betaFeedback

WorksPhysics and Astronomy

Effective T-matrix of a cylinder filled with a random 2 dimensional particulate

Abstract

When a wave, such as sound or light, scatters within a densely packed particulate, it can be rescattered many times between the particles, which is called multiple scattering. Multiple scattering can be unavoidable when: trying to use sound waves to measure a dense particulate, such as a composite with reinforcing fibers. Here we solve from first principles multiple scattering of scalar waves, including acoustic, for any frequency from a set of 2D particles confined in a circular area. This case has not been solved yet, and its solution is important to perform numerical validation, as particles within a cylinder require only a finite number of particles to perform direct numerical simulations. The method we use involves ensemble averaging over particle configurations, which leads us to deduce an effective T-matrix for the whole cylinder, which can be used to easily describe the scattering from any incident wave. In the specific case when the particles are monopole scatters, the expression of this effective T-matrix simplifies and reduces to the T-matrix of a homogeneous cylinder with an effective wavenumber k⋆k_\star. To validate our theoretical predictions we develop an efficient Monte Carlo method and conclude that our theoretical predictions are highly accurate for a broad range of frequencies.

Introduction

Ensemble averaging. Multiple scattering is unavoidable when using waves to characterise a particulate composite, or designing metamaterials to control wave propagation. Further, the number of particles in most applications makes direct numerical simulations impossible for current computing power, though there are some notable attempts [13, 15]. Even if such simulations were possible, the positions of the particles are impossible to know for one particular sample. A mathematical method to solve this problem, called ensemble average, is to take an average over all possible particle positions. This type of averaging can occur naturally for fluids and gases, as the particles are moving. That is, taking an average of the scattered signal over time can be equivalent to an ensemble average1. Assuming that ensemble averaging is equivalent to averaging over time and space, which is often the case, is called the ergodicity assumption [22].

What is known. The scenarios that are best understood are: 1) waves in an infinite media with no boundaries [36] [27], and 2) plane waves incidence on a half-space or plate filled with particles [33, 34]. Both scenarios have been considered to obtain effective wavenumbers [37] [6, 19]. Though we mention the methods that use Lipmann-Schwinger for acoustics, they involve an extra integral which is often omitted [20] and complicates the calculations. Our approach in this work is valid for any type of scatterer and scalar waves. Both scenarios of using plane waves and an infinite media have several applications (typically when considering layered media such as planetary atmospheres, layers in the ocean, or soils) but one significant drawback is that it has been very challenging to numerically validate the assumptions used for these methods. Both use statistical assumptions, such as the Quasi-Crystalline Assumption (QCA) [17] that is not based on an asymptotic approximation. Validation is needed to establish the range of validity of these assumptions. However, direct numerical simulations of scattering from a configuration of particles for both planes and infinite media require a huge number of particles [14], [9] or the introduction of periodic boundaries which can introduce artifacts [3].

The cylindrical setting. The methods developed to describe the average plane wave propagating in a disordered particulate plate or half-space can now be extended to other geometries [10]. The ideal scenario to compare theoretical predictions with direct numerical simulations is to have cylindrical particles inside a cylinder, as this reduces the problem to two dimensions and we need only a finite number of particles for the direct numerical simulations. See Figure 1 for an illustration. This is the simplest case to perform numerical validation of a very general theory [10]. Further, we show in this work that the effective dispersion equation for the cylindrical geometry is the same as the plane-wave case. So numerically validating the cylindrical geometry will also serve as numerical validation of the dispersion equation for plane waves and all geometries. We also note that it appears that the cylindrical setting has never been solved from a first principles approach. The formulas we provide are also valid for any inter-particle pair correlation.

Image not extracted

Figure 1. Consider a circular region filled with sound-hard particles. Both panels compare the average scattered wave ⟨usc⟩\langle u_{\mathrm{sc}}\rangle, for a source with only one mode: uinc=VN(r)u_{\mathrm{inc}} = V_N(r), when using our effective wave method and a brute force Monte Carlo approach. The Monte Carlo approach simulates the scattered field from either one, M=1M = 1, or two hundred, M=200M = 200, configurations of particles and then takes the average over these configurations. In general, the average scattered wave is given by (17). The left panel shows that when using an incident wave uinc=V0(r)u_{\mathrm{inc}} = V_0(r), with radial symmetry, then the only scattered mode also has radial symmetry after average. The right panel shows how a source with a 60° rotation symmetry also leads to a scattered wave with the same symmetry.

Industrial applications. Beyond numerical validation, there are industrial applications that need a method to calculate waves scattered from a cylinder with particles. Examples of cylinders filled with cylindrical particles include concrete beams reinforced with iron, cables filled with wires, or fibre-reinforced composite [1]. Applications include designing cylinders with exotic effective properties or developing methods to measure the cylindrical particles [24, 25]. In terms of measurement, it is likely that more information can be extracted from waves scattered from a cylinder filled with a particulate than just plane wave reflection from a plate filled with the same particulate.

Effective properties. The most common approach to model the average scattering from, say, a spherical or cylindrical region with particles is to assume the region is homogeneous with some effective properties [32, 35] [31, 26, 5, 18, 30], and then use the standard boundary conditions such as continuity of displacement. This approach is valid for low-frequency [31] but for finite frequencies is incorrect in three dimensions [10], and we demonstrate the same here for two dimensions in this work. To obtain an accurate model for a finite frequency the boundary condition needs to be deduced from first principles, together with an eigen system for the effective wavenumbers [10]. Although the process is more involved, the final expression for the average scattered wave from a cylindrical region is simple: the average scattered wave can be calculated from an effective T-matrix for any incident frequency, source, and particle properties. We stress that without deducing the results from first principles, as we do here, it would not be possible to just guess the form of this effective T-matrix.

Monopole scatterers. One somewhat surprising result we deduce is that if the particles scatter only monopole waves, that is waves that have radial symmetry, then the material as a whole behaves as homogeneous, where the mass density is the same as the background, and the bulk density is given by a simple formula. We deduce this for particles in a cylinder and hypothesize that it is true for any material filled with monopole scatterers, even when including all orders of multiple scattering. Beyond just curiosity, there are many particles that behave approximately like a monopole scatterer, and therefore the simple formulas we deduced are appropriate. For example, in acoustics void-like particles are approximately monopoles for a broad frequency range, see the Dirichlet case in Figure 2. In elasticity, particles become approximately monopole when the bulk modulus is much greater than the shear modulus [4]. Other cases include resonators such as a split ring resonator [28].

Image not extracted

Figure 2. Compares various methods to calculate the component T0T_0 of the T-matrix of a cylinder filled with particles. The solid red line is our effective waves method (EWM) (50), the black points are from a brute force Monte Carlo method (MC), and the dashed blue line is our method when only monopole scattering is accounted for (EWM-MA) (1). The left (right) graph shows the results for sound soft (hard) particles. In both cases, the general expression of the effective T-matrix matches the Monte Carlo results. The EWM-MA method only matches well with the Monte Carlo for sound soft scatterers for low frequencies.

Overview of the method. After ensemble averaging over particle configurations within a cylindrical region, the system inherits cylindrical symmetry. For example, if the source has radial symmetry then the average scattered field will also have radial symmetry. This is also true for sources with more general rotational symmetry resulting in scattered fields of the same rotational symmetry (cf. Figure 1). In this paper, we take advantage of this mode-to-mode symmetry to analyze the general behaviour of the random particulate material independently from the incident field.

After denoting with VnV_n and UnU_n the regular and outgoing cylindrical waves of order nn (cf. 2.2), the cylindrical symmetry translates as follows: when the exciting source is VnV_n, then the average scattered field is TnUnT_nU_n where the complex number TnT_n only depends on the properties of the random particulate cylinder (radius and properties of the particles). Since the scattering problem is linear, the knowledge of the TnT_n allows us to describe the scattering from any incident field, after decomposing the latter into the modes VnV_n. Having simple expressions for TnT_n is crucial to help guide methods to characterize or design particulate materials. We do so by using the effective wave method approach [10]. Finally, we validate our results with an adapted Monte Carlo method in which the rate of convergence is accelerated thanks to the cylindrical symmetry.

Overview of this paper

In § 2, we first introduce the statistics of the random particulate material and the required notations for the ensemble averaging. We then define the T-matrix of the effective cylinder whose exact formula depends on the solutions of the averaged Foldy-Lax equations.

The latter is solved in § 3 by using the effective waves method, which consists in finding solutions that are isotropic waves with a complex wavenumber k⋆k_{\star}. The method leads to an eigenvalue problem called the dispersion equation, whose eigenvalue provides k⋆k_{\star}.

In § 4, we use the expression of the solutions of the averaged Foldy-Lax equations to deduce a formula of the effective T-matrix. The latter is very simple when the particles are monopole scatterers:

Tn=−CnDnwithCn=kJn′(kR~)Jn(k⋆R~)−k⋆Jn(kR~)Jn′(k⋆R~),Dn=kHn′(kR~)Jn(k⋆R~)−k⋆Hn(kR~)Jn′(k⋆R~)(1)T_n=-\frac{C_n}{D_n} \qquad\text{with}\qquad \begin{aligned} C_n&=kJ_n'(k\widetilde{R})J_n(k_{\star}\widetilde{R})-k_{\star}J_n(k\widetilde{R})J_n'(k_{\star}\widetilde{R}),\\ D_n&=kH_n'(k\widetilde{R})J_n(k_{\star}\widetilde{R})-k_{\star}H_n(k\widetilde{R})J_n'(k_{\star}\widetilde{R}) \end{aligned} \tag*{(1)}

where R~:=R−a\widetilde{R}:=R-a. This result is remarkable because the above expression corresponds to the T-matrix of a homogeneous acoustic cylinder of radius R~\widetilde{R}, sound speed c⋆=ω/k⋆c_{\star}=\omega/k_{\star} and density ρ\rho which is equal to the background medium. Particles that are approximately monopole scatterers appear in mainly two scenarios: either small sound soft particles or resonators [29]. For all these cases, the T-matrix (1) shows us that the effective wavenumber k⋆k_{\star} suffices to describe the random material.

When the particles are not monopole scatterers, (1) is not exact. The exact formula is given by

Tn=−∑n′Cn−n′Fn′∑n′Dn−n′Fn′(2)T_n=-\frac{\sum_{n'}C_{n-n'}F_{n'}}{\sum_{n'}D_{n-n'}F_{n'}} \tag*{(2)}

where the CnC_n and DnD_n are the same as before, and the weights FnF_n are the eigenfunctions of the dispersion equation, associated with the effective wavenumber k⋆k_{\star}.

For monopole scatterers, we have that (n′=0)(n'=0), and the above reduces to (1). We note that often F0F_0 is the largest term, which explains why (1) can give accurate results for non-monopole scatterers, see for example the numerical results for sound soft (Dirichlet) particles shown in Figure 2 in the low-frequency regime. In this same figure, we also see how for sound hard particles (Neumann), which are dipole scatters, the Monte Carlo results and the formula (2) closely match, whereas the T-matrix for monopole scatters doesn’t match.

Random particulate material

Deterministic scattering from J particles

Here we summarise some results for multiple scattering of acoustic waves, in the time-harmonic regime, by a collection of JJ cylinders, referred as to particles, which axis are aligned with the zz-axis. The centre of the particle ii is identified by ri∈R2\mathbf{r}_i\in\mathbb{R}^{2} as shown in Figure 3 and is assumed to be confined in an area denoted by Ri\mathcal{R}_i.

Image not extracted

Figure 3. Illustration of a possible configuration of four particles. The particles are cylinders whose axes are aligned with the zz-axis. The position ri∈R\mathbf{r}_i \in\mathbb{R} of the ithi^{\mathrm{th}} particle is determined by the intersection of its axis with the plane xOyxOy. The particles are of radius aa and physically contained in a cylinder of radius RR. In this specific example, all the particles are the same and the centres ri\mathbf{r}_i are therefore confined in the same disc R1\mathcal{R}_1 shown in green.

The propagation of waves in free space is governed by the 2D Helmholtz equation

Δu+k2u=0(3)\Delta u+k^{2}u=0 \tag*{(3)}

where Δ:=∂x2+∂y2\Delta:=\partial_x^{2}+\partial_y^{2} is the 2D Laplace operator and k∈Rk\in\mathbb{R} is the wavenumber of the homogeneous background. Consider the following two independent bases of the Helmholtz equation VnV_n and

UnU_n defined by

Un(kr):=Hn(kr)einθ∀r∈R2∖{0}Vn(kr):=Jn(kr)einθ∀r∈R2(4)\begin{aligned} U_n(kr) &:= H_n(kr)e^{in\theta} && \forall r \in\mathbb{R}^2 \setminus\{0\}\\ V_n(kr) &:= J_n(kr)e^{in\theta} && \forall r \in\mathbb{R}^2 \tag*{(4)} \end{aligned}

where kk is the background wavenumber and (r,θ)(r,\theta) are the polar coordinates of rr, ie r=(rcos⁡θ,rsin⁡θ)r=(r\cos\theta,r\sin\theta). The specific solutions Vn(kr)V_n(kr) have the particularity of being smooth while Un(kr)U_n(kr) have a singularity at the origin and are outgoing solutions.

Both the incident field uincu_{\mathrm{inc}} and the scattered field uscu_{\mathrm{sc}} are solutions of (3). While we assume that uincu_{\mathrm{inc}} is smooth and regular in the region that covers the particles2 the scattered field has to be a sum of outgoing fields from each particles centered at rir_i, as a result we can write

uinc(r)=∑n=−∞+∞gnVn(kr)(5)u_{\mathrm{inc}}(r)=\sum_{n=-\infty}^{+\infty}g_nV_n(kr) \tag*{(5)}
usc(r)=∑i=1J∑n=−∞+∞fniUn(kr−kri).(6)u_{\mathrm{sc}}(r)=\sum_{i=1}^{J}\sum_{n=-\infty}^{+\infty}f_n^iU_n(kr-kr_i). \tag*{(6)}

The coefficients gn∈Cg_n\in\mathbb{C} are given parameters of the system and fni∈Cf_n^i\in\mathbb{C} are unknowns that can be determined by following the Foldy-Lax self-consistent method [21]. The latter leads to the following system of equations

fni=Tni∑n′Vn′−n(kri)gn′+Tni∑j≠i∑n′Un′−n(kri−rj)fn′j∀n∈Z, ∀i=1…J,(7)f_n^i=T_n^i\sum_{n'}V_{n'-n}(kr_i)g_{n'}+T_n^i\sum_{j\ne i}\sum_{n'}U_{n'-n}(kr_i-r_j)f_{n'}^j\qquad\forall n\in\mathbb{Z},\ \forall i=1\ldots J, \tag*{(7)}

where TniT_n^i is the T-matrix of particle ii, which can represent a wide range of particles [7].

To give an example, the expression of a T-matrix of a homogeneous particle with wavenumber kik_i, density ρi\rho_i and radius aia_i is given by

Tni=−ρikJn′(kai)Jn(kiai)−ρkiJn(kai)Jn′(kiai)ρikHn′(kai)Jn(kiai)−ρkiHn(kai)Jn′(kiai).(8)T_n^i=-\frac{\rho_i k J_n'(k a_i)J_n(k_i a_i)-\rho k_i J_n(k a_i)J_n'(k_i a_i)}{\rho_i k H_n'(k a_i)J_n(k_i a_i)-\rho k_i H_n(k a_i)J_n'(k_i a_i)}. \tag*{(8)}

The system (7) totally determines the scattering coefficients fnif_n^i, which allows to solve the scattering problem from the JJ given particles. However, this result is not very useful in practice for two main reasons: first, the position of the particles is often unknown, and second, there can be a very large number of particles in most industrial applications [2]. The ensemble average over all particle positions, that we summarize below, solves both these problems from the computational standpoint.

Particulate distribution

We describe each particle, say particle ii, with two random variables: ri\mathbf{r}_i is the probability distribution of the centre of the particle in space, and λi∈S\lambda_i \in\mathcal{S} describes all other properties of the particle (radius aia_i, density ρi\rho_i, etc.).

The centre of a particle ri\mathbf{r}_i is restricted in a region Ri\mathcal{R}_i which depends on the particle’s properties λi\lambda_i. For example, ri\mathbf{r}_i has to be one radius away from the boundary ∂R\partial\mathcal{R} which completely contains all particles. In this paper we assume that ri\mathbf{r}_i is distributed uniformly over the set Ri(λi)\mathcal{R}_i(\lambda_i) defined by

Ri:={r∈R:dist⁡(r,∂R)>ai}.(9)\mathcal{R}_i:=\left\{\mathbf{r}\in\mathcal{R}:\operatorname{dist}(\mathbf{r},\partial\mathcal{R})>a_i\right\}. \tag*{(9)}

As a consequence,

p(ri,λi)=p(ri∣λi)p(λi)=p(λi)∣Ri∣,(10)p(\mathbf{r}_i,\lambda_i)=p(\mathbf{r}_i\mid\lambda_i)p(\lambda_i)=\frac{p(\lambda_i)}{|\mathcal{R}_i|}, \tag*{(10)}

where ∣Ri∣|\mathcal{R}_i| is the volume of Ri\mathcal{R}_i.

Joint particle distribution

Each particle is described by two random variables (λi,ri)(\lambda_i, r_i), i=1,…,Ji = 1,\ldots,J. In general, the particle positions are correlated, for example, no particles can overlap. The most commonly used term to describe inter-particle correlation is the pair correlation gg, which satisfies (cf. [16, eq. 8.1.2]):

p(r1;r2∣λ1;λ2)=g(r1,λ1;r2,λ2)J∣R1∣∣R2∣1J−1,(11)p(\mathbf{r}_1;\mathbf{r}_2|\lambda_1;\lambda_2)=g(\mathbf{r}_1,\lambda_1;\mathbf{r}_2,\lambda_2)\frac{J}{|\mathcal{R}_1||\mathcal{R}_2|}\frac{1}{J-1}, \tag*{(11)}

where on the left is the joint law of two particles’ position when their properties are known. The pair correlation gg describes how correlated any two particles are, when the positions and properties of all other particles are unknown. For example, if g=1g=1 for all values of its arguments then both rir_i and rjr_j in the above are independent and uniformly distributed over Ri\mathcal{R}_i and Rj\mathcal{R}_j, respectively (in the limit J→∞J\to\infty).

Finally, we introduce the density

n(λi):=J∣Ri∣p(λi)(number of λi types particles per unit volume).(12)n(\lambda_i):=\frac{J}{|\mathcal{R}_i|}p(\lambda_i)\quad\text{(number of }\lambda_i\text{ types particles per unit volume)}. \tag*{(12)}

Then we derive the following useful relation

p(rj,λj∣ri,λi)=p(rj,λj;ri,λi)p(ri,λi)=∣Ri∣p(λj)p(ri;rj∣λi;λj)=n(λj)J−1g(ri,rj),(13)p(\mathbf{r}_j,\lambda_j|\mathbf{r}_i,\lambda_i)=\frac{p(\mathbf{r}_j,\lambda_j;\mathbf{r}_i,\lambda_i)}{p(\mathbf{r}_i,\lambda_i)}=|\mathcal{R}_i|p(\lambda_j)p(\mathbf{r}_i;\mathbf{r}_j|\lambda_i;\lambda_j)=\frac{n(\lambda_j)}{J-1}g(\mathbf{r}_i,\mathbf{r}_j), \tag*{(13)}

where we used (10), (11) and (12).

The pair correlation

In this paper, we consider that the particles have a distribution which is isotropic and homogeneous in space. As a consequence, the pair correlation is of the form g(ri,λi;rj,λj)=g(∣rj−ri∣,λi,λj)g(\mathbf{r}_i,\lambda_i;\mathbf{r}_j,\lambda_j)=g(|\mathbf{r}_j-\mathbf{r}_i|,\lambda_i,\lambda_j). We assume the pair correlation is of the following form:

g(r,λ1,λ2)={0,r<a12,1+δg(r,λ1,λ2),a12<r<b12,1,r>b12,(14)g(r,\lambda_1,\lambda_2)= \begin{cases} 0, & r<a_{12},\\ 1+\delta g(r,\lambda_1,\lambda_2), & a_{12}<r<b_{12},\\ 1, & r>b_{12}, \end{cases} \tag*{(14)}

where g(r,λ1,λ2)=0g(r,\lambda_1,\lambda_2)=0 when particles overlap, with a12≥a1+a2a_{12}\geq a_1+a_2 and in general a12a_{12} can depend on λ1\lambda_1 and λ2\lambda_2. This form of the pair correlation moreover assumes that at a certain distance b12b_{12} from each other, the particles become uncorrelated, that is g(r,λ1,λ2)=1g(r,\lambda_1,\lambda_2)=1 for r>b12r>b_{12}. This assumption will lead to analytic simplifications, as well as being a good approximation for most disordered materials. A typical plot of the pair correlation is illustrated in Figure 4.

Image not extracted

Figure 4. Typical plot of the pair correlation function.

Definition of the effective T-matrix

The scattered field (2.3b) can be simplified after using Graf’s addition theorem (A.1,ii) with x=krx=kr and d=−krid=-k\mathbf{r}_i

usc(r)=∑n=−∞+∞F~nUn(kr)(15)u_{\mathrm{sc}}(r)=\sum_{n=-\infty}^{+\infty}\widetilde{F}_nU_n(kr) \tag*{(15)}
F~n:=∑i=1J∑n′=−∞+∞Vn′−n(−kri)fn′i.(16)\widetilde{F}_n:=\sum_{i=1}^{J}\sum_{n'=-\infty}^{+\infty}V_{n'-n}(-k\mathbf{r}_i)f_{n'}^i. \tag*{(16)}

The pair-correlation g(r)g(r) is zero for r<a12r<a_{12} because particles cannot overlap. The local minima (maxima) indicate that there is a distance rr from one fixed particle, say at r1r_1, where it is less (more) likely to find other particles. Finally, assuming g(r)=1g(r)=1 for r≥b12r\geq b_{12} means that a particle at r1r_1 becomes uncorrelated to particles which are further than b12b_{12}.

Note that (2.12) is only valid for r>Rr>R, nevertheless this is enough for the definition of the effective T-matrix below. Taking the ensemble average of the equation above, as defined in Appendix § 9, leads to

⟨u~sc(r)⟩=∑n=−∞+∞⟨f~n⟩Un(kr).(17)\langle\tilde{u}_{\mathrm{sc}}(r)\rangle=\sum_{n=-\infty}^{+\infty}\langle\tilde{f}_n\rangle U_n(kr). \tag*{(17)}

Since the scattering problem is linear with respect to the incident field, ⟨f~n⟩\langle\tilde{f}_n\rangle depends linearly on the modes gng_n of the incident field (5):

⟨f~n⟩=∑N=−∞+∞Tn,NgN(effective T-matrix definition).(18)\langle\tilde{f}_n\rangle=\sum_{N=-\infty}^{+\infty}\mathcal{T}_{n,N}g_N \quad\text{(effective T-matrix definition)}. \tag*{(18)}

This relation defines the T-matrix T\mathcal{T} of the averaged material which connects the modes of the incident field with the ones of the averaged scattered field. It allows us to describe the scattering from any incident field, provided the coefficients gng_n in (5).

In the specific case when the area where the particles are confined is circular, we have Tn,N=0\mathcal{T}_{n,N}=0 if n≠Nn\ne N (cf. Appendix § 12), so that (17) becomes

⟨u~sc(r)⟩=∑n=−∞+∞TngnUn(kr).(19)\langle\tilde{u}_{\mathrm{sc}}(r)\rangle=\sum_{n=-\infty}^{+\infty}\mathcal{T}_n g_n U_n(kr). \tag*{(19)}

where Tn:=Tn,n\mathcal{T}_n:=\mathcal{T}_{n,n}.

Computing Tn,N\mathcal{T}_{n,N} requires to compute ⟨f~n⟩\langle\tilde{f}_n\rangle. From (15), (B.1), and (B.3) we obtain

⟨f~n⟩=∫Sn(λ1)∫R1∑n′Vn′−n(−kr1)⟨fn′1⟩(r1,λ1) dr1dλ1.(20)\langle\tilde{f}_n\rangle=\int_{\mathcal{S}}\mathbf{n}(\lambda_1)\int_{\mathcal{R}_1}\sum_{n'}V_{n'-n}(-kr_1)\langle f_{n'}^{1}\rangle(r_1,\lambda_1)\,\mathrm{d}r_1\mathrm{d}\lambda_1. \tag*{(20)}

Here we used (B.7) to substitute ⟨fn⟩(r1,λ1)=⟨fn1⟩(r1,λ1)\langle f_n\rangle(r_1,\lambda_1)=\langle f_n^1\rangle(r_1,\lambda_1).

The function ⟨fn⟩(r1,λ1)\langle f_n\rangle(\mathbf{r}_1,\lambda_1) needs to be determined before we can compute ⟨Fn⟩\langle\mathcal{F}_n\rangle. In Appendix § 9.1 we show how to obtain the governing equation:

⟨fn⟩(r1,λ1)=Tn(λ1)∑n′Vn′−n(kr1)gn′+Tn(λ1)∑n′∫Sn(λ2)∫R2Un′−n(kr1−kr2)⟨fn′⟩(r1,λ1)g(∣r1−r2∣,λ1λ2) dr2dλ2(21)\langle f_n\rangle(\mathbf{r}_1,\lambda_1)=T_n(\lambda_1)\sum_{n'}V_{n'-n}(k\mathbf{r}_1)g_{n'}+T_n(\lambda_1)\sum_{n'}\int_{\mathcal{S}}n(\lambda_2)\int_{\mathcal{R}_2}U_{n'-n}(k\mathbf{r}_1-k\mathbf{r}_2)\langle f_{n'}\rangle(\mathbf{r}_1,\lambda_1)g(\lvert\mathbf{r}_1-\mathbf{r}_2\rvert,\lambda_1\lambda_2)\,\mathrm{d}r_2\mathrm{d}\lambda_2 \tag*{(21)}

where we also used the simpler pair-correlation (14).

In the next section, we use the effective wave method to solve (21). This method introduced in [10] proved to be successful in 3D and provides a closed formula for ⟨fn⟩(r1,λ1)\langle f_n\rangle(\mathbf{r}_1,\lambda_1). This in turn allows us to compute ⟨Fn⟩\langle\mathcal{F}_n\rangle (20) and reach an explicit formula for Tn,NT_{n,N} by specifying gn=δn,Ng_n=\delta_{n,N} in (18).

The effective wave method

To solve the general integral governing equation (21), we use the effective waves method [10] as summarised below.

Overview of the effective waves method

The starting point is to assume that there exists k⋆∈Ck_{\star}\in\mathbb{C} such that

(Δ+k⋆2)⟨fn⟩(r1,λ1)=0.(22)\left(\Delta+k_{\star}^{2}\right)\langle f_n\rangle(\mathbf{r}_1,\lambda_1)=0. \tag*{(22)}

This assumption will greatly simplify the governing integral equation (21), from which we will be able to determine both k⋆k_{\star} and ⟨fn,N⟩(r1,λ1)\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1). To summarise, the method has three steps:

  1. Separate microstructure and boundary. The assumption (22) is used in the governing equation (21) to derive two separate equations called the ensemble wave equation and the ensemble boundary conditions. The first one only depends on the microstructure of the random material while, in contrast, the second one takes into account the incident field and the shape of the random material, acting much like a boundary condition.

  1. The effective eigensystem. We decompose ⟨fn,N⟩(r1,λ1)\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1) in the basis of functions Vn(k⋆r1)V_n(k_{\star}\mathbf{r}_1) and substitute the decomposition into the ensemble wave equation. The coefficients F(λ1)F(\lambda_1) of the decomposition are then shown to be the eigenfunctions of an eigensystem which eigenvalue is k⋆k_{\star}.

  1. The ensemble boundary condition. To determine the amplitudes of the eigenfunctions F(λ1)F(\lambda_1) we then use the ensemble boundary condition, which takes into account the incident field and the shape of the random material.

Separate microstructure and boundary

We follow the steps described above to determine the solutions ⟨fn⟩(r1,λ1)\langle f_n\rangle(\mathbf{r}_1,\lambda_1) of (21). The first step uses (22), and some algebraic manipulations shown in Appendix § 10.1 to rewrite the governing equation (21) into two separate equations:

⟨fn⟩(r1,λ1)+∑n′∈ZTn(λ1)∫S[Jn′n(r1)k2−k⋆2−Kn′n(r1)]n(λ2) dλ2=0(ensemble wave equation)(23)\langle f_n\rangle(\mathbf{r}_1,\lambda_1)+\sum_{n'\in\mathbb{Z}}T_n(\lambda_1)\int_S\left[\frac{\mathcal{J}_{n'n}(\mathbf{r}_1)}{k^2-k_\star^2}-\mathcal{K}_{n'n}(\mathbf{r}_1)\right]n(\lambda_2)\,\mathrm{d}\lambda_2=0\quad\text{(ensemble wave equation)} \tag*{(23)}
∑n′Vn′−n(kr1)gn′+∑n′∫SIn′n(r1)k2−k⋆2n(λ2) dλ2=0(ensemble boundary condition).(24)\sum_{n'}V_{n'-n}(k r_1)g_{n'}+\sum_{n'}\int_S\frac{\mathcal{I}_{n'n}(\mathbf{r}_1)}{k^2-k_\star^2}n(\lambda_2)\,\mathrm{d}\lambda_2=0\quad\text{(ensemble boundary condition)}. \tag*{(24)}

The terms Jn′n(r1)\mathcal{J}_{n'n}(\mathbf{r}_1), In′n(r1)\mathcal{I}_{n'n}(\mathbf{r}_1), and Kn′n(r1)\mathcal{K}_{n'n}(\mathbf{r}_1), respectively defined in (C.6) and (C.3), involve the function ⟨fn,N⟩(r1,λ1)\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1).

One of the key advantages of splitting the integral equation (21) into the two separate equations above is that the ensemble wave equation (23) does not depend on the shape of the region RR nor the incident wave. As we will see below, the ensemble wave equation (23) can be used to determine the effective wavenumber k⋆k_\star, which implies that k⋆k_\star only depends on the microstructure: the density of the particles n(λ1)n(\lambda_1), their properties provided by the T-matrix Tn(λ1)T_n(\lambda_1), and the pair correlation gg which explicitly appears in the quantity Kn,n′(r1)\mathcal{K}_{n,n'}(\mathbf{r}_1) (C.3). We further discuss how to interpret k⋆k_\star in § 4.6.

On the other hand, the ensemble boundary condition (24) acts like a boundary condition, and shows how the incident wave and material boundary affect the overall solution.

The effective eigensystem

Since ⟨fn⟩(r1,λ1)\langle f_n\rangle(\mathbf{r}_1,\lambda_1) satisfies (22), it can be decomposed into the modes

⟨fn⟩(r1,λ1)=∑n1Fnn1(λ1)Vn1(k⋆r1),(25)\langle f_n\rangle(\mathbf{r}_1,\lambda_1)=\sum_{n_1}F_{nn_1}(\lambda_1)V_{n_1}(k_\star r_1), \tag*{(25)}

where Vn1V_{n_1} is defined in (4).

The unknowns k⋆k_\star and Fnn1(λ1)F_{nn_1}(\lambda_1) can be determined by substituting (25) into (23). The details are shown in Appendix § 10.§ 2 with the resulting equation being

Fnn1(λ1)+∑n′n2δn2−n1+n′−nTn(λ1)∫SNn′−n12(k,k⋆)Fn′n2(λ2)n(λ2) dλ2=0,(26)F_{nn_1}(\lambda_1)+\sum_{n'n_2}\delta_{n_2-n_1+n'-n}T_n(\lambda_1)\int_S\mathcal{N}^{12}_{n'-n}(k,k_\star)F_{n'n_2}(\lambda_2)n(\lambda_2)\,\mathrm{d}\lambda_2=0, \tag*{(26)}

where

Nl12(k,k⋆):=2πNl(ka12,k⋆a12)k⋆2−k2−2π∫a12b12Jl(k⋆r)Hl(kr)δg(r)r dr,(27)\mathcal{N}^{12}_{l}(k,k_\star):=2\pi\frac{N_l(k a_{12},k_\star a_{12})}{k_\star^2-k^2}-2\pi\int_{a_{12}}^{b_{12}}J_l(k_\star r)H_l(k r)\delta g(r)r\,\mathrm{d}r, \tag*{(27)}
Nl(x,y):=xHl′(x)Jl(y)−yHl(x)Jl′(y).(28)N_l(x,y):=xH'_l(x)J_l(y)-yH_l(x)J'_l(y). \tag*{(28)}

The above is a nonlinear eigenvalue problem, which is why we refer to k⋆k_\star as an eigenvalue. After calculating k⋆k_\star we can calculate the eigenfunctions Fn′n2(λ1)F_{n'n_2}(\lambda_1) by solving the linear system (26), though in practice it is far better to calculate both k⋆k_\star and Fn′n2(λ1)F_{n'n_2}(\lambda_1) using the modal decomposition as we do in § 4.

The ensemble boundary condition

The eigensystem (26) is not enough to fully determine the Fnn1F_{nn_1}, for instance if Fnn1F_{nn_1} is a solution to the eigensystem then so is αFnn1\alpha F_{nn_1} for any scalar α\alpha.

To fully determine Fnn1F_{nn_1} we need to substitute (25) into the ensemble boundary condition (24). The details are shown in Appendix § 11, where we deduce the ensemble boundary conditions (D.5) without specifying the region R2\mathcal{R}_2, followed by choosing R2\mathcal{R}_2 to be a cylinder with radius R−a2R-a_2 which results in the boundary condition:

gN+2πk⋆2−k2∑n′∫SFn′(N−n′)(λ2)NN−n′(kR2,k⋆R2)n(λ2) dλ2=0,(29)g_N+\frac{2\pi}{k_\star^2-k^2}\sum_{n'}\int_{\mathcal{S}}F_{n'(N-n')}(\lambda_2)N_{N-n'}(kR_2,k_\star R_2)n(\lambda_2)\,\mathrm{d}\lambda_2=0, \tag*{(29)}

where NlN_l is defined by (27).

An effective cylinder

Modal decomposition of the problem

In this section, we assume that the particles are confined in a circular disc R\mathcal{R} with radius RR. Since (21) is linear with respect to gng_n, it can be decomposed in simpler and independent equations: let ⟨fn,N⟩(r1,λ1)\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1) be the solution when substituting gn=δn,Ng_n=\delta_{n,N} in (21), then

⟨fn,N⟩(r1,λ1)=Tn(λ1)VN−n(kr1)+Tn(λ1)∑n′∫Sn(λ2)∫R2Un′−n(kr1−kr2)⟨fn′,N⟩(r1,λ1)g(∣r1−r2∣,λ1,λ2) dr2dλ2,(30)\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1)=T_n(\lambda_1)V_{N-n}(k\mathbf{r}_1)+T_n(\lambda_1)\sum_{n'}\int_{\mathcal{S}}n(\lambda_2)\int_{\mathcal{R}_2}U_{n'-n}(k\mathbf{r}_1-k\mathbf{r}_2)\langle f_{n',N}\rangle(\mathbf{r}_1,\lambda_1)g(\lvert\mathbf{r}_1-\mathbf{r}_2\rvert,\lambda_1,\lambda_2)\,\mathrm{d}\mathbf{r}_2\mathrm{d}\lambda_2, \tag*{(30)}

where we can recover the solution for any incident wave by using

⟨fn⟩(r1,λ1)=∑NgN⟨fn,N⟩(r1,λ1).(31)\langle f_n\rangle(\mathbf{r}_1,\lambda_1)=\sum_N g_N\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1). \tag*{(31)}

From the 3D version of these effective equations [10] we know that (26) can be reduced by using symmetry. We show how to do this for the modal decomposition below, with the result being:

⟨fn,N⟩(r1,λ1)=αNFn,N(λ1)VN−n(k⋆r1)(32)\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1)=\alpha_N F_{n,N}(\lambda_1)V_{N-n}(k_{\star}\mathbf{r}_1) \tag*{(32)}

where αN∈C\alpha_N \in\mathbb{C} is some amplitude which is introduced for later convenience. Below is the proof of (32).

Proof. Using the rotational symmetry of the modal source, we simplify the form of the modal solutions ⟨fn,N⟩(r1,λ1)\langle f_{n,N}\rangle(\mathbf{r}_1,\lambda_1). To this end, we denote by Rϕ\mathbf{R}_{\phi} the rotation matrix of angle ϕ\phi and replace r1\mathbf{r}_1 with Rϕr1\mathbf{R}_{\phi}\mathbf{r}_1 in (30):

⟨fn,N⟩(Rϕr1,λ1)=Tn(λ1)VN−n(kRϕr1)+Tn(λ1)∑n′∫Sn(λ2)∫R2Un′−n(kRϕr1−kRϕr2)⟨fn′,N⟩(Rϕr2,λ2)g(∣r1−r2∣) dr2dλ2.(33)\begin{aligned} \langle f_{n,N}\rangle(\mathbf{R}_{\phi}\mathbf{r}_1,\lambda_1) ={}&T_n(\lambda_1)V_{N-n}(k\mathbf{R}_{\phi}\mathbf{r}_1)\\ &+T_n(\lambda_1)\sum_{n'}\int_S n(\lambda_2)\int_{\mathcal{R}_2}U_{n'-n}(k\mathbf{R}_{\phi}\mathbf{r}_1-k\mathbf{R}_{\phi}\mathbf{r}_2)\\ &\qquad\qquad\langle f_{n',N}\rangle(\mathbf{R}_{\phi}\mathbf{r}_2,\lambda_2)g(\lvert\mathbf{r}_1-\mathbf{r}_2\rvert)\,\mathrm{d}\mathbf{r}_2\mathrm{d}\lambda_2. \tag*{(33)} \end{aligned}

where we changed the integration variable from r2\mathbf{r}_2 to Rϕr2\mathbf{R}_{\phi}\mathbf{r}_2, which is possible for any rotation as R2\mathcal{R}_2 is a disc. Then from (4) we deduce the property

Un(Rϕr1)=Un(r1)einϕandVn(Rϕr1)=Vn(r1)einϕ,U_n(\mathbf{R}_{\phi}\mathbf{r}_1)=U_n(\mathbf{r}_1)e^{in\phi}\quad\text{and}\quad V_n(\mathbf{R}_{\phi}\mathbf{r}_1)=V_n(\mathbf{r}_1)e^{in\phi},

which used in (33), and then multiplying both sides of the equation with e−i(N−n)ϕe^{-i(N-n)\phi} leads to

⟨fn,N⟩(Rϕr1,λ1)e−i(N−n)ϕ=Tn(λ1)VN−n(kr1)+Tn(λ1)∑n′∫Sn(λ2)∫R2Un′−n(kr1−kr2)e−i(N−n′)ϕ⟨fn′,N⟩(Rϕr2,λ2)g(∣r1−r2∣) dr2dλ2.(34)\begin{aligned} \langle f_{n,N}\rangle(\mathbf{R}_{\phi}\mathbf{r}_1,\lambda_1)e^{-i(N-n)\phi} ={}&T_n(\lambda_1)V_{N-n}(k\mathbf{r}_1)\\ &+T_n(\lambda_1)\sum_{n'}\int_S n(\lambda_2)\int_{\mathcal{R}_2}U_{n'-n}(k\mathbf{r}_1-k\mathbf{r}_2)e^{-i(N-n')\phi}\\ &\qquad\qquad\langle f_{n',N}\rangle(\mathbf{R}_{\phi}\mathbf{r}_2,\lambda_2)g(\lvert\mathbf{r}_1-\mathbf{r}_2\rvert)\,\mathrm{d}\mathbf{r}_2\mathrm{d}\lambda_2. \tag*{(34)} \end{aligned}

Now note that both ⟨fn,N⟩(Rϕr1,λ1)e−i(N−n)ϕ\langle f_{n,N}\rangle(\mathbf{R}_{\phi}\mathbf{r}_1,\lambda_1)e^{-i(N-n)\phi} and ⟨fn,N⟩(r1)\langle f_{n,N}\rangle(\mathbf{r}_1) solve exactly the same integral equation. So by assuming uniqueness, i.e. that there is only one solution to the above, we conclude that

⟨fn,N⟩(r1)=⟨fn,N⟩(Rϕr1)e−i(N−n)ϕ,(35)\langle f_{n,N}\rangle(\mathbf{r}_1)=\langle f_{n,N}\rangle(\mathbf{R}_{\phi}\mathbf{r}_1)e^{-i(N-n)\phi}, \tag*{(35)}

for any r1\mathbf{r}_1 and ϕ\phi. Let (r1,θ1)(r_1,\theta_1) be the polar coordinates of r1\mathbf{r}_1, then, without lose of generality, we then choose ϕ=−θ1\phi=-\theta_1 which leads to

⟨fn,N⟩(r1,θ1,λ1)=⟨fn,N⟩(r1,0,λ1)ei(N−n)θ1.(36)\langle f_{n,N}\rangle(r_1,\theta_1,\lambda_1)=\langle f_{n,N}\rangle(r_1,0,\lambda_1)e^{i(N-n)\theta_1}. \tag*{(36)}

Finally, because ⟨fn,N⟩(r1)\langle f_{n,N}\rangle(\mathbf{r}_1) satisfies a wave equation (22) it can be written in a series of the form (25). Using the symmetry above we see that only one term in this series remains shown by (32).

The modal dispersion equation

We can deduce a simpler effective eigensystem and dispersion equation by using the symmetry (32).

To start we substitute the expansions (C.8) and (32) into the modal decomposition (31) to obtain

∑n1Fnn1(λ1)Vn1(k⋆r1)=∑N1αN1gN1Fn,N1(λ1)VN1−n(k⋆r1)(37)\sum_{n_1}F_{nn_1}(\lambda_1)V_{n_1}(k_{\star}\mathbf{r}_1)=\sum_{N_1}\alpha_{N_1}g_{N_1}F_{n,N_1}(\lambda_1)V_{N_1-n}(k_{\star}\mathbf{r}_1) \tag*{(37)}

Then since VnV_n form an orthogonal basis of functions,

Fnn1=∑N1δN1−n1−nαN1gN1Fn,N1.(38)F_{nn_1}=\sum_{N_1}\delta_{N_1-n_1-n}\alpha_{N_1}g_{N_1}F_{n,N_1}. \tag*{(38)}

To simplify the effective eigensystem (26) we consider one mode at a time by taking gN1=δN1−Ng_{N_1}=\delta_{N_1-N}, which used in (38) implies that we can substitute Fnn1=δN−n1−nαNFn,NF_{nn_1}=\delta_{N-n_1-n}\alpha_NF_{n,N} into (26). The result after some algebraic manipulations is

Fn,N(λ1)+∑n′Tn(λ1)∫SNn′−n12(k,k⋆)Fn′,N(λ2)n(λ2) dλ2=0.(39)F_{n,N}(\lambda_1)+\sum_{n'}\mathrm{T}_n(\lambda_1)\int_S\mathcal{N}^{12}_{n'-n}(k,k_\star)F_{n',N}(\lambda_2)n(\lambda_2)\,\mathrm{d}\lambda_2=0. \tag*{(39)}

The above equation is identical to the case of the eigensystem for plane waves [12], and matches also the eigensystems for a single type of particle [11, 19] when taking n(λ)=δ(λ−λ1)n(\lambda)=\delta(\lambda-\lambda_1). This result is somewhat expected as the ensemble wave equation (23) does not depend on the incident wave and material geometry, which also explains why the modal index NN only appears in Fn,NF_{n,N} in the above.

Instead of solving (26), it is far simpler to solve the above, and then write the general solution in the modal form (38). In practice, to solve (39), we can discretise the integral over SS as a set of reals {t1,…,tS}\{t_1,\ldots,t_S\}. Then define a block vector F\mathbf{F} containing the entries Fn(ts)F_n(t_s) for n=−M,−M+1,…,M−1,Mn=-M,-M+1,\ldots,M-1,M, for some finite MM, and for s=1,…,Ss=1,\ldots,S, so that the eigensystem becomes

(I+M)⋅F=0(40)(\mathbf{I}+\mathbf{M})\cdot\mathbf{F}=0 \tag*{(40)}

where

Mnn′,ss′(k⋆)=Tn(ts)Nn′−n12(k,k⋆)n(ts′).(41)M_{nn',ss'}(k_\star)=\mathrm{T}_n(t_s)\mathcal{N}^{12}_{n'-n}(k,k_\star)n(t_{s'}). \tag*{(41)}

The parameter k⋆k_\star is then obtained by solving the equation

det⁡[I+M(k⋆)]=0.(42)\det[\mathbf{I}+\mathbf{M}(k_\star)]=0. \tag*{(42)}

Multiple wavenumbers

The dispersion equation (39) has infinitely many eigenvalues kpk_p with p=1,2,…p=1,2,\ldots. Consequently, ⟨fn⟩(r1)\langle f_n\rangle(\mathbf{r}_1) can more generally be written as a sum over all the eigenvalues kpk_p and their corresponding eigenfunctions [8], which leads to more accurate solutions. However, only a small difference in comparison to using just the eigenvalue k⋆k_\star with the smallest imaginary part is observed (cf. [11, 8, 10] for details). For this reason, and for simplicity, we only account for the one wavenumber k⋆k_\star in this paper. See Figure 5 for a typical distribution of the many eigenvalues of (39).

Image not extracted

Figure 5. Shows 11 of the eigenvalues kpk_p of (42) for sound soft particles of radius a=1a=1 confined in a cylinder of radius R=10R=10 when ω=1\omega=1. k⋆k_\star corresponds to the one with the smallest imaginary part. The xx-axis is the real part and yy-axis is the imaginary part. The other wavenumbers have a much larger imaginary part leading to evanescent waves inside the random material.

The modal ensemble boundary condition

To determine the αN\alpha_N that appear in (38) we need to use the ensemble boundary condition. The simplest way to do this is again to take gN1=δN1−Ng_{N_1}=\delta_{N_1-N} in (38) and then substitute the result into (29) to obtain

1+2παNk⋆2−k2∑n′∫SFn′,N(λ2)NN−n′(kR2,k⋆R2)n(λ2) dλ2=0.(43)1+\frac{2\pi\alpha_N}{k_\star^2-k^2}\sum_{n'}\int_S F_{n',N}(\lambda_2)\mathcal{N}_{N-n'}(kR_2,k_\star R_2)n(\lambda_2)\,\mathrm{d}\lambda_2=0. \tag*{(43)}

which we can use to determine:

αN=−k⋆2−k22π(∑n′∫SFn′,N(λ2)NN−n′(kR2,k⋆R2)n(λ2) dλ2)−1.(44)\alpha_N=-\frac{k_\star^2-k^2}{2\pi}\left(\sum_{n'}\int_S F_{n',N}(\lambda_2)\mathcal{N}_{N-n'}(kR_2,k_\star R_2)n(\lambda_2)\,\mathrm{d}\lambda_2\right)^{-1}. \tag*{(44)}

The effective T-matrix

As discussed § 4, the effective T-matrix Tn,N\mathcal{T}_{n,N} can easily describe the average scattered wave for any incident wave through:

⟨Fn⟩=∑N=−∞+∞Tn,NgN,(45)\langle\mathcal{F}_n\rangle= \sum_{N=-\infty}^{+\infty} \mathcal{T}_{n,N}g_N, \tag*{(45)}

where the ⟨Fn⟩\langle\mathcal{F}_n\rangle, given by (20), are the average coefficients of the waves scattered from the whole cylinder R\mathcal{R}. Note the above holds for any choice of gNg_N and Tn,N\mathcal{T}_{n,N} does not depend on gNg_N.

To calculate Tn,N\mathcal{T}_{n,N} we substitute the modal decomposition (31) into (20) to obtain

⟨Fn⟩=∑NgN∫Sn(λ1)∫R1∑n′Vn′−n(−kr1)⟨fn′,N′⟩(r1,λ1) dr1dλ1.(46)\langle\mathcal{F}_n\rangle= \sum_N g_N \int_S n(\lambda_1)\int_{\mathcal{R}_1}\sum_{n'}\mathrm{V}_{n'-n}(-k\mathbf{r}_1)\langle f_{n',N}'\rangle(\mathbf{r}_1,\lambda_1)\,\mathrm{d}\mathbf{r}_1\mathrm{d}\lambda_1. \tag*{(46)}

Comparing the above with the definition of the effective T-matrix (45), it is clear that

Tn,N=∫Sn(λ1)∫R1∑n′Vn′−n(−kr1)⟨fn′,N′⟩(r1,λ1) dr1dλ1.(47)\mathcal{T}_{n,N} = \int_S n(\lambda_1)\int_{\mathcal{R}_1}\sum_{n'}\mathrm{V}_{n'-n}(-k\mathbf{r}_1)\langle f_{n',N}'\rangle(\mathbf{r}_1,\lambda_1)\,\mathrm{d}\mathbf{r}_1\mathrm{d}\lambda_1. \tag*{(47)}

To calculate the above, we follow the same steps shown in Appendix § 10. Specifically, use use Green’s second identity (C.4), the regular expansion (32), and the orthogonality of the Vn\mathrm{V}_n functions to conclude that

Tn,N=δN−n2παnk⋆2−k2∫Sn(λ1)∑n′Fn′,n(λ1)Qn−n′(kR1,k⋆R1) dλ1(48)\mathcal{T}_{n,N}=\delta_{N-n}\frac{2\pi\alpha_n}{k_\star^2-k^2}\int_S n(\lambda_1)\sum_{n'}F_{n',n}(\lambda_1)Q_{n-n'}(kR_1,k_\star R_1)\,\mathrm{d}\lambda_1 \tag*{(48)}

where R1R_1 is the radius of the disc R1\mathcal{R}_1 and we defined:

Ql(x,y):=xJl′(x)Jl(y)−yJl(x)Jl′(y).(49)Q_l(x,y):=xJ_l'(x)J_l(y)-yJ_l(x)J_l'(y). \tag*{(49)}

Finally, we substitute αn\alpha_n given by (44), which results in

Tn,N=−δN−n∫Sn(λ1)∑n′Fn′,n(λ1)Qn−n′(kR1,k⋆R1) dλ1∫Sn(λ1)∑n′Fn′,n(λ1)Nn−n′(kR1,k⋆R1) dλ1.(50)\mathcal{T}_{n,N}=-\delta_{N-n}\frac{\int_S n(\lambda_1)\sum_{n'}F_{n',n}(\lambda_1)Q_{n-n'}(kR_1,k_\star R_1)\,\mathrm{d}\lambda_1}{\int_S n(\lambda_1)\sum_{n'}F_{n',n}(\lambda_1)N_{n-n'}(kR_1,k_\star R_1)\,\mathrm{d}\lambda_1}. \tag*{(50)}

Note that once the effective T-matrix is known, the scattering from any incident field can be computed with (19) after decomposing the incident in modes (5). For example, in Figure 6, we have plotted the total pressure field u:=uinc+uscu:=u_{\mathrm{inc}}+u_{\mathrm{sc}} resulting from an incident plane wave and a point source.

Image not extracted

Figure 6. Average field for two different sources. Average pressure field when the incident field is a plane wave (left) and a point source (right). The material is made of sound-hard particles of radius one confined in a disc of radius 10.

Monopole particles only

The effective T-matrix (50) resembles the T-matrix for a homogeneous cylinder, see for example (8). In fact, it is a weighted average of the factors of a homogeneous T-matrix, as explained in the introduction. From this observation we see that if the particles are monopole scatterers we obtain a significant simplification.

Let us assume here that the particles scatter only monopole waves, in which case the scattered field (6) becomes

usc(r)=∑i=1Jf0iH0(kr−kri),(51)u_{\mathrm{sc}}(\mathbf{r})=\sum_{i=1}^{J}f_0^iH_0(k\mathbf{r}-k\mathbf{r}_i), \tag*{(51)}

which leads to ⟨fn,N⟩(r1)=0\langle f_{n,N}\rangle(r_1)=0 if n≠0n\ne0 and, as a result of (32), Fn,N=0F_{n,N}=0 if n≠0n\ne0. Substituting this result into the effective T-matrix (50), and assuming that the radius R1R_1 of the region R1\mathcal{R}_1 is the same for every type of particle λ1\lambda_1, we obtain:

TnM=−kJn′(kR~)Jn(k⋆R~)−k⋆Jn(kR~)Jn′(k⋆R~)kHn′(kR~)Jn(k⋆R~)−k⋆Hn(kR~)Jn′(k⋆R~)(monopole scatterers).(52)\mathcal{T}^{\mathrm{M}}_n=-\frac{kJ'_n(k\widetilde{R})J_n(k_{\star}\widetilde{R})-k_{\star}J_n(k\widetilde{R})J'_n(k_{\star}\widetilde{R})}{kH'_n(k\widetilde{R})J_n(k_{\star}\widetilde{R})-k_{\star}H_n(k\widetilde{R})J'_n(k_{\star}\widetilde{R})}\quad\text{(monopole scatterers).} \tag*{(52)}

This corresponds to the T-matrix of a homogeneous cylinder of radius R~=R−a\widetilde{R}=R-a, sound speed c⋆=ω/k⋆c_{\star}=\omega/k_{\star} and mass density ρ⋆=ρ\rho_{\star}=\rho where ρ\rho is the density of the host medium (cf. (8)).

Numerical results

The numerical results obtained in this paper use the open-source package EffectiveTMatrix.jl written in Julia [23]. More details on the simulations below can be found on the GitHub page https://github.com/Kevish-Napal/EffectiveTMatrix.jl.

Optimized Monte Carlo simulations

We use a Monte Carlo method to validate our theoretical results (50) and (52). To develop an efficient Monte Carlo method we rely on the following symmetry of the modes:

uinc(r)=JN(kr)eiNθ⟹⟨usc⟩(r)=TNHN(kr)eiNθ.(53)u_{\mathrm{inc}}(r)=J_N(kr)e^{iN\theta}\quad\Longrightarrow\quad\langle u_{\mathrm{sc}}\rangle(r)=T_NH_N(kr)e^{iN\theta}. \tag*{(53)}

This result is easily obtained from (19) with the specific choice gn=δn,Ng_n=\delta_{n,N}, which substituted into (18) leads

TN=⟨FN⟩.(54)T_N=\langle\mathcal{F}_N\rangle. \tag*{(54)}

In other words, TNT_N can be numerically estimated by simulating the waves scattered from one particle configuration at a time by using (7), and then taking the average of FN\mathcal{F}_N (16) over many different particles configurations.

To illustrate the efficiency of this Monte Carlo method we compare it with another method which directly computes ⟨usc⟩\langle u_{\mathrm{sc}}\rangle, which has been commonly used in the literature [16]. For this second method we use (6) to compute ⟨usc⟩(R,0)\langle u_{\mathrm{sc}}\rangle(R,0). Then from (53) we can also compute TNT_N with

TN=HN−1(kR)⟨usc⟩(R,0).(55)T_N=H_N^{-1}(kR)\langle u_{\mathrm{sc}}\rangle(R,0). \tag*{(55)}

The two methods (54) and (55) are compared in Figure 7. The standard deviation of the mean of the second method is larger than the first one, resulting in a slower convergence. The reason is that (55), in contrary to (54), includes all the modes of each scattered field computed for a specific particles configuration:

Image not extracted

Figure 7. Computation of T0T_0 using the two different methods (54) (modal) and (55) (naive). We computed 5000 realisations of these quantities for different configurations of particles, the plots on the top and bottom respectively correspond to the real and imaginary parts of the results. The distribution of the results is reported on the histograms on the left. The plots on the right correspond to the cumulative average of the realisations. Both methods converge to the same limit, however, the modal method converges faster and presents a lower standard deviation of the mean.

⟨usc⟩(R,0)=⟨∑nFnHn(kR)⟩.(56)\langle u_{\mathrm{sc}\rangle}(R,0)=\left\langle\sum_n\mathcal{F}_nH_n(kR)\right\rangle. \tag*{(56)}

While the terms n≠Nn\ne N of the sum vanish on average, they significantly contribute to the standard deviation of the mean in (55).

Validation of the Effective Waves Method

We validate the Effective Waves method (47) against the Monte Carlo method (54) for several frequencies ω\omega ranging in Ω:=[0.05:0.015:1.5]\Omega:= [0.05 : 0.015 : 1.5]. To this end we define the relative error, averaged over frequencies:

ϵn:=1∣Ω∣∑ω∈Ω∣TnMC(ω)−TnEWM(ω)∣∣TnMC(ω)∣.(57)\epsilon_n:=\frac{1}{|\Omega|}\sum_{\omega\in\Omega}\frac{\left|T_n^{\mathrm{MC}}(\omega)-T_n^{\mathrm{EWM}}(\omega)\right|}{\left|T_n^{\mathrm{MC}}(\omega)\right|}. \tag*{(57)}

where TnMC(ω)T_n^{\mathrm{MC}}(\omega) is obtained following the Monte Carlo method (54) and TnEWM(ω)T_n^{\mathrm{EWM}}(\omega) following the Effective Waves method (47). A few plots of TnMC(ω)T_n^{\mathrm{MC}}(\omega) and TnEWM(ω)T_n^{\mathrm{EWM}}(\omega) are provided in Figure 2 and Figure 8. The values of ϵ0\epsilon_0, ϵ1\epsilon_1, ϵ2\epsilon_2, ϵ3\epsilon_3, ϵ4\epsilon_4 for the cases of sound soft and sound hard particles are reported in Table 1.

Image not extracted

Figure 8. Figure 8: Compares various methods to calculate the components T1T_1, T2T_2, T3T_3 and T4T_4 of the T-matrix of a cylinder filled with sound soft particles. The solid red line is our effective waves method (EWM) (50), the black points are from the Monte Carlo method (MC) (54), and the dashed blue line is our method when only monopole scattering is accounted for (EWM-MA) (52). The general expression of the effective T-matrix matches the Monte Carlo results. The EWM-MA method only matches well with the Monte Carlo for low frequencies.

ε0\varepsilon_0ε1\varepsilon_1ε2\varepsilon_2ε3\varepsilon_3ε4\varepsilon_4
sound soft particles6.75e−26.75\mathrm{e}{-2}4.59e−24.59\mathrm{e}{-2}5.40e−25.40\mathrm{e}{-2}6.45e−26.45\mathrm{e}{-2}7.45e−27.45\mathrm{e}{-2}
sound hard particles2.66e−22.66\mathrm{e}{-2}2.43e−22.43\mathrm{e}{-2}2.33e−22.33\mathrm{e}{-2}2.35e−22.35\mathrm{e}{-2}2.37e−22.37\mathrm{e}{-2}

Table 1. Relative errors εn\varepsilon_n defined by (57) in the cases of sound soft and sound hard. The particles are of radius 1 and confined in a circular area of radius 20. The volume fraction is 0.05 which corresponds to 19.7 particles on average.

Conclusion

Main goal. Our main goal was to describe how an incident wave is scattered from a cylinder filled with smaller cylinders, which we have called particles, that are placed in a disordered but correlated way. We describe this correlation through the inter-particle pair correlation, see (11). The literature so far has focused on plane waves scattered from a halfspace or plate filled with a particulate [14]. There has been at least one paper on solving this scenario, but used an ad hoc method, whereas here everything is deduced from first principles making only two assumptions: such as the Quasi-Crystalline Approximation (QCA) [17], and expressing the average field as a sum of effective waves, which has been shown to be the analytic solution [8]. One of the key advantages of describing the scattering from a cylinder with 2D particles is that it is far easier to validate this scenario with direct numerical simulations. Validation is still necessary as the theory requires the use of QCA whose validity is not clearly established. Further, validating this

ε0\varepsilon_0ε1\varepsilon_1ε2\varepsilon_2ε3\varepsilon_3ε4\varepsilon_4
sound soft particles6.75e−26.75\mathrm{e}{-2}4.59e−24.59\mathrm{e}{-2}5.40e−25.40\mathrm{e}{-2}6.45e−26.45\mathrm{e}{-2}7.45e−27.45\mathrm{e}{-2}
sound hard particles2.66e−22.66\mathrm{e}{-2}2.43e−22.43\mathrm{e}{-2}2.33e−22.33\mathrm{e}{-2}2.35e−22.35\mathrm{e}{-2}2.37e−22.37\mathrm{e}{-2}

scenario also serves to validate the predicted effective wavenumbers from any material geometry [10].

Modal scattering. A method that allowed us to greatly simplify both the theoretical and Monte Carlo calculations which we use to validate this work, was to make use of all symmetries present. We achieved this by solving for each polar mode of the incident wave separately. This simple, but effective technique, leads us to an effective T-matrix given by equation (50) that can be used to calculate the average scattered wave from any incident wave, see § 3 for a brief overview. We note that we were able to describe the scattered field without calculating the average transmitted field. In future work, this may be interesting to do, for example, to clearly identify when different effective wavenumbers are excited [14].

Effective T-matrix. The result of our theoretical work is summarised by the T-matrix (50). Beyond using just QCA, to calculate this T-matrix, we also assumed that only one effective wavenumber k⋆k_\star is excited. This is true for a wide range of parameters but it is not always the case. In particular it appears that very strong scattering at moderate frequencies can trigger more than one effective wavenumber to be excited [14, 11, 10]. One possible extension to our work is to include the effect of more than one effective wavenumber.

Monopole Scatterers. One surprising result, is that if the particles only scatter monopole waves, then the effective T-matrix greatly simplifies and becomes (52). This form is exactly the same as the T-matrix for a homogenous cylinder, one with constant material parameters. We hypothesise that any material filled with monopole scatterers would, on average, respond like a homogeneous material. Monopole scatterers are a good approximation for many types of resonant particles [4]. Figure 8 compares the results for the monopole scattering approximation with Monte Carlo results for sound-soft particles, which does not assume the particles scatter like monopoles.

Monte Carlo. Beyond deducing an effective T-matrix for a particulate cylinder, we also developed an efficient Monte-Carlo method, which matched our theoretical predictions very accurately see Figure 2 and Figure 8. Our numerical validation was for a broad frequency range, but we did not cover a broad range of parameters. Doing this would be valuable future work, and could help clearly identify the limits of QCA and different approximations for the inter-particle pair-correlation.

A prototype for new materials. The setting we deduce is the ideal case to test new disordered particulates. That is, to use exotic inter-particle pair-correlation to achieve effects such as band-gaps, or impedance matching. The scenario of a cylinder filled with 2D particles is ideal for testing new types of particles, and inter-particle pair-correlations, because they can be easily validated with Monte-Carlo methods. When designing new materials with exotic responses, and stretching the limits of the theory, we need to have a way to validate those predictions. And this paper provides that.

Appendices

Bessel functions and translation matrices

Given two points x,y∈R2\boldsymbol{x},\boldsymbol{y}\in\mathbb{R}^{2}, we have the following identities where d=x−y\boldsymbol{d}=\boldsymbol{x}-\boldsymbol{y}

i)Vn(y)=∑n′=−∞+∞Vn−n′(d)Vn′(x),for all x,d∈R2ii)Un(y)=∑n′=−∞+∞Vn−n′(d)Un′(x),for all ∣x∣>∣d∣iii)Un(y)=∑n′=−∞+∞Un−n′(d)Vn′(x),for all ∣x∣<∣d∣.(58)\begin{aligned} \mathrm{i)}\quad V_n(\boldsymbol{y}) &= \sum_{n'=-\infty}^{+\infty} V_{n-n'}(\boldsymbol{d})V_{n'}(\boldsymbol{x}), \quad\text{for all }\boldsymbol{x},\boldsymbol{d}\in\mathbb{R}^{2}\\ \mathrm{ii)}\quad U_n(\boldsymbol{y}) &= \sum_{n'=-\infty}^{+\infty} V_{n-n'}(\boldsymbol{d})U_{n'}(\boldsymbol{x}), \quad\text{for all }|\boldsymbol{x}|>|\boldsymbol{d}|\\ \mathrm{iii)}\quad U_n(\boldsymbol{y}) &= \sum_{n'=-\infty}^{+\infty} U_{n-n'}(\boldsymbol{d})V_{n'}(\boldsymbol{x}), \quad\text{for all }|\boldsymbol{x}|<|\boldsymbol{d}|. \tag*{(58)} \end{aligned}

The above formulas are direct consequences of Graf’s theorem (see [21, Th. 2.11-2.12] for instance).

Ensemble averaging

Here we give a brief overview of ensemble averaging so that this paper is more self-contained. For more details, see [10, 21] and the references within. To simplify computations, we represent one particle configurations with:

Λ=r1,λ1…rJ,λJ,Λ(1)=r2,λ2…rJ,λJ,\Lambda=\boldsymbol{r}_{1},\lambda_{1}\ldots\boldsymbol{r}_{J},\lambda_{J},\qquad\Lambda^{(1)}=\boldsymbol{r}_{2},\lambda_{2}\ldots\boldsymbol{r}_{J},\lambda_{J},
Λ(1,j)=r2,λ2,…,rj−1,λj−1,rj+1,λj+1,…,rJ,λJ.\Lambda^{(1,j)}=\boldsymbol{r}_{2},\lambda_{2},\ldots,\boldsymbol{r}_{j-1},\lambda_{j-1},\boldsymbol{r}_{j+1},\lambda_{j+1},\ldots,\boldsymbol{r}_{J},\lambda_{J}.

Using these definitions, we define the ensemble average, and conditional ensemble averages, of a quantity A, which can depend on the positions and properties of all the particles, as

⟨A⟩:=∫A(Λ) p(Λ) dΛ⟨A⟩(ri,λi):=∫A(Λ) p(Λ(i)∣riλi) dΛ(i),(59)\langle A\rangle:= \int A(\Lambda)\,p(\Lambda)\,\mathrm{d}\Lambda\qquad\langle A\rangle(\boldsymbol{r}_{i},\lambda_{i}) := \int A(\Lambda)\,p(\Lambda^{(i)}|\boldsymbol{r}_{i}\lambda_{i})\,\mathrm{d}\Lambda^{(i)}, \tag*{(59)}
⟨A⟩(ri,λi;rj,λj):=∫A(Λ) p(Λ(i,j)∣ri,λi;rj,λj) dΛ(i,j),(60)\langle A\rangle(\boldsymbol{r}_{i},\lambda_{i};\boldsymbol{r}_{j},\lambda_{j}) := \int A(\Lambda)\,p(\Lambda^{(i,j)}|\boldsymbol{r}_{i},\lambda_{i};\boldsymbol{r}_{j},\lambda_{j})\,\mathrm{d}\Lambda^{(i,j)}, \tag*{(60)}

where the domain of integration for Λ\Lambda is over all possible particles positions and properties. The term ⟨A⟩(ri,λi;rj,λj)\langle A\rangle(\boldsymbol{r}_{i},\lambda_{i};\boldsymbol{r}_{j},\lambda_{j}) is the ensemble average of A conditional to (r1,λ1,r2,λ2)(\boldsymbol{r}_{1},\lambda_{1},\boldsymbol{r}_{2},\lambda_{2}).

Taking an ensemble average (B.1) on both sides of (15) leads to

⟨Fn⟩=∑i=1J∑n′=−∞+∞∫(−1)n−n′Vn−n′(kri)⟨fn′i⟩(ri,λi)p(ri,λi) dridλi=J∑n′=−∞+∞∫(−1)n−n′Vn−n′(kr1)⟨fn′1⟩(r1,λ1)p(r1,λ1) dr1dλ1(61)\begin{aligned} \langle F_n\rangle&= \sum_{i=1}^{J}\sum_{n'=-\infty}^{+\infty}\int(-1)^{n-n'}V_{n-n'}(k\boldsymbol{r}_{i})\langle f_{n'}^{i}\rangle(\boldsymbol{r}_{i},\lambda_{i})p(\boldsymbol{r}_{i},\lambda_{i})\,\mathrm{d}\boldsymbol{r}_{i}\mathrm{d}\lambda_{i}\\ &=J\sum_{n'=-\infty}^{+\infty}\int(-1)^{n-n'}V_{n-n'}(k\boldsymbol{r}_{1})\langle f_{n'}^{1}\rangle(\boldsymbol{r}_{1},\lambda_{1})p(\boldsymbol{r}_{1},\lambda_{1})\,\mathrm{d}\boldsymbol{r}_{1}\mathrm{d}\lambda_{1} \tag*{(61)} \end{aligned}

where we used the definition of conditional probability: p(Λ)=p(ri,λi)p(Λ(i))p(\Lambda)=p(\boldsymbol{r}_{i},\lambda_{i})p(\Lambda^{(i)}), the definition of conditional average (B.2) to introduce the term ⟨fn′i⟩(ri,λi)\langle f_{n'}^{i}\rangle(\boldsymbol{r}_{i},\lambda_{i}), and then used that particles are indistinguishable3. Finally, using (10) and (12) leads to formula (2.16).

Average governing equation

Here we briefly show how to reach the averaging governing equation by using just one assumption, the Quasi-Crystalline Approximation (QCA). For more details see [10].

Taking the conditional average (B.1)2_2 of (7) with i=1i=1 gives

⟨fn1⟩(r1,λ1)=Tn(λ1)∑n′Vn′−n(kr1)gn′+Tn(λ1)∑j≠1∑n′∫Un′−n(kr1−krj)fn′j(r1,λ1)p(Λ(1)∣r1,λ1) dΛ(1).(62)\left\langle f_n^1\right\rangle(\mathbf{r}_1,\lambda_1)=T_n(\lambda_1)\sum_{n'}V_{n'-n}(k\mathbf{r}_1)g_{n'}+T_n(\lambda_1)\sum_{j\ne1}\sum_{n'}\int U_{n'-n}(k\mathbf{r}_1-k\mathbf{r}_j)f_{n'}^j(\mathbf{r}_1,\lambda_1)p(\Lambda^{(1)}\mid\mathbf{r}_1,\lambda_1)\,\mathrm{d}\Lambda^{(1)}. \tag*{(62)}

We can simplify the above by using the definition of the conditional average (B.2) of fn′j(r1,λ1;rj,λj)f_{n'}^j(\mathbf{r}_1,\lambda_1;\mathbf{r}_j,\lambda_j), using (13), and that particles are indistinguishable to obtain

⟨fn1⟩(r1,λ1)=Tn(λ1)∑n′Vn′−n(kr1)gn′+Tn(λ1)∑n′∫Sn(λ2)∫R2Un′−n(kr1−kr2)⟨fn′2⟩(r2,λ2;r1,λ1)g(r1,λ1;r2,λ2) dr2dλ2.(63)\left\langle f_n^1\right\rangle(\mathbf{r}_1,\lambda_1)=T_n(\lambda_1)\sum_{n'}V_{n'-n}(k\mathbf{r}_1)g_{n'}+T_n(\lambda_1)\sum_{n'}\int_{\mathcal{S}}n(\lambda_2)\int_{R_2}U_{n'-n}(k\mathbf{r}_1-k\mathbf{r}_2)\left\langle f_{n'}^2\right\rangle(\mathbf{r}_2,\lambda_2;\mathbf{r}_1,\lambda_1)g(\mathbf{r}_1,\lambda_1;\mathbf{r}_2,\lambda_2)\,\mathrm{d}\mathbf{r}_2\mathrm{d}\lambda_2. \tag*{(63)}

However, this equation is not a closed form equation for ⟨fn1⟩(r1,λ1)\left\langle f_n^1\right\rangle(\mathbf{r}_1,\lambda_1) and an extra assumption is required to proceed further.

A standard solution found in the literature to tackle the problem mentioned above is to use the quasi-crystalline approximation (QCA). It is stated as follows:

⟨fn2⟩(r1,λ1;r2,λ2)≈⟨fn2⟩(r2,λ2),∣r1−r2∣≥a12(QCA).(64)\left\langle f_n^2\right\rangle(\mathbf{r}_1,\lambda_1;\mathbf{r}_2,\lambda_2)\approx\left\langle f_n^2\right\rangle(\mathbf{r}_2,\lambda_2),\qquad\lvert\mathbf{r}_1-\mathbf{r}_2\rvert\ge a_{12}\quad\text{(QCA)}. \tag*{(64)}

See [12] for a brief discussion on this approximation.

Finally we use that particles are indistinguishable which implies that ⟨fn1⟩(r1,λ1)=⟨fn2⟩(r2,λ2)\left\langle f_n^1\right\rangle(\mathbf{r}_1,\lambda_1)=\left\langle f_n^2\right\rangle(\mathbf{r}_2,\lambda_2) when r1=r2\mathbf{r}_1=\mathbf{r}_2 and λ1=λ2\lambda_1=\lambda_2 to substitute

⟨fn2⟩(r2,λ2)=⟨fn⟩(r2,λ2)and⟨fn1⟩(r1,λ1)=⟨fn⟩(r1,λ1)(65)\left\langle f_n^2\right\rangle(\mathbf{r}_2,\lambda_2)=\langle f_n\rangle(\mathbf{r}_2,\lambda_2)\quad\text{and}\quad\left\langle f_n^1\right\rangle(\mathbf{r}_1,\lambda_1)=\langle f_n\rangle(\mathbf{r}_1,\lambda_1) \tag*{(65)}

into (B.5), which together with QCA (B.6) leads to the average governing equation (21).

The effective waves method

Derivation of two ensemble equations

Here we show how to use the effective wave assumption to rewrite the governing equation (30) into two separate equations: the effective wave equation and the effective boundary conditions. To achieve this we define using set builder notation:

D(y,a):={x∈R:∣x−y∣≤a}.D(\mathbf{y},a):=\{\mathbf{x}\in\mathbb{R}:\lvert\mathbf{x}-\mathbf{y}\rvert\le a\}.

Using the decomposition of the pair correlation function (14), we split the domain of integration in the governing equation (21) into two integrals: one over D(r1,a12)D(\mathbf{r}_1,a_{12}), another one over R2∖D(r1,a12)\mathcal{R}_2 \setminus D(r_1,a_{12}) in the form

⟨fn⟩(r1,λ1)=Tn(λ1)∑n′VN−n(kr1)gn′+Tn(λ1)∑n′∫Sn(λ2)∫R2∖D(r1,a12)Un′−n(kr1−kr2)⟨fn′⟩(r2,λ2) dr2dλ2+Tn(λ1)∑n′∫Sn(λ2)∫D(r1,a12,b12)Un′−n(kr1−kr2)⟨fn′⟩(r2,λ2)δg(∣r1−r2∣,λ1,λ2) dr2dλ2.(66)\begin{aligned} \langle f_n\rangle(r_1,\lambda_1)={}&T_n(\lambda_1)\sum_{n'}V_{N-n}(kr_1)g_{n'}\\ &+T_n(\lambda_1)\sum_{n'}\int_{\mathcal{S}}\mathbf{n}(\lambda_2)\int_{\mathcal{R}_2\setminus D(r_1,a_{12})}U_{n'-n}(kr_1-kr_2)\langle f_{n'}\rangle(r_2,\lambda_2)\,dr_2d\lambda_2\\ &+T_n(\lambda_1)\sum_{n'}\int_{\mathcal{S}}\mathbf{n}(\lambda_2)\int_{D(r_1,a_{12},b_{12})}U_{n'-n}(kr_1-kr_2)\langle f_{n'}\rangle(r_2,\lambda_2)\delta g(\lvert r_1-r_2\rvert,\lambda_1,\lambda_2)\,dr_2d\lambda_2. \tag*{(66)} \end{aligned}

where D(r1,a12,b12):=D(r1,b12)∖D(r1,a12)D(r_1,a_{12},b_{12}):=D(r_1,b_{12})\setminus D(r_1,a_{12}) and the annulus D(r1,a12,b12)D(r_1,a_{12},b_{12}) is completely contain within R2\mathcal{R}_2 when

dist⁡(r1,∂R2)≥b12.(67)\operatorname{dist}(r_1,\partial\mathcal{R}_2)\ge b_{12}. \tag*{(67)}

In this section, and in the whole paper, we only solve (30) for r1r_1 that satisfies the above. This avoids the boundary layer [11, 8] which greatly complicates the solution and is only needed when there is a large particle volume fraction, moderate frequencies, and strongly scattering particles.

The last integral in (C.1) can be simplified by changing the variable of integration to r=r2−r1r=r_2-r_1 which leads to the integral

Kn′n(r1,λ2):=∫D(0,a12,b12)Un′−n(−kr)⟨fn′⟩(r+r1,λ2)δg(r,λ1,λ2) dr.(68)K_{n'n}(r_1,\lambda_2):=\int_{D(0,a_{12},b_{12})}U_{n'-n}(-kr)\langle f_{n'}\rangle(r+r_1,\lambda_2)\delta g(r,\lambda_1,\lambda_2)\,dr. \tag*{(68)}

The first integral over r2r_2 in (C.1) can be further simplified by using Green’s theorem to replace the volume integral over R2∖D(r1,a12)\mathcal{R}_2\setminus D(r_1,a_{12}) by surface integrals: given any two function smooth functions u,vu,v which satisfy

Δu(r)+k⋆u(r)=0andΔv(r)+kv(r)=0,\Delta u(r)+k_{\star}u(r)=0\quad\text{and}\quad\Delta v(r)+kv(r)=0,

over a set Ω\Omega we have that

(k2−k⋆2)∫Ωuv dr=∫Ω(Δuv−uΔv) dr=∫∂Ω(∂νuv−u∂νv) ds(r).(69)(k^2-k_{\star}^2)\int_{\Omega}uv\,dr=\int_{\Omega}(\Delta uv-u\Delta v)\,dr=\int_{\partial\Omega}(\partial_{\nu}uv-u\partial_{\nu}v)\,ds(r). \tag*{(69)}

With u(r2)u(r_2) substituted for ⟨fn′,N⟩(r2,λ2)\langle f_{n',N}\rangle(r_2,\lambda_2) and v(r2)v(r_2) substituted for Un′−n(kr1−kr2)U_{n'-n}(kr_1-kr_2) we can use the above to deduce:

∫R2∖D(r1,a12)Un′−n(kr1−kr2)⟨fn′⟩(r2,λ2) dr2=In′n(r1)−Jn′n(r1)k2−k⋆2.(70)\int_{\mathcal{R}_2\setminus D(r_1,a_{12})}U_{n'-n}(kr_1-kr_2)\langle f_{n'}\rangle(r_2,\lambda_2)\,dr_2=\frac{I_{n'n}(r_1)-J_{n'n}(r_1)}{k^2-k_{\star}^2}. \tag*{(70)}

where we defined

In′n(r1):=∫∂R2Un′−n(kr1−kr2)∂⟨fn′⟩(r2,λ2)∂ν2−∂Un′−n(kr1−kr2)∂ν2⟨fn′⟩(r2,λ2) dA2,I_{n'n}(r_1):=\int_{\partial\mathcal{R}_2}U_{n'-n}(kr_1-kr_2)\frac{\partial\langle f_{n'}\rangle(r_2,\lambda_2)}{\partial\nu_2}-\frac{\partial U_{n'-n}(kr_1-kr_2)}{\partial\nu_2}\langle f_{n'}\rangle(r_2,\lambda_2)\,dA_2,
Jn′n(r1):=∫∂D(0,a12)Un′−n(−kr)∂⟨fn′⟩(r+r1,λ2)∂ν−∂Un′−n(−kr)∂ν⟨fn′⟩(r+r1,λ2) dA.(71)J_{n'n}(r_1):=\int_{\partial D(0,a_{12})}U_{n'-n}(-kr)\frac{\partial\langle f_{n'}\rangle(r+r_1,\lambda_2)}{\partial\nu}-\frac{\partial U_{n'-n}(-kr)}{\partial\nu}\langle f_{n'}\rangle(r+r_1,\lambda_2)\,dA. \tag*{(71)}

Finally, substituting (C.3) and (C.5) into the governing equation (21) gives

⟨fn⟩(r1,λ1)=Tn(λ1)∑n′Vn′−n(kr1)gn′+∑n′Tn(λ1)∫S[In′n(r1)−Jn′n(r1)k2−k⋆2+Kn′n(r1)]n(λ2)dλ2.(72)\begin{aligned} \langle f_n\rangle(r_1,\lambda_1)={}&T_n(\lambda_1)\sum_{n'}V_{n'-n}(kr_1)g_{n'}\\ &+\sum_{n'}T_n(\lambda_1)\int_{\mathcal{S}}\left[\frac{I_{n'n}(r_1)-J_{n'n}(r_1)}{k^2-k_{\star}^2}+K_{n'n}(r_1)\right]\mathbf{n}(\lambda_2)d\lambda_2. \tag*{(72)} \end{aligned}

The above now can be split into two separate equations by noting that the functions ⟨fn⟩(r1)\langle f_n\rangle(r_1), Jn′n(r1)\mathcal{J}_{n'n}(r_1) and Kn′n(r1)\mathcal{K}_{n'n}(r_1) satisfy the wave equation with wavenumber k⋆k_\star, while Vn(kr1)V_n(kr_1) and In′n(r1)\mathcal{I}_{n'n}(r_1) satisfy the wave equation with wavenumber kk. Since solutions of the Helmholtz equation with different wavenumbers are independent, see [10] for details, (C.7) can be split into the ensemble wave equation (23) containing the terms with wavenumber k⋆k_\star and the ensemble boundary conditions (24) containing the terms with wavenumber kk.

The effective eigensystem

Here we deduce a general eigensystem which can be used to determine the effective wavenumber k⋆k_\star and write ⟨fn⟩(r1,λ1)\langle f_n\rangle(r_1,\lambda_1) in terms of eigenfunctions.

Since ⟨fn⟩(r1,λ1)\langle f_n\rangle(r_1,\lambda_1) satisfies the wave equation (22), it can be decomposed into the modes

⟨fn⟩(r1,λ1)=∑n1Fnn1(λ1)Vn1(k⋆r1),(73)\langle f_n\rangle(r_1,\lambda_1)=\sum_{n_1}F_{nn_1}(\lambda_1)V_{n_1}(k_\star r_1), \tag*{(73)}

where Vn1V_{n_1} is defined in (4).

The unknowns k⋆k_\star and Fnn1(λ1)F_{nn_1}(\lambda_1) can be determined by substituting (C.8) into (23), which requires the term

⟨fn⟩(r+r1,λ2)=∑n1n2Fnn1(λ2)Vn1−n2(k⋆r)Vn2(k⋆r1),(74)\langle f_n\rangle(r+r_1,\lambda_2)=\sum_{n_1n_2}F_{nn_1}(\lambda_2)V_{n_1-n_2}(k_\star r)V_{n_2}(k_\star r_1), \tag*{(74)}

where the right side is a result of using Graf’s addition theorem (A.1, i) in (25).

By substituting (C.9) in Kn′n(r1)\mathcal{K}_{n'n}(r_1) (C.3) we can use the orthogonality of the cylindrical Bessel functions to remove the sum over n2n_2, because only the cases (n1−n2)=(n−n′)(n_1-n_2)=(n-n') are non-zero. Likewise, we can perform the same simplification by substituting (C.9) in Jn′n(r1)\mathcal{J}_{n'n}(r_1) (C.6). The simplifications result in

Kn′n(r1)=2πWn′−n(k,k⋆)∑n1Fn′n1(λ2)Vn1+n′−n(k⋆r1),\mathcal{K}_{n'n}(r_1)=2\pi W_{n'-n}(k,k_\star)\sum_{n_1}F_{n'n_1}(\lambda_2)V_{n_1+n'-n}(k_\star r_1),
Jn′n(r1)=−2πNn′−n(ka12,k⋆a12)∑n1Fn′n1(λ2)Vn1+n′−n(k⋆r1),(75)\mathcal{J}_{n'n}(r_1)=-2\pi N_{n'-n}(ka_{12},k_\star a_{12})\sum_{n_1}F_{n'n_1}(\lambda_2)V_{n_1+n'-n}(k_\star r_1), \tag*{(75)}

where we introduced the notations

Wl(k,k⋆):=∫a12b12Hl(kr)Jl(k⋆r)δg(r,λ1,λ2)r dr,W_l(k,k_\star):=\int_{a_{12}}^{b_{12}}H_l(kr)J_l(k_\star r)\delta g(r,\lambda_1,\lambda_2)r\,dr,
Nl(x,y):=xHl′(x)Jl(y)−yHl(x)Jl′(y).(76)N_l(x,y):=xH'_l(x)J_l(y)-yH_l(x)J'_l(y). \tag*{(76)}

Finally, substituting (C.10) in the ensemble wave equation (23), and again using the orthogonality of the cylindrical Bessel functions, we reach equation (26) where the following term appears:

Nl12(k,k⋆)=2πNl(ka12,k⋆a12)k⋆2−k2−2πWl(k,k⋆).(77)N_l^{12}(k,k_\star)=2\pi\frac{N_l(ka_{12},k_\star a_{12})}{k_\star^2-k^2}-2\pi W_l(k,k_\star). \tag*{(77)}

Boundary condition for effective waves

The eigensystem (26) is not enough to fully determine the Fnn1F_{nn_1}, to do so we need to substitute (25) into the ensemble boundary condition (24). To achieve this the first step is to use Graf’s addition theorem (A.1, iii) with x=r1x=r_1 and y=−r2y=-r_2 to obtain

Un′−n(kr1−kr2)=∑n3Vn3(kr1)Un′−n−n3(−kr2)=∑n2Vn2−n+n′(kr1)U−n2(−kr2).(78)U_{n'-n}(kr_1-kr_2)=\sum_{n_3}V_{n_3}(kr_1)U_{n'-n-n_3}(-kr_2)=\sum_{n_2}V_{n_2-n+n'}(kr_1)U_{-n_2}(-kr_2). \tag*{(78)}

where we used the change of variable n2=n+n3−n′n_{2}=n+n_{3}-n^{\prime}, and that ∣r1∣<∣r2∣\lvert r_{1}\rvert<\lvert r_{2}\rvert because for In′n\mathcal{I}_{n^{\prime}n} the variable r2r_{2} is on the boundary ∂R2\partial\mathcal{R}_{2}, whereas r1r_{1} satisfies (C.2). Substituting (25) in In′n\mathcal{I}_{n^{\prime}n} defined by (C.6) and using the above (D.1) gives

In′n(r1)=∑n2n1Fn′n1(λ2)Bn1n2Vn2−n+n′(kr1)(79)\mathcal{I}_{n^{\prime}n}(r_{1})=\sum_{n_{2}n_{1}}F_{n^{\prime}n_{1}}(\lambda_{2})B_{n_{1}n_{2}}V_{n_{2}-n+n^{\prime}}(kr_{1}) \tag*{(79)}

where

Bn1n2=(−1)n2∫∂R2[U−n2(kr2)∂Vn1(k⋆r2)∂ν2−∂U−n2(kr2)∂ν2Vn1(k⋆r2)] dA2(80)B_{n_{1}n_{2}}=(-1)^{n_{2}}\int_{\partial\mathcal{R}_{2}}\left[U_{-n_{2}}(kr_{2})\frac{\partial V_{n_{1}}(k_{\star}r_{2})}{\partial\nu_{2}}-\frac{\partial U_{-n_{2}}(kr_{2})}{\partial\nu_{2}}V_{n_{1}}(k_{\star}r_{2})\right]\,dA_{2} \tag*{(80)}

then substituting (D.2) in (24) leads to

∑n′Vn′−n(kr1)gn′+∑n′n2n1∫SFn′n1(λ2)Vn2−n+n′(kr1)Bn1n2k2−k⋆2n(λ2) dλ2=0.(81)\sum_{n^{\prime}}V_{n^{\prime}-n}(kr_{1})g_{n^{\prime}}+\sum_{n^{\prime}n_{2}n_{1}}\int_{\mathcal{S}}F_{n^{\prime}n_{1}}(\lambda_{2})V_{n_{2}-n+n^{\prime}}(kr_{1})\frac{B_{n_{1}n_{2}}}{k^{2}-k_{\star}^{2}}\mathfrak{n}(\lambda_{2})\,d\lambda_{2}=0. \tag*{(81)}

We can further simplify the above by using the orthogonality of the functions VnV_{n} to obtain

gN+∑n′n1∫SFn′n1(λ2)Bn1(N−n′)k2−k⋆2n(λ2) dλ2=0,(82)g_{N}+\sum_{n^{\prime}n_{1}}\int_{\mathcal{S}}F_{n^{\prime}n_{1}}(\lambda_{2})\frac{B_{n_{1}(N-n^{\prime})}}{k^{2}-k_{\star}^{2}}\mathfrak{n}(\lambda_{2})\,d\lambda_{2}=0, \tag*{(82)}

which holds for every NN.

When all particles are in a disk, then ∂R2\partial\mathcal{R}_{2} is a circle and the above simplifies. This is the only case we completely resolve in this paper. Let R2R_{2} be the radius of the disk R2\mathcal{R}_{2}, then n2=n1n_{2}=n_{1} in (D.3) which reduces to

Bn1n2=−2πδn1−n2Nn1(kR2,k⋆R2),(83)B_{n_{1}n_{2}}=-2\pi\delta_{n_{1}-n_{2}}N_{n_{1}}(kR_{2},k_{\star}R_{2}), \tag*{(83)}

where Nn1N_{n_{1}} is defined by (C.11)2_2. Substituting this into (D.5) leads to (29).

Elementary proof that thee effective T-matrix is diagonal

We provide an elementary proof that T\mathcal{T} is diagonal when the particles are confined in a disk of radius RR. To this end, we consider the scattering from the modal source uincNu_{\mathrm{inc}}^{N} obtained for gn=δn,Ng_{n}=\delta_{n,N}:

The notation Fn,N\mathcal{F}_{n,N} is the corresponding Fn\mathcal{F}_{n} to the specific incident field with gn=δn,Ng_{n}=\delta_{n,N} (compare with (15)). We then denote by fn,Ni(σ)f_{n,N}^{i}(\sigma) the resulting solution of (7) for the specific configuration σ=r1,…,rJ\sigma=r_{1},\ldots,r_{J}. The rotation by angle ϕ\phi of the particles r1,…,rJr_{1},\ldots,r_{J} correspond to another valid configuration (because the random material is cylindrical), for which the solutions are given by

fn,Ni(Rϕσ)=ei(N−n)ϕfn,Ni(σ)(84)f_{n,N}^{i}(R_{\phi}\sigma)=e^{\mathrm{i}(N-n)\phi}f_{n,N}^{i}(\sigma) \tag*{(84)}

This tells us how the rotation of a configuration modifies the coefficient Fn,N\mathcal{F}_{n,N}, using (15):

Fn,N(Rϕσ)=∑i=1J∑n′=−∞+∞Vn−n′(kri)ei(N−n)ϕfn′,Ni(σ).(85)\mathcal{F}_{n,N}(R_{\phi}\sigma)=\sum_{i=1}^{J}\sum_{n^{\prime}=-\infty}^{+\infty}V_{n-n^{\prime}}(kr_{i})e^{\mathrm{i}(N-n)\phi}f_{n^{\prime},N}^{i}(\sigma). \tag*{(85)}

Consequently,

⟨Fn,N⟩=∫σFn,N(σ)p(σ) dσ=∫σ12π∫02πFn,N(Rϕσ)p(σ) dσ dϕ=δn,N∫σFn,N(σ)p(σ) dσ(86)\left\langle\mathcal{F}_{n,N}\right\rangle=\int_{\sigma}\mathcal{F}_{n,N}(\sigma)p(\sigma)\,d\sigma=\int_{\sigma}\frac{1}{2\pi}\int_{0}^{2\pi}\mathcal{F}_{n,N}(R_{\phi}\sigma)p(\sigma)\,d\sigma\,d\phi=\delta_{n,N}\int_{\sigma}\mathcal{F}_{n,N}(\sigma)p(\sigma)\,d\sigma \tag*{(86)}

Finally, we deduce

Tn,N=δn,N∫Fn,N(r1,…,rJ)p(r1,…,rJ) dr1,…,drJ.(87)T_{n,N} = \delta_{n,N} \int\mathcal{F}_{n,N}(r_1,\ldots,r_J)p(r_1,\ldots,r_J)\,\mathrm{d}r_1,\ldots,\mathrm{d}r_J. \tag*{(87)}

This analysis proves that only the diagonal terms of the effective T-matrix are nonzero and can be estimated by Tn,n=⟨Fn,n⟩T_{n,n} = \langle\mathcal{F}_{n,n}\rangle.

References

  1. [32]L. Tsang, J. A. Kong, and T. Habashy. Multiple scattering of acoustic waves by random distribution of discrete spherical scatterers with the quasicrystalline and Percus–Yevick approximation. The Journal of the Acoustical Society of America, 71(3):552–558, March 1982.DOI
  2. [33]Victor Twersky. On scattering of waves by random distributions. i. free-space scatterer formalism. Journal of Mathematical Physics, 3(4):700–715, 1962.DOI
  3. [34]Victor Twersky. On scattering of waves by random distributions. ii. two-space scatterer formalism. Journal of Mathematical Physics, 3(4):724–734, 1962.DOI
  4. [35]V. K. Varadan, V. N. Bringi, V. V. Varadan, and A. Ishimaru. Multiple scattering theory for waves in discrete random media and comparison with experiments. Radio Science, 18(3):321–327, May 1983.DOI
  5. [36]Kevin Vynck, Romain Pierrat, Rémi Carminati, Luis S Froufe-Pérez, Frank Scheffold, Riccardo Sapienza, Silvia Vignolini, and Juan José Sáenz. Light in correlated disordered media. arXiv preprint arXiv:2106.13892, 2021.DOI
  6. [37]Peter Cary Waterman and Rohn Truell. Multiple scattering of waves. Journal of mathematical physics, 2(4):512–537, 1961.DOI

Paper details

vlnc:W_s34e5fta
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents