not the classical free conditions but are sucient conditions to impose an edge free of loads in the in-plane direction [23].

Modal approach
A modal approach is used for the spatial discretization of the problem, following the general procedure 145 described in [24]. Here the diculty comes from the plate's variable thickness which, in turn, gives rise to a more complex formulation for the linear operators given in Eqs. (12). The eigenmodes are computed for the unforced, conservative problem. For that purpose, the real bending stiness D, dened in (8) is used. The conservative and linear associated problems to be solved nally read: to which are added the boundary conditions (15). These are eigenvalue problems with solutions (ω k , Φ k (x, y)) and (ζ n , Ψ n (x, y)), respectively. The displacement and the Airy stress function are then expressed as: and where q k (t) and β n (t) are the modal coordinates for the transverse displacement and the Airy stress function, respectively. The integers N Φ and N Ψ represent the number of transverse and in-plane modes kept in the modal 155 truncation.
It is emphasized that the modes are orthogonal with respect to the mass: and normed such that: The projection on the linear modes gives two equivalent formulations of the problem. The formulation (22) is quadratic in (q, β) and can be written as: with M s = S ρ(x)h(x)Φ s (x, y) 2 dS the modal masses, F s = 1 Ms S pΦ s dS the modal force, and the coupling coecients: and The formulation (25), where variables β have been eliminated to keep only the modal amplitudes q of the transverse displacement w, shows a cubic nonlinearity in q: The Γ s k,m,n coecients are here calculated from H l m,n and E s k,l : Although completely equivalent, the two dierent formulations can be used for dierent purposes. As shown 165 in section 3.5, an energy-conserving scheme can be adopted for a robust and accurate time integration of the system. As shown in [24], this scheme particularly suits the quadratic formulation. On the other hand, the cubic formulation (25) shows that in-plane variables β can be eliminated from the procedure. As two dierent integers have been introduced in the modal truncation given in Eqs. (18)- (19), one can use the cubic formulation in order to verify the convergence of the solution as function of the number of in-plane modes N Ψ . Indeed, 170 monitoring the convergence of the terms of the tensor Γ s k,m,n versus N Ψ gives the sucient number of in-plane modes needed for convergence. This will be shown in more details in the result section, see 4.2.
One can remark that the properties H p i,j = H p j,i and H p i,j = E i p,j already described in [23,24] for the coupling tensors are also valid in this case and allow one to reduce the number of computed coecients and thus the computational burden associated with this preliminary step of oine calculations. 175

Computation of the linear modes
Due to the variable thickness, the solution to the eigenvalue problems (16)(17) is particularly dicult and needs a dedicated numerical approach. The solutions are obtained using a nite dierence method thoroughly described in [9]. The use of a uniform grid for a structure where the wavelength varies spatially can result in a high computational cost. This problem is solved using a coordinate change and a transformed problem, 180 which is equivalent to using a grid adapted to the exural wavelength variation. Since the wavelength λ = 2π 4 Eh 2 /12ρ(1 − ν 2 )ω 2 varies as the square root of the thickness according to classical plate theory [25], the coordinate change is selected as follows: wherex denes the transformed coordinate.
First the displacement w(x, y), the bending stiness D(x), the mass density ρ(x) and the thickness h(x) are approximated with grid functions w n,q , D n , ρ n and h n , respectively, where n and q are the spatial indexes such that x = n∆x and y = q∆ y , with ∆x and ∆ y the spatial steps along thex and y directions. Then the eigenvalue problem in transverse displacement (16) is approximated with a transformed discrete problem, which can be written as [26,9]: where δx − , δx + , µx − , µx + and δ yy are the nite dierence operators dened in A. 185 The in-plane problem is treated in the same manner. The convergence of the method has been extensively studied in [9]. The results presented in this article have been obtained using a 2000 × 100 grid. It is reminded that nite dierence methods (especially those using second-order approximations), although converged, still presents an inherent numerical dispersion [26]. Other numerical methods, such as Finite Element Analysis (FEA), are indeed also suitable to nd the desired solutions, as it has for instance been recently done in the 190 current context in [13] for circular ABH.
The resolution of the discrete eigenvalue problems yields the out-of-plane and in-plane modes and the associated eigenvalues. The coecients H p i,j , E s k,l and Γ s k,m,n are computed with the expressions (23,24,26) from the numerical mode shapes, following the general procedure also described in [27] for nite element shell models. 195

Losses
In accordance with the general modal framework used to solve the problem, the losses are taken into account by using a modal damping ratio for each oscillator equation. Assuming they remain small, the modal damping ratios ξ s are obtained by solving the eigenvalue problem in displacement with complex bending stiness: According to the literature [28], the eigenvalues ω * k of the problem dened in Eq. (29) can be written as: whereω s are the natural frequencies and ξ s are the modal damping ratios. The modal damping ratios are then directly used in the system (22a-22b), which can nally be written as: This method is advantageous since it yields a dissipative time evolution problem that can be solved in the same manner as for the conservative one. Note that modal damping have been introduced only on Eq. (32) for the transverse motions w. This can be justied by considering the cubic formulation given in Eq. (25). Indeed the in-plane motions can be condensed in order to use a closed formulation for the modal displacements q s only where damping has to be introduced. Secondly, inertia eects are not taken into account for the in-plane 205 motions, which gives a second justications for neglecting possible losses in Eq. (31).

Time integration scheme
The system of Eqs. (31-32) is solved using an energy-conserving scheme described in [29,24], that is conditionally stable. The rst step is to discretize the modal coecients q s (t) → q n s (where t = n∆t and ∆t is the time step) and the derivation operators with the proper nite dierence operators. The conservative scheme can then be written [24]: where the operators δ tt , δ t. , µ t− , µ t. and e t− are dened in A.
The stability condition of such a scheme is [24]: where f e = 1/∆t is the sampling rate and f max is the largest frequency retained in the modal truncation for 210 the transverse motion w, i.e. f max = f NΦ .
4 Numerical results

Parameters of the simulated beams
The model is used to analyze and compare the damping performance of thirteen dierent beams, denoted U, E0, E0D0, E1, E2, E3, E4, E4D0, E4D1, E4D2, E4D3 and E4D4, and represented schematically in Fig. 4. 215 A special emphasis is put on their ability to improve the eciency of the ABH in the low frequency regime. All the tested beams have common parameters that are gathered in Tab. 1, while the distinctive features of each of these congurations are given in Tab. 2. Conguration U is a uniform beam and will be used as a reference case for comparing the dierent eects brought respectively by the ABH, the nonlinear additional length and the damping layer. Beams E0 to E4 are ABH beams without viscoelastic layer. Their additional length l add is 220 gradually increased in order to extend the interval of the beam where the thickness is minimal. Their behavior is expected to be more and more nonlinear. Conguration E0D0 has the same geometry than E0, the only dierence being the presence of the viscoelastic layer. Finally congurations E4D0 to E4D4 are all based on E4, with the presence of the damping layer of increasing length. As a memory aid, E stands for extension, D stands for damping and the number immediately after the letter is an indication for the concerned length. 225

Convergence study
The convergence of the numerical scheme is rst investigated. The main free parameters that have to be set are the number of modes kept in the truncation, both for transverse and in-plane modes, N Φ and N Ψ , and the time step ∆t, or, equivalently, the sampling rate f e . In order to ensure a converged representation of the in-plane motions, the convergence of coecient Γ p p,p,p with respect to N Ψ is investigated. An example is shown 230 in Fig. 6, for p=29 (which correspond to the transverse mode (23,0) with 23 nodal lines along x and no nodal lines along y, the modes being sorted by increasing eigenfrequencies), for two dierent congurations : the reference beam U and the ABH E0 with a small length of decreasing thickness and no damping layer. The corresponding modal shape for E0 is plotted in Fig. 5(a).
In all cases, the coupling term Γ p p,p,p for the simple conguration U converges very quickly: approximately 235 100 in-plane modes are sucient in order to ensure a converged value and thus a ne representation of the inplane motions. On the other hand, the convergence for the beam E0 is much slower, as evidenced in Fig. 6(a). Note also that the value of the coupling coecient is also much larger in case E0 than U, which was awaited since the small thickness region is designed in order to increase the nonlinearity in the system. The rst consequence is thus the values of the nonlinear coupling coecients, which are substantially increased. The convergence of 240 Γ p p,p,p for mode p=29 is slow and behaves incrementally; which suggests that some in-plane modes contributes strongly to the nal value of the coecients while others do not. In fact, the convergence can be made much faster by sorting the in-plane modes by families instead of sorting them by increasing eigenfrequencies. Let us dene in-plane family n the family containing the modes with n+1 nodal lines across the y-dimension (including the boundaries). Two examples are shown in Fig. 5(c-d): mode Ψ 101,2 in Fig. 5(c) belongs to family 1 while 245 mode Ψ 61,3 in Fig. 5(d) belongs to family 2. Fig. 6(b) shows the result of the computation when the in-plane modes are sorted by families. Remarkably enough, only modes of family 1 actually contribute to the value of the coupling term for this specic transverse mode p=29. The same remarks can be made for other value of p and it is veried that these result hold for N Φ = 200 out-of-plane modes in the modal truncation. The sorting procedure allows to eliminate non-contributing families from the computation, which greatly reduces 250 the computational cost. The number of transverse modes N Φ depends on the frequency band one would like to cover in the dynamical simulations. It depends on the nonlinear regime at hand and the largest frequency present in the spectrum of the solution. For all the presented simulations, N Φ =200 has been selected, which gives an accurate representation of the solution up to the frequency 7 kHz for the most nonlinear conguration E4. This number has been found to be sucient for the tested regimes. Finally, once the number of in-plane modes chosen, the sampling rate f e can be simply selected thanks to the stability rule given by Eq. (35). However a larger value has been retained with f e =100 kHz in all the computations, in order to have a more rened time representation.

Nonlinear regimes
In this section, the behavior of the model is investigated with respect to a harmonic forcing: localized at point (x 0 ,y 0 ), with xed frequency Ω and increasing amplitude G(t), in order to show the ability of the model in retrieving the dierent nonlinear phenomena experimentally observed. Two congurations are selected for that purpose, U and E0, with a forcing frequency of 115 Hz. This forcing frequency has been arbitrarily chosen to underline how the nonlinear eects appear, it is nonetheless close to 120 Hz which corresponds to an eigenfrequency of conguration E0; it is also close to 109 Hz which corresponds to an 265 eigenfrequency of conguration U. The forcing amplitude G(t) is linearly increasing from 0 to 20 N in 20 s. The spectra of output displacement at the excitation point (x 0 =0.6 m, y 0 =0.01 m) are plotted in Fig. 7. Fig. 7(a) shows that the uniform beam behaves linearly as expected, since the spectrum only contains the excitation frequency. It can be observed in Fig. 7(b) that the spectrum of E0 contains two distinct regimes: there is rst a periodic regime showing an enrichment of the spectrum with odd harmonics, which qualitatively 270 reproduces the results of the experiment of Sec. 2.2. Note that, as compared to the experiments, numerical spectra display only odd harmonics while experimental spectra contain both even and odd harmonics. The presence of only odd harmonics in the simulation is related to the fact that the model contains only cubic nonlinearities. Experimentally, geometric imperfections are unavoidable in real plates, especially in the ABH case where the thickness at the termination is very small. The presence of these imperfections creates in turn 275 quadratic nonlinearity, which gives rise to even harmonics [30,31]. The second regime observed in Fig. 7(b) is a chaotic one, associated with the Wave Turbulence (WT) phenomenon, as discussed in Sec. 2.2. Note that the von Kármán model is a moderate perturbation of the motions around the equilibrium position, with linearized angles. Hence the model is not able to produce periodic orbits around other positions than the conguration at rest. Consequently in all the simulated cases, the return to periodic orbits sometimes observed experimentally 280 when still increasing the forcing amplitude, has not been retrieved. The choice of not considering imperfections in the model has been made both for the sake of simplicity and also mainly because in turbulent vibratory regimes with large amplitude, the cubic nonlinearities dominate the quadratic ones, as shown theoretically and numerically in [32]. In order to observe the potential for an increased eciency of the ABH in low frequency, the beams U, E0 and E0D0 are excited with a ltered white noise on the 100-500 Hz range during 10 s, for several amplitudes, i.e.
several standard deviation σ, as it is done experimentally in Sec. 2.3. The Power Spectral Density of the output velocity of beams E0 and E0D0 at the excitation point (x 0 =0.6 m, y 0 =0.01 m) is displayed in Fig. 8 for dierent amplitudes of excitation. For the covered ABH beam E0D0, experimental results are qualitatively reproduced 290 since energy leaks outside the excitation range, and the transfer becomes more important with the amplitude level, without however aecting the resonance peaks in the excitation range. This energy transfer towards higher frequency (energy cascade due to the Wave Turbulence [15,33]) is much more noticeable and important in the E0 case, since the level outside excitation range easily reaches the level within, and a larger frequency range is concerned (see Fig. 8(c)). At large amplitude, the level of the resonance peaks within the excitation range is 295 slightly reduced because this energy is transferred to higher frequencies where it is dissipated. It is important to note that while very little damping is present in the beam, nonlinearities allow for an interesting reduction of vibration in this case.
In order to evaluate the performance gain of the ABH due to the aforementioned nonlinear eects, an indicator I c,σ is built based on the responses to the white noise excitation of Fig. 8. It is dened as: where c refers to the current conguration, S aa is the PSD of acceleration output a(t) and S pp is the PSD of the forcing p(t). The comparison is drawn out with a reference case denoted as I ref .
This reference case is selected as the uniform beam, conguration U, excited with the minimum amplitude (σ = 0.11N ), hence giving rise to the indicator denoted as E c,σ : Indicator E c,σ is plotted in Fig. 9 for the three congurations U, E0 and E0D0. A quasi-constant value is 305 obtained for the uniform beam, as it is expected. Although the PSD is averaged over 10 s, the variability of the white noise excitation induces that the same exact value cannot be obtained for a linear case at dierent excitation levels, as it is shown by the small variations of case U. The value of the indicator for beam E0 is constant for small σ and then decreases, indicating that less energy is present in the 100-500 Hz range, as it was shown by the spectrum of Fig. 8(c). Finally, the indicator has a constant and smaller value for beam E0D0, 310 reecting a better performance due to the damping layer but also smaller nonlinear eects and no signicant energy transfer towards the high frequency domain, as it is shown in Fig. 8(b). This rst result shows that a compromise has to be found between two competing eects : nonlinearity and damping. A too large value of damping prevents the nonlinear eect to settle down and thus creates an overdamped situation with no dependence on vibration amplitude. The next sections are devoted to a quantitative study of these two eects 315 in order to optimize the low-frequency eciency of the ABH.

Eect of the thickness prole
The importance of the nonlinear eects is rst investigated by simulating the responses of congurations E1 to E4, which have increasing additional length of constant and minimum thickness. Increasing this length should enhance the nonlinear behavior. Moreover, the modal density will also increase in the excited frequency 320 band, hence favoring the turbulent behavior and the energy transfer.
The procedure described in Sec. 4.4 is thus applied to beams E1 to E4 with gradually increasing length of extensions, according to Fig. 4. The results for the indicator E c,σ are plotted in Fig. 10. Conguration E1 shows that adding a small extension positively aects the nonlinear behavior since the structure becomes more sensitive to the excitation level: Indicator E c,σ decreases with respect to the amplitude of excitation and 325 reaches values close to -9 dB. Congurations E2 to E4 are even more interesting from this point of view since the indicator reaches values around -12 dB. However, the similar values obtained for these congurations seem to indicate a saturation eect, i.e. that a longer extension does not result in a better energy transfer or a better dissipation due to nonlinearities.
As known from wave turbulence theory, nite-size eects alters the energy cascade due to the fact that a 330 continuum of wavevectors is not at hand anymore, so that the resonance relationships shall then be veried   on a discrete set composed of the modes of the system [19,34,35]. In our case where energy transfer has to be optimized, the best situation would then be to have a nonlinear termination as large as possible in order to ensure the best modal density in the low-frequency range, and thus optimize the energy ux toward high frequencies. This situation is out of range for the selected beams where the number of eigenmodes in the tested 335 frequency band 100-500 Hz increases from 8 (E0 conguration) to 17 (conguration E4). Fig. 11 shows the energy spectra for beams E0 to E4, and for the largest excitation amplitude tested, σ=2.6 N, highlighting the fact that the energy is redistributed over the prominent peaks dened by the eigenfrequencies, when extending the additional length, which explains the observed saturation on the indicator. However, better performance could be awaited for two-dimensional ABH as described in [5,36], where the extension of the minimum thickness 340 region can be signicantly larger.

Balance of damping and nonlinearity
As shown in the previous section, the additional length l add has to be chosen as large as possible in order to enhance the nonlinear eects. The length of the damping layer is now investigated, in order to nd the best compromise between nonlinearity and damping, and to ensure an increasing performance of the ABH thanks to 345 the nonlinear behavior.
In order to observe the competition between damping and nonlinearities in the ABH beam, dierent lengths of damping layer are added to conguration E4, resulting in congurations E4D0 to E4D4 (see Fig. 4). The behavior of the indicator E c,σ for the associated simulations with a ltered white noise excitation are plotted in Fig. 12. Beam E4D0 reveals an interesting combination since the indicator E c,σ is lower than for E4 for small 350 excitation levels, due to the more important damping in the structure. When increasing the vibration amplitude, indicator E c,σ for E4D0 shows then a decreasing behavior indicating that the nonlinearity is activated and a better eciency of the ABH is obtained. The eect can clearly be seen in Fig. 13, which reveals reductions of the PSD down to 14 dB in the 100-500 Hz range (see Fig. 13(b)), with the σ 2 normalization. More specically one can observe that the resonance frequencies tend to larger values when increasing the nonlinearity, reecting 355 the hardening behavior of the plate. Secondly, all resonance peaks widen due to the appearance of the strongly nonlinear behavior associated with the chaotic dynamics, and globally decrease as energy has own to higher frequencies.
When the damping layer length is increased, gradually covering the nonlinear extension, nonlinear eects tend to disappear, as shown in Fig. 12 for E4D2 to E4D4 congurations, where the indicator is seen to stay at a 360 constant value whatever the excitation level. A longer layer, hence a more important damping, however results in a better performance of the ABH. This study shows that both damping and nonlinear eect can be decoupled as the additional length (or the length of the ABH region) and the length (or the material characteristics) of the damping layer can be independently adjusted in order to nd an optimal behavior, depending of the context.

Conclusions and perspectives 365
This study is devoted to the nonlinear eects in ABH beams, and the potential benet that can be obtained from using them as a vector for transferring energy from the low frequencies, where ABH is known to be usually inecient, to the high frequencies; this transfer is provided by the wave turbulence regime and its associated energy cascade. Experimental results on an ABH beam have been rst reported, evidencing that nonlinearities are activated, giving rise to typical nonlinear phenomena. The thinness of the ABH termination are the support 370 for the geometric nonlinearity, given the amplitude of motions involved.
A von Kármán plate model with variable thickness has then been derived in order to study numerically the main eects involved, with the aim of optimizing the energy transfer and improving the eciency of the ABH in the low frequency range. A full numerical model has been developed, including the geometry variations, the damping due to the added viscoelastic layer, and the geometric nonlinearity due to large-amplitude vibrations. 375 A modal approach with an energy-conserving scheme has been used for the resolution, this method allowing one to ensure a very good accuracy and convergence of the model for reasonable computational burden, together with a ne description of the frequency-dependent losses present in the structure.
A parametric study on a number of selected geometries has then been led, in order to investigate how the two competing eects represented by both the nonlinearity and the damping, can be tuned for obtaining an important 380 gain in eciency for the ABH. A good compromise has been found by using a long additional termination such that the nonlinearity can develop into a wave turbulence regime, together with a moderate length of the damping layer. Indeed, using a too large value of the damping overdamps all vibrations and counteracts the nonlinear behavior such that no improvement in eciency is obtained with increased amplitudes. On the other hand, letting the chaotic dynamics settle down allows for the energy transfer in high frequencies where vibrations are 385 damped, and thus improves the eciency as amplitude increases. More generally, the results presented in this study show that previous linear models for the ABH may not apply anymore to some particular congurations, or need to be used with great care as they neglect an important nonlinear eect present in the dynamics. Further work dealing with plates embedding two-dimensional circular ABH with nonlinear eects could also be very interesting and closer to possible applications in mechanical engineering. As noted in the course of this study, the low-frequency range. This fact should optimize the energy cascade, and could be more advantageous for evacuating the energy toward high frequencies.