2.1 Low Mach number expansion

Equation (2.1) can describe a range of physical phenomena, which is excessive in this case because the system’s behaviour is governed to first order by only two phenomena. The first is steady flow in a channel with rigid boundaries, with the inlet velocity of order $\bar{U}=0.1~\text{m}~\text{s}^{-1}$ and $Re\approx 1$ . The second is periodic acoustic oscillation, with a small displacement amplitude at the boundary $\unicode[STIX]{x1D6E5}\leqslant 0.1~\unicode[STIX]{x03BC}\text{m}$ and a high oscillation frequency $\unicode[STIX]{x1D714}\approx 100~\text{kHz}$ . The characteristic oscillation velocity is of order $\tilde{U} =\unicode[STIX]{x1D714}\unicode[STIX]{x1D6E5}\approx 0.01~\text{m}~\text{s}^{-1}$ .

We choose the ambient state density $\unicode[STIX]{x1D70C}^{b}$ and the speed of sound $(c_{s}^{b})^{2}=(\unicode[STIX]{x2202}P/\unicode[STIX]{x2202}\unicode[STIX]{x1D70C})_{s}$ as the reference dimensional density and velocity, and the characteristic domain size $L$ as the reference length. The reference pressure $P^{b}$ is chosen as a function of density and the speed of sound: $P^{b}=\unicode[STIX]{x1D70C}^{b}(c_{s}^{b})^{2}$ , and the reference temperature $T^{b}$ is the ambient temperature.

In this problem, we assume that the local Mach number is small:

(2.6) $$\begin{eqnarray}M\equiv \frac{|u|}{c_{s}^{b}}\ll 1.\end{eqnarray}$$

The characteristic velocity amplitudes of the steady flow, $\bar{U}$ , and the oscillating flow, $\tilde{U}$ , are also small in comparison to the speed of sound, which allows us to introduce two small parameters: the steady flow Mach number, $\unicode[STIX]{x1D707}$ , and the oscillating flow Mach number, $\unicode[STIX]{x1D716}$ :

(2.7a ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D707}\equiv \frac{\bar{U}}{c_{s}^{b}}\simeq \frac{0.1}{1000}\ll 1, & \displaystyle\end{eqnarray}$$

(2.7b ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D716}\equiv \frac{\tilde{U} }{c_{s}^{b}}\simeq \frac{0.01}{1000}\ll 1. & \displaystyle\end{eqnarray}$$

The oscillating flow time scale differs greatly from the steady flow time scale. The oscillating time scale is $t_{ac}\sim L/c_{s}$ , and the steady flow time scale is $t_{hyd}\sim L/\bar{U}=\unicode[STIX]{x1D707}^{-1}t_{ac}\gg t_{ac}$ . This allows us to decouple two phenomena and study them independently. We consider a generic state variable $\unicode[STIX]{x1D713}(\boldsymbol{x},t)=(\unicode[STIX]{x1D70C},u_{i},P)$ . We denote a zero-order state variable by $\unicode[STIX]{x1D713}_{0}(\boldsymbol{x},t)$ , as if the steady flow and the oscillating flow were absent, i.e. $\unicode[STIX]{x1D716}=\unicode[STIX]{x1D707}=0$ . If there is no external energy and momentum production (by imposed temperature gradients, heat release or body forces), then $\unicode[STIX]{x1D713}_{0}$ is uniform in space and constant in time. We then assume that the perturbation $\unicode[STIX]{x1D719}(\boldsymbol{x},t)$ of $\unicode[STIX]{x1D713}_{0}$ is at least linearly proportional to $\unicode[STIX]{x1D707}$ and $\unicode[STIX]{x1D716}$ , such that:

(2.8) $$\begin{eqnarray}\unicode[STIX]{x1D713}(\boldsymbol{x},t)=\unicode[STIX]{x1D713}_{0}+\unicode[STIX]{x1D719}(\boldsymbol{x},t,\unicode[STIX]{x1D707},\unicode[STIX]{x1D716}).\end{eqnarray}$$

We assume that a flow state perturbation related to a particular phenomenon depends solely on the phenomenon’s temporal scale, such that $\unicode[STIX]{x1D719}(\boldsymbol{x},t)$ becomes a sum of the slow hydrodynamic perturbation $\bar{\unicode[STIX]{x1D719}}(\boldsymbol{x},t,\unicode[STIX]{x1D707})$ , labelled the steady flow, and the fast acoustic perturbation $\tilde{\unicode[STIX]{x1D719}}(\boldsymbol{x},t.\unicode[STIX]{x1D716})$ , labelled the oscillating flow:

(2.9a ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D719}(\boldsymbol{x},t,\unicode[STIX]{x1D707},\unicode[STIX]{x1D716})=\bar{\unicode[STIX]{x1D719}}(\boldsymbol{x},t,\unicode[STIX]{x1D707})+\tilde{\unicode[STIX]{x1D719}}(\boldsymbol{x},t,\unicode[STIX]{x1D716}), & \displaystyle\end{eqnarray}$$

(2.9b ) $$\begin{eqnarray}\displaystyle & \displaystyle \bar{\unicode[STIX]{x1D719}}(\boldsymbol{x},t,\unicode[STIX]{x1D707})=\frac{1}{T_{ac}}\int _{T_{ac}}\unicode[STIX]{x1D719}(\boldsymbol{x},t,\unicode[STIX]{x1D707},\unicode[STIX]{x1D716})\,\text{d}t. & \displaystyle\end{eqnarray}$$

$\unicode[STIX]{x1D713}(\boldsymbol{x},t)$

$\unicode[STIX]{x1D713}_{0}$

$\bar{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{hyd},\unicode[STIX]{x1D707})$

$\tilde{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{ac},\unicode[STIX]{x1D716})$

(2.10) $$\begin{eqnarray}\unicode[STIX]{x1D713}(\boldsymbol{x},t)=\unicode[STIX]{x1D713}_{0}+\bar{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{hyd},\unicode[STIX]{x1D707})+\tilde{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{ac},\unicode[STIX]{x1D716}).\end{eqnarray}$$

We can perform a low Mach number expansion in terms of $\unicode[STIX]{x1D707}$ and $\unicode[STIX]{x1D716}$ because they are both small. The state perturbations $\bar{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{hyd})$ and $\tilde{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{ac})$ independently tend to zero as $\unicode[STIX]{x1D707}\rightarrow 0$ and $\unicode[STIX]{x1D716}\rightarrow 0$ , so we assume low Mach number decompositions of the form $\bar{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{hyd})=\Vert \bar{\unicode[STIX]{x1D719}}\Vert \sum \unicode[STIX]{x1D707}^{k}\bar{\unicode[STIX]{x1D719}}^{(k)}(\boldsymbol{x},t_{hyd})$ and $\tilde{\unicode[STIX]{x1D719}}(\boldsymbol{x},t_{ac})=\Vert \tilde{\unicode[STIX]{x1D719}}\Vert \sum \unicode[STIX]{x1D716}^{k}\tilde{\unicode[STIX]{x1D719}}^{(k)}(\boldsymbol{x},t_{ac})$ , where $\bar{\unicode[STIX]{x1D719}}^{(k)},\tilde{\unicode[STIX]{x1D719}}^{(k)}$ are the $k$ th-order non-dimensional perturbation shapes, and $\Vert \bar{\unicode[STIX]{x1D719}}\Vert ,\Vert \tilde{\unicode[STIX]{x1D719}}\Vert$ are the characteristic dimensional magnitudes of the variables: $\Vert \bar{\unicode[STIX]{x1D70C}}\Vert =\Vert \tilde{\unicode[STIX]{x1D70C}}\Vert =\unicode[STIX]{x1D70C}^{b},\Vert \bar{u}\Vert =\bar{U}$ , $\Vert \tilde{u} \Vert =\tilde{U}$ , $\Vert \bar{P}\Vert =\Vert \tilde{P}\Vert =P^{b}$ .

We neglect the interaction between the steady flow and the oscillating flow given by the higher-order mixed terms $\sum \unicode[STIX]{x1D707}^{n}\unicode[STIX]{x1D716}^{m}\unicode[STIX]{x1D719}^{(m+n)}(x,t_{ac},t_{hyd})$ ; $m,n\geqslant 1$ because $\unicode[STIX]{x1D716}$ and $\unicode[STIX]{x1D707}$ are both small. The expansion of the primal variables is therefore:

(2.11a ) $$\begin{eqnarray}\displaystyle \unicode[STIX]{x1D70C}(x,t) & = & \displaystyle \unicode[STIX]{x1D70C}^{b}\unicode[STIX]{x1D70C}_{0}+\unicode[STIX]{x1D70C}^{b}(\unicode[STIX]{x1D707}\bar{\unicode[STIX]{x1D70C}}^{(1)}+\unicode[STIX]{x1D716}\tilde{\unicode[STIX]{x1D70C}}^{(1)})+\mathit{O}(\unicode[STIX]{x1D707}^{2},\unicode[STIX]{x1D716}^{2},\unicode[STIX]{x1D707}\unicode[STIX]{x1D716}),\end{eqnarray}$$

(2.11b ) $$\begin{eqnarray}\displaystyle u(x,t) & = & \displaystyle c_{s}^{b}(\unicode[STIX]{x1D707}\bar{u}^{(1)}+\unicode[STIX]{x1D716}\tilde{u} ^{(1)})+\mathit{O}(\unicode[STIX]{x1D707}^{2},\unicode[STIX]{x1D716}^{2},\unicode[STIX]{x1D707}\unicode[STIX]{x1D716}),\end{eqnarray}$$

(2.11c ) $$\begin{eqnarray}\displaystyle P(x,t) & = & \displaystyle P^{b}P_{0}+P^{b}(\unicode[STIX]{x1D707}\bar{P}^{(1)}+\unicode[STIX]{x1D707}^{2}\bar{P}^{(2)}+\unicode[STIX]{x1D716}\tilde{P}^{(1)})+\mathit{O}(\unicode[STIX]{x1D707}^{3},\unicode[STIX]{x1D716}^{2},\unicode[STIX]{x1D707}\unicode[STIX]{x1D716}).\end{eqnarray}$$

$\unicode[STIX]{x1D707}^{2}\bar{P}^{(2)}$

$\bar{P}^{(1)}$

2.2 Zero Mach number limit

Substituting the primal variables expansion (2.11) into (2.1) and (2.5), and collecting the zero-order terms, we obtain:

(2.12a ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D735}_{i}P^{(0)}(x)=0, & \displaystyle\end{eqnarray}$$

(2.12b ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D735}_{k}(\unicode[STIX]{x1D705}\unicode[STIX]{x1D735}_{k}T^{(0)}(x))=0, & \displaystyle\end{eqnarray}$$

(2.12c ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D70C}^{(0)}(x)=\unicode[STIX]{x1D70C}(P^{(0)},T^{(0)}). & \displaystyle\end{eqnarray}$$

The zeroth-order equations describe the ambient state, $\unicode[STIX]{x1D716}=\unicode[STIX]{x1D707}=0$ . Equation (2.12a ) shows that the ambient pressure $P_{0}$ is spatially uniform, and (2.12b ) describes the temperature distribution of the ambient state. If all the boundaries have uniform and constant temperature, then the ambient temperature and density are uniform and non-dimensionalized as $T^{(0)}(x)=1,\unicode[STIX]{x1D70C}^{(0)}(x)=1$ .

2.3 Low Mach number steady flow

Collecting the first-order terms of $\unicode[STIX]{x1D707}$ in the continuity equation and the second-order terms of $\unicode[STIX]{x1D707}^{2}$ in the momentum equations (2.2) and assuming a Newtonian fluid results in the incompressible Navier–Stokes equation:

(2.13a ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D735}_{i}\bar{u}_{i}^{(1)}=0, & \displaystyle\end{eqnarray}$$

(2.13b ) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t_{hyd}}\bar{u}_{i}^{(1)}+(\bar{u}_{j}^{(1)}\unicode[STIX]{x1D735}_{j})\bar{u}_{i}^{(1)}+\unicode[STIX]{x1D735}_{i}\bar{P}^{(2)}-\frac{1}{\overline{Re}}\unicode[STIX]{x0394}\bar{u}_{i}^{(1)}=0. & \displaystyle\end{eqnarray}$$

$\bar{P}$

$\bar{P}=\unicode[STIX]{x1D707}^{2}\bar{P}^{(2)}+\mathit{O}(\unicode[STIX]{x1D707}^{3})$

$\overline{Re}\equiv \unicode[STIX]{x1D70C}^{b}LU_{in}/\unicode[STIX]{x1D707}_{vis}$

We supplement the steady flow equations with a prescribed velocity boundary condition at the inlet, $\unicode[STIX]{x1D6E4}_{in}$ , a no-slip boundary condition on the walls, $\unicode[STIX]{x1D6E4}_{w}$ , and a zero stress boundary condition at the outlet, $\unicode[STIX]{x1D6E4}_{out}$ :

(2.14a ) $$\begin{eqnarray}\displaystyle \bar{u}_{i}^{(1)} & = & \displaystyle U_{i}\quad \text{on }\unicode[STIX]{x1D6E4}_{in},\end{eqnarray}$$

(2.14b ) $$\begin{eqnarray}\displaystyle \bar{u}_{i}^{(1)} & = & \displaystyle 0\quad \text{on }\unicode[STIX]{x1D6E4}_{w},\end{eqnarray}$$

(2.14c ) $$\begin{eqnarray}\displaystyle -\bar{P}^{(2)}n_{i}+\frac{1}{Re}\frac{\unicode[STIX]{x2202}\bar{u}_{i}^{(1)}}{\unicode[STIX]{x2202}n} & = & \displaystyle 0\quad \text{on }\unicode[STIX]{x1D6E4}_{out}.\end{eqnarray}$$

2.4 Low Mach number oscillating flow

Collecting the first-order terms of $\unicode[STIX]{x1D716}$ , the oscillating flow continuity, momentum and energy equations are governed by:

(2.15a ) $$\begin{eqnarray}\displaystyle \frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t_{ac}}\tilde{\unicode[STIX]{x1D70C}}^{(1)}+\unicode[STIX]{x1D735}_{i}\tilde{u} _{i}^{(1)} & = & \displaystyle 0,\end{eqnarray}$$

(2.15b ) $$\begin{eqnarray}\displaystyle \frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t_{ac}}\tilde{u} _{i}^{(1)}+\unicode[STIX]{x1D735}_{i}\tilde{P}^{(1)} & = & \displaystyle \frac{1}{\tilde{Re}}\unicode[STIX]{x1D735}_{j}\tilde{\unicode[STIX]{x1D70F}}_{ij}^{(1)},\end{eqnarray}$$

(2.15c ) $$\begin{eqnarray}\displaystyle \frac{s^{b}}{c_{p}}\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t_{ac}}\tilde{s}^{(1)} & = & \displaystyle \frac{1}{\tilde{Pe}}\unicode[STIX]{x0394}\tilde{T}^{(1)}.\end{eqnarray}$$

The Reynolds and Péclet numbers based on the speed of sound are $\tilde{Re}\equiv \unicode[STIX]{x1D70C}^{b}Lc_{s}^{b}/\unicode[STIX]{x1D707}_{vis}$ and $\tilde{Pe}\equiv \unicode[STIX]{x1D70C}^{b}Lc_{s}^{b}c_{p}/\unicode[STIX]{x1D705}$ . The heat capacity ratio is $\unicode[STIX]{x1D6FE}\equiv c_{p}/c_{v}$ , where $c_{p}$ and $c_{v}$ are the specific heats at constant pressure and constant volume, and $s^{b}$ is the dimensional ambient state entropy. The viscous contribution to the mechanical energy dissipation $\unicode[STIX]{x1D735}_{k}(\tilde{\unicode[STIX]{x1D70F}}_{kj}^{(1)}\tilde{u} _{j}^{(1)})$ is absent in (2.15c ) because it is second order in $\unicode[STIX]{x1D716}$ and therefore negligible. We solve the oscillating flow equations in terms of the oscillating flow pressure $\tilde{P}^{(1)}$ , velocity $\tilde{u} _{i}^{(1)}$ and temperature $\tilde{T}^{(1)}$ , and express the flow density and entropy as $\tilde{s}^{(1)}=\tilde{s}(\tilde{P}^{(1)},\tilde{T}^{(1)})$ , and $\tilde{\unicode[STIX]{x1D70C}}^{(1)}=\tilde{\unicode[STIX]{x1D70C}}(\tilde{P}^{(1)},\tilde{T}^{(1)})$ using the following thermodynamic equalities:

(2.16a ) $$\begin{eqnarray}\displaystyle & \displaystyle s^{b}\tilde{s}^{(1)}=\left(\frac{\unicode[STIX]{x2202}S}{\unicode[STIX]{x2202}P}\right)_{T}P^{b}\tilde{P}^{(1)}+\left(\frac{\unicode[STIX]{x2202}S}{\unicode[STIX]{x2202}T}\right)_{P}T^{b}\tilde{T}^{(1)}=-\frac{\unicode[STIX]{x1D6FC}_{p}}{\unicode[STIX]{x1D70C}_{b}}P^{b}\tilde{P}^{(1)}+\frac{c_{p}}{T_{b}}T^{b}\tilde{T}^{(1)}, & \displaystyle\end{eqnarray}$$

(2.16b ) $$\begin{eqnarray}\displaystyle & \displaystyle \unicode[STIX]{x1D70C}^{b}\tilde{\unicode[STIX]{x1D70C}}^{(1)}=\left(\frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D70C}}{\unicode[STIX]{x2202}P}\right)_{T}P^{b}\tilde{P}^{(1)}+\left(\frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D70C}}{\unicode[STIX]{x2202}T}\right)_{P}T^{b}\tilde{T}^{(1)}=\frac{\unicode[STIX]{x1D6FE}}{(c_{s}^{b})^{2}}P^{b}\tilde{P}^{(1)}-\unicode[STIX]{x1D70C}^{b}\unicode[STIX]{x1D6FC}_{p}T^{b}\tilde{T}^{(1)}, & \displaystyle\end{eqnarray}$$

$\unicode[STIX]{x1D6FC}_{p}\equiv \unicode[STIX]{x1D70C}^{b}(\unicode[STIX]{x2202}V/\unicode[STIX]{x2202}T)_{P}$

$\unicode[STIX]{x1D6FC}_{p}T^{b}\tilde{T}^{(1)}\rightarrow \tilde{T}^{(1)}$

$c_{p}-c_{v}=T^{b}(c_{s}^{b})^{2}\unicode[STIX]{x1D6FC}_{p}^{2}/\unicode[STIX]{x1D6FE}$

(2.17a ) $$\begin{eqnarray}\displaystyle \frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t_{ac}}(\unicode[STIX]{x1D6FE}\tilde{P}^{(1)}-\tilde{T}^{(1)})+\unicode[STIX]{x1D735}_{i}\tilde{u} _{i}^{(1)} & = & \displaystyle 0,\end{eqnarray}$$

(2.17b ) $$\begin{eqnarray}\displaystyle \frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t_{ac}}\left(\frac{\tilde{T}^{(1)}}{\unicode[STIX]{x1D6FE}-1}-\tilde{P}^{(1)}\right) & = & \displaystyle \frac{1}{\tilde{Pe}}\unicode[STIX]{x0394}\tilde{T}^{(1)}.\end{eqnarray}$$

(2.18a,b ) $$\begin{eqnarray}\tilde{\unicode[STIX]{x1D70C}}^{(1)}\equiv \unicode[STIX]{x1D6FE}\tilde{P}^{(1)}-\tilde{T}^{(1)},\quad \tilde{s}^{(1)}\equiv \frac{\tilde{T}^{(1)}}{\unicode[STIX]{x1D6FE}-1}-\tilde{P}^{(1)}.\end{eqnarray}$$

No-slip velocity $\tilde{u} ^{(1)}=0$ and isothermal $\tilde{T}^{(1)}=0$ boundary conditions induce viscous and thermal boundary layers, which damp the acoustic waves. The thickness of the viscous boundary layer $\unicode[STIX]{x1D6FF}_{\unicode[STIX]{x1D708}}$ and the thermal boundary layer $\unicode[STIX]{x1D6FF}_{T}$ depends on the oscillation frequency (Beltman Reference Beltman1999): $\unicode[STIX]{x1D6FF}_{\unicode[STIX]{x1D708}}(\unicode[STIX]{x1D714})=\sqrt{\unicode[STIX]{x1D707}_{vis}/(\unicode[STIX]{x1D70C}^{b}\unicode[STIX]{x1D714})}=\unicode[STIX]{x1D6FF}_{T}\sqrt{Pr}$ . The non-dimensional viscous and thermal boundary layers thicknesses, $\tilde{\unicode[STIX]{x1D6FF}}_{\unicode[STIX]{x1D708}}(\unicode[STIX]{x1D714}),\tilde{\unicode[STIX]{x1D6FF}}_{T}(\unicode[STIX]{x1D714})$ , are:

(2.19a,b ) $$\begin{eqnarray}\tilde{\unicode[STIX]{x1D6FF}}_{\unicode[STIX]{x1D708}}^{2}(\unicode[STIX]{x1D714})=\frac{\unicode[STIX]{x1D6FF}_{\unicode[STIX]{x1D708}}^{2}(\unicode[STIX]{x1D714})}{L^{2}}=\frac{1}{\tilde{Re}}\frac{\unicode[STIX]{x1D714}_{ac}}{\unicode[STIX]{x1D714}},\quad \tilde{\unicode[STIX]{x1D6FF}}_{T}^{2}(\unicode[STIX]{x1D714})=\frac{\unicode[STIX]{x1D6FF}_{T}^{2}(\unicode[STIX]{x1D714})}{L^{2}}=\frac{1}{\tilde{Pe}}\frac{\unicode[STIX]{x1D714}_{ac}}{\unicode[STIX]{x1D714}},\end{eqnarray}$$

where $\unicode[STIX]{x1D714}_{ac}=t_{ac}^{-1}$ is the characteristic acoustic frequency. If the oscillation frequency, $\unicode[STIX]{x1D714}$ , is similar to or smaller than the acoustic frequency, $\unicode[STIX]{x1D714}_{ac}$ , then the viscothermal effects cannot be ignored for general $\tilde{Re},\tilde{Pe}$ . For an inkjet printhead, the fluid viscosity is of order $10^{-2}~\text{Pa}~\text{s}$ , the speed of sound is $10^{3}~\text{ms}^{-1}$ and the channel width is of order $100~\unicode[STIX]{x03BC}\text{m}$ , which results in $\unicode[STIX]{x1D714}_{ac}=10~\text{MHz}$ . At the typical operational frequency of $\unicode[STIX]{x1D714}=100~\text{kHz}$ the viscous boundary-layer thickness is then $\tilde{\unicode[STIX]{x1D6FF}}_{\unicode[STIX]{x1D708}}\sim 0.1$ . The thermal boundary layer thickness is smaller by a factor of $\sqrt{Pr}$ . For inks used in inkjet printers, with $10<Pr<30$ (Seccombe et al. Reference Seccombe1997) $\tilde{\unicode[STIX]{x1D6FF}}_{T}\sim 0.025$ .

We perform a modal decomposition of the oscillating flow state vector $\tilde{\boldsymbol{q}}(x,t)=\hat{\boldsymbol{q}}(x)\text{e}^{st}$ . The Laplace transform of (2.15) results in an eigenvalue problem $sB\hat{\boldsymbol{q}}+A\hat{\boldsymbol{q}}=0$ in terms of $\hat{\boldsymbol{q}}=(\hat{u} _{i},\hat{P},\hat{T})$ , where $\hat{\boldsymbol{q}}$ is the complex eigenfunction, and $s$ is the complex eigenvalue:

(2.20a ) $$\begin{eqnarray}\displaystyle & \displaystyle s\left[\begin{array}{@{}ccc@{}}1 & 0 & 0\\ 0 & \unicode[STIX]{x1D6FE} & -1\\ 0 & -1 & {\displaystyle \frac{1}{\unicode[STIX]{x1D6FE}-1}}\end{array}\right]\left[\begin{array}{@{}c@{}}\hat{u} _{i}\\ \hat{P}\\ \hat{T}\end{array}\right]+\left[\begin{array}{@{}ccc@{}}-{\displaystyle \frac{1}{\tilde{Re}}}\unicode[STIX]{x1D735}_{j}\hat{\unicode[STIX]{x1D749}}_{ij} & \unicode[STIX]{x1D735}_{i} & 0\\ \unicode[STIX]{x1D735}_{i} & 0 & 0\\ 0 & 0 & -{\displaystyle \frac{\unicode[STIX]{x1D6E5}}{(\unicode[STIX]{x1D6FE}-1)\tilde{Pe}}}\end{array}\right]\left[\begin{array}{@{}c@{}}\hat{u} _{i}\\ \hat{P}\\ \hat{T}\end{array}\right]=0,\qquad & \displaystyle\end{eqnarray}$$

(2.20b ) $$\begin{eqnarray}\displaystyle & \displaystyle s=\unicode[STIX]{x1D70E}+\text{i}\unicode[STIX]{x1D714}, & \displaystyle\end{eqnarray}$$

$-\unicode[STIX]{x1D70E}$

$\unicode[STIX]{x1D714}$

$B=B^{H},A=A^{H}$

2.4.1 Boundary conditions

For rigid boundaries we apply a no-slip boundary condition and for open boundaries we apply a stress-free boundary condition:

(2.21a ) $$\begin{eqnarray}\displaystyle \hat{u} _{i} & = & \displaystyle 0\quad \text{on }\unicode[STIX]{x1D6E4}_{nsl},\end{eqnarray}$$

(2.21b ) $$\begin{eqnarray}\displaystyle (-\hat{P}\unicode[STIX]{x1D6FF}_{ij}+\tilde{Re}^{-1}\hat{\unicode[STIX]{x1D70F}}_{ij})n_{j} & = & \displaystyle 0\quad \text{on }\unicode[STIX]{x1D6E4}_{free}.\end{eqnarray}$$

We also apply isothermal and adiabatic boundary conditions for temperature:

(2.22a,b ) $$\begin{eqnarray}\hat{T}=0\quad \text{on }\unicode[STIX]{x1D6E4}_{iso},\quad \frac{\unicode[STIX]{x2202}\hat{T}}{\unicode[STIX]{x2202}n}=0\quad \text{on }\unicode[STIX]{x1D6E4}_{ad}.\end{eqnarray}$$

If the boundaries are not rigid, they displace in reaction to the flow on the boundary. For inviscid flow, the boundary impedance, $Z$ , links the pressure to the velocity on the boundary (Myers Reference Myers1980). $Z$ is typically frequency dependent: $Z=Z(s)$ . For viscous flow, the force at the boundary needs to include the viscous stress, such that the impedance boundary condition is

(2.23) $$\begin{eqnarray}Z\hat{u} _{i}=(-\hat{P}\unicode[STIX]{x1D6FF}_{ij}+\tilde{Re}^{-1}\hat{\unicode[STIX]{x1D70F}}_{ij})n_{j}.\end{eqnarray}$$

Here we neither restrict the tangential velocity to be zero nor forbid tangential displacements of the compliant boundary. As $Z\rightarrow 0$ , the boundary becomes a free surface, $\hat{\unicode[STIX]{x1D70E}}_{ij}n_{j}\rightarrow 0$ . As $Z\rightarrow \infty$ , the boundary becomes a no-slip rigid wall, $\hat{u} _{i}\rightarrow 0$ .

Similarly, the thermal accommodation coefficient $\unicode[STIX]{x1D6FC}_{w}\geqslant 0$ can be introduced to describe the temperature boundary condition (Carslaw & Jaeger Reference Carslaw and Jaeger1986; Beltman Reference Beltman1999),

(2.24) $$\begin{eqnarray}\hat{T}=-\unicode[STIX]{x1D6FC}_{w}\frac{\unicode[STIX]{x2202}\hat{T}}{\unicode[STIX]{x2202}n}.\end{eqnarray}$$

As $\unicode[STIX]{x1D6FC}_{w}\rightarrow 0$ , the boundary becomes isothermal. As $\unicode[STIX]{x1D6FC}_{w}\rightarrow \infty$ , the boundary becomes adiabatic.

The domain boundaries can have non-uniform compliance and thermal properties. The boundary impedance and thermal accommodation coefficients are non-uniform frequency-dependent functions, $Z=Z(s),\unicode[STIX]{x1D6FC}_{w}=\unicode[STIX]{x1D6FC}_{w}(s)$ on $\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}$ . In summary, the velocity and temperature boundary conditions can be generalized to Robin boundary conditions (2.23), (2.24), with special cases for rigid and open boundaries:

(2.25a,b ) $$\begin{eqnarray}Z=0\quad \text{on }\unicode[STIX]{x1D6E4}_{nsl},\quad Z^{-1}=0\quad \text{on }\unicode[STIX]{x1D6E4}_{free},\end{eqnarray}$$

(2.25c,d ) $$\begin{eqnarray}\unicode[STIX]{x1D6FC}_{w}=0\quad \text{on }\unicode[STIX]{x1D6E4}_{iso},\quad \unicode[STIX]{x1D6FC}_{w}^{-1}=0\quad \text{on }\unicode[STIX]{x1D6E4}_{ad}.\end{eqnarray}$$

2.4.2 Energy of the acoustic oscillation

The thermoviscous acoustic problem (2.15) is dissipative, and later we will investigate how and where this dissipation occurs. For this we introduce an energy norm (Chu Reference Chu1965), $\hat{E}$ , where $^{\ast }$ denotes the complex conjugate:

(2.26) $$\begin{eqnarray}\hat{E}=\int _{\unicode[STIX]{x1D6FA}}\hat{u} _{i}\hat{u} _{i}^{\ast }+\hat{\unicode[STIX]{x1D70C}}\hat{P}^{\ast }+{\hat{s}}\hat{T}^{\ast }\,\text{d}\boldsymbol{x},\end{eqnarray}$$

such that the total energy $\tilde{E}=\hat{E}\text{e}^{2st}$ decays in time as $\text{e}^{2\unicode[STIX]{x1D70E}t}$ (2.20b ). We premultiply the oscillating flow governing equation (2.20) by the state vector $\hat{\boldsymbol{q}}^{\text{T}}$ and integrate it over the volume, $\int _{\unicode[STIX]{x1D6FA}}s\hat{\boldsymbol{q}}^{\text{T}}B\hat{\boldsymbol{q}}+\hat{\boldsymbol{q}}^{\text{T}}A\hat{\boldsymbol{q}}=0$ . We integrate the second term by parts once and apply the boundary conditions (2.23), (2.24). The first term is the energy norm, $\int _{\unicode[STIX]{x1D6FA}}\hat{\boldsymbol{q}}^{\text{T}}B\hat{\boldsymbol{q}}=\hat{E}$ , and the second term is the energy dissipation inside the domain and the energy flux through the boundary. We take the real part of this volume integral to express the decay rate of the mode as the sum of volumetric energy dissipation $\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}}$ and the surface energy transfer $\unicode[STIX]{x1D70E}_{\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}}$ :

(2.27) $$\begin{eqnarray}\displaystyle \unicode[STIX]{x1D70E}\equiv \text{Re}(s) & = & \displaystyle \frac{1}{\hat{E}}\int _{\unicode[STIX]{x1D6FA}}-\frac{1}{\tilde{Re}}\hat{\unicode[STIX]{x1D70F}}_{ij}\unicode[STIX]{x1D735}_{j}\hat{u} _{i}^{\ast }-\frac{1}{(\unicode[STIX]{x1D6FE}-1)\tilde{Pe}}\Vert \unicode[STIX]{x1D735}_{i}\hat{T}\Vert ^{2}\,\text{d}\boldsymbol{x}\nonumber\\ \displaystyle & & \displaystyle +\,\frac{1}{\hat{E}}\int _{\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}}-\frac{\text{Re}(\unicode[STIX]{x1D6FC}_{w})}{(\unicode[STIX]{x1D6FE}-1)\tilde{Pe}}\Vert \frac{\unicode[STIX]{x2202}\hat{T}}{\unicode[STIX]{x2202}n}\Vert ^{2}+\text{Re}(Z)\Vert \hat{u} _{i}\Vert ^{2}\,\text{d}s\nonumber\\ \displaystyle & \equiv & \displaystyle \frac{1}{\hat{E}}\left(\int _{\unicode[STIX]{x1D6FA}}\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}}\,\text{d}\boldsymbol{x}+\int _{\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}}\unicode[STIX]{x1D70E}_{\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}}\,\text{d}s\right).\end{eqnarray}$$

The volumetric energy dissipation of the acoustic perturbation consists of viscous and thermal dissipation and is always negative, $\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}}\leqslant 0$ , while the surface energy transfer of the acoustic perturbation depends on the heat losses through the boundary and the work done by or on the fluid at the boundary. For the rigid and open boundary conditions (2.25) the surface energy transfer vanishes.

3.1 Shape gradients in Hadamard form

We consider a governing equation ${\mathcal{R}}(q,\boldsymbol{a})=0$ satisfied over a domain $\unicode[STIX]{x1D6FA}$ , with solution $q$ for parameters $\boldsymbol{a}$ . We define an objective function $J(q,\unicode[STIX]{x1D6FA},\boldsymbol{a})$ . The gradient of $J$ with respect to a parameter variation, $\unicode[STIX]{x1D6FF}\boldsymbol{a}$ , at $\boldsymbol{a}=\boldsymbol{a}_{0},q_{0}=q(\boldsymbol{a}_{0}),\unicode[STIX]{x1D6FA}_{0}=\unicode[STIX]{x1D6FA}(\boldsymbol{a}_{0})$ , is denoted with a square bracket $J^{\prime }[\unicode[STIX]{x1D6FF}\boldsymbol{a}]$ :

(3.1) $$\begin{eqnarray}J^{\prime }(q_{0},\unicode[STIX]{x1D6FA}_{0},\boldsymbol{a}_{0})[\unicode[STIX]{x1D6FF}\boldsymbol{a}]=\lim _{\unicode[STIX]{x1D709}\rightarrow 0^{+}}\frac{J(q(\boldsymbol{a}_{0}+\unicode[STIX]{x1D709}\unicode[STIX]{x1D6FF}\boldsymbol{a}),\unicode[STIX]{x1D6FA}(\boldsymbol{a}_{0}+\unicode[STIX]{x1D709}\unicode[STIX]{x1D6FF}\boldsymbol{a}))-J_{0}}{\unicode[STIX]{x1D709}}.\end{eqnarray}$$

In shape optimization, the parameters $\boldsymbol{a}$ also determine the domain boundary $\unicode[STIX]{x1D6E4}=\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}$ . In two dimensions, a displacement field $V:\mathbb{R}^{2}\rightarrow \mathbb{R}^{2}$ defined in $\unicode[STIX]{x1D6FA}$ represents the domain deformation, and $\unicode[STIX]{x1D709}$ is the displacement amplitude. We denote the perturbed domain as $\unicode[STIX]{x1D6FA}_{\unicode[STIX]{x1D709}}$ , and $q_{\unicode[STIX]{x1D709}}$ as the corresponding perturbed flow state. A perturbed boundary $\unicode[STIX]{x1D6E4}_{\unicode[STIX]{x1D709}}=\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}_{\unicode[STIX]{x1D709}}$ is given by

(3.2) $$\begin{eqnarray}\unicode[STIX]{x1D6E4}_{\unicode[STIX]{x1D709}}=\unicode[STIX]{x1D6E4}+\unicode[STIX]{x1D709}V(\boldsymbol{x})\quad \text{for }\boldsymbol{x}\in \unicode[STIX]{x1D6E4}.\end{eqnarray}$$

If the domain boundary $\unicode[STIX]{x1D6E4}$ is sufficiently smooth, any tangential displacement only changes the boundary parametrization but not the actual shape. Therefore the boundary displacements in the direction of $V$ and its normal component $(V\boldsymbol{\cdot }\boldsymbol{n})\boldsymbol{n}$ are equivalent, where $\boldsymbol{n}$ is the boundary unit normal vector.

The shape derivative of a general boundary condition independent of the geometry, in particular independent of the surface normal, can be calculated as follows. Given a boundary condition $\boldsymbol{g}(q_{0})=g_{0}$ on the unperturbed boundary $\unicode[STIX]{x1D6E4}_{0}$ , the perturbed boundary condition $\boldsymbol{g}(q_{\unicode[STIX]{x1D709}})=g_{\unicode[STIX]{x1D709}}$ on $\unicode[STIX]{x1D6E4}_{\unicode[STIX]{x1D709}}$ can be linearized around $\unicode[STIX]{x1D6E4}_{0}$ for a small shape deformation with magnitude $\unicode[STIX]{x1D709}\ll 1$ . We expand the perturbed solution as

(3.3) $$\begin{eqnarray}\displaystyle q_{\unicode[STIX]{x1D709}}(\unicode[STIX]{x1D6E4}_{\unicode[STIX]{x1D709}})\! & = & \displaystyle \!(1+\unicode[STIX]{x1D709}(V\boldsymbol{\cdot }\unicode[STIX]{x1D735}))q_{\unicode[STIX]{x1D709}}(\unicode[STIX]{x1D6E4}_{0})+O(\unicode[STIX]{x1D709}^{2})=(1+\unicode[STIX]{x1D709}(V\boldsymbol{\cdot }\unicode[STIX]{x1D735}))(q_{0}(\unicode[STIX]{x1D6E4}_{0})+\unicode[STIX]{x1D709}q_{0}^{\prime }[V](\unicode[STIX]{x1D6E4}_{0}))+O(\unicode[STIX]{x1D709}^{2})\nonumber\\ \displaystyle \! & = & \displaystyle \!q_{0}(\unicode[STIX]{x1D6E4}_{0})+\unicode[STIX]{x1D709}q_{0}^{\prime }[V](\unicode[STIX]{x1D6E4}_{0})+\unicode[STIX]{x1D709}(V\boldsymbol{\cdot }\unicode[STIX]{x1D735})q_{0}(\unicode[STIX]{x1D6E4}_{0})+O(\unicode[STIX]{x1D709}^{2}),\end{eqnarray}$$

such that the total derivative of the solution with respect to the shape perturbation $V$ is $\text{d}q[V]\equiv q_{0}^{\prime }[V](\unicode[STIX]{x1D6E4}_{0})+(V\boldsymbol{\cdot }\unicode[STIX]{x1D735})q_{\unicode[STIX]{x1D709}}(\unicode[STIX]{x1D6E4}_{0})$ and $q_{0}^{\prime }[V]$ is the local shape derivative. The linearization of the boundary condition is

(3.4) $$\begin{eqnarray}\boldsymbol{g}(q_{\unicode[STIX]{x1D709}}(\unicode[STIX]{x1D6E4}_{\unicode[STIX]{x1D709}}))=\boldsymbol{g}(q_{0}+\unicode[STIX]{x1D709}\,\text{d}q[V])=\boldsymbol{g}(q_{0},\unicode[STIX]{x1D6E4}_{0})+\unicode[STIX]{x1D709}\left(\left.\frac{\unicode[STIX]{x2202}\boldsymbol{g}}{\unicode[STIX]{x2202}q}\right|_{0}(q_{0}^{\prime }[V]+(V\boldsymbol{\cdot }\unicode[STIX]{x1D735})q_{0})\right)+O(\unicode[STIX]{x1D709}^{2}),\end{eqnarray}$$

where the subscript $|_{0}$ indicates the value at $q=q_{0},\unicode[STIX]{x1D6E4}=\unicode[STIX]{x1D6E4}_{0}$ . The term $(\unicode[STIX]{x2202}\boldsymbol{g}/\unicode[STIX]{x2202}q)|_{0}q^{\prime }[V]$ represents the boundary condition of the first-order solution’s response to shape deformation on the unperturbed boundary. It can be expressed in terms of the initial solution $q_{0}$ as

(3.5) $$\begin{eqnarray}\left.\frac{\unicode[STIX]{x2202}\boldsymbol{g}}{\unicode[STIX]{x2202}q}\right|_{0}q^{\prime }[V]=\lim _{\unicode[STIX]{x1D709}\rightarrow 0^{+}}\frac{g_{b,\unicode[STIX]{x1D709}}-g_{b,0}}{\unicode[STIX]{x1D709}}-\left.\frac{\unicode[STIX]{x2202}\boldsymbol{g}}{\unicode[STIX]{x2202}q}\right|_{0}(V\boldsymbol{\cdot }\unicode[STIX]{x1D735})q_{0}=(V\boldsymbol{\cdot }\unicode[STIX]{x1D735})g_{b,0}-\left.\frac{\unicode[STIX]{x2202}\boldsymbol{g}}{\unicode[STIX]{x2202}q}\right|_{0}(V\boldsymbol{\cdot }\unicode[STIX]{x1D735})q_{0}.\end{eqnarray}$$

A shape derivative $J^{\prime }(\unicode[STIX]{x1D6FA})[V]$ can be written in Hadamard form as a scalar product of a sensitivity functional $G(q,q^{+})$ and the normal component of the deformation field $V$ , where $q^{+}$ is the adjoint state:

(3.6) $$\begin{eqnarray}J^{\prime }[V]=\int _{\unicode[STIX]{x1D6E4}_{0}}(V\boldsymbol{\cdot }\boldsymbol{n})G(q,q^{+})\,\text{d}s.\end{eqnarray}$$

In the following sections, we discuss the choice of the objective functions for the steady and the oscillating flows, construct the adjoint states and derive the corresponding sensitivity functionals.

3.2 Steady flow shape sensitivity

For the steady flow, we wish to minimize the viscous dissipation, $J_{vd}$ , in the domain:

(3.7) $$\begin{eqnarray}J_{vd}(\bar{u},\unicode[STIX]{x1D6FA})=\int _{\unicode[STIX]{x1D6FA}}\frac{1}{\overline{Re}}(\unicode[STIX]{x1D735}_{j}\bar{u}_{i})^{2}\,\text{d}\boldsymbol{x},\end{eqnarray}$$

where $\bar{u}$ satisfies the momentum equation and the divergence-free condition given by (2.13), (2.14). As discussed by Schmidt & Schulz (Reference Schmidt and Schulz2010), the viscous dissipation sensitivity functional $G_{vd}$ for a shape displacement defined on a no-slip surface is

(3.8) $$\begin{eqnarray}G_{vd}(\bar{u}_{i},\unicode[STIX]{x1D706}_{i})=\frac{1}{\overline{Re}}\frac{\unicode[STIX]{x2202}\bar{u}_{i}}{\unicode[STIX]{x2202}n}\frac{\unicode[STIX]{x2202}(\unicode[STIX]{x1D706}_{i}-\bar{u}_{i})}{\unicode[STIX]{x2202}n},\end{eqnarray}$$

where $\unicode[STIX]{x1D706}_{i}$ and $\unicode[STIX]{x1D706}_{p}$ are the adjoint velocity and pressure states satisfying

(3.9a ) $$\begin{eqnarray}\displaystyle \bar{u}_{j}\unicode[STIX]{x1D735}_{j}\unicode[STIX]{x1D706}_{i}+\bar{u}_{j}\unicode[STIX]{x1D735}_{i}\unicode[STIX]{x1D706}_{j}+\unicode[STIX]{x1D708}\unicode[STIX]{x0394}\unicode[STIX]{x1D706}_{i}+\unicode[STIX]{x1D735}_{i}\unicode[STIX]{x1D706}_{p} & = & \displaystyle \frac{2}{\overline{Re}}\unicode[STIX]{x0394}\bar{u}_{i},\end{eqnarray}$$

(3.9b ) $$\begin{eqnarray}\displaystyle \unicode[STIX]{x1D735}_{i}\unicode[STIX]{x1D706}_{i} & = & \displaystyle 0,\quad \text{in }\unicode[STIX]{x1D6FA},\end{eqnarray}$$

(3.9c ) $$\begin{eqnarray}\displaystyle \unicode[STIX]{x1D706}_{i} & = & \displaystyle 0\quad \text{on }\unicode[STIX]{x1D6E4}_{in}\cup \unicode[STIX]{x1D6E4}_{w},\end{eqnarray}$$

(3.9d ) $$\begin{eqnarray}\displaystyle \unicode[STIX]{x1D706}_{i}\bar{u}_{j}n_{j}+\unicode[STIX]{x1D706}_{j}\bar{u}_{j}n_{i}+\unicode[STIX]{x1D708}\frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D706}_{i}}{\unicode[STIX]{x2202}n}+\unicode[STIX]{x1D706}_{p}n_{i} & = & \displaystyle \frac{2}{\overline{Re}}\frac{\unicode[STIX]{x2202}\bar{u}_{i}}{\unicode[STIX]{x2202}n}\quad \text{on }\unicode[STIX]{x1D6E4}_{out}.\end{eqnarray}$$

3.3 Oscillating flow shape sensitivity

For the oscillating flow we wish to control the decay rate and frequency (Luchini & Bottaro Reference Luchini and Bottaro2014), so the objective function is the complex natural frequency, $s$ , of the thermoviscous acoustic flow (2.20):

(3.10) $$\begin{eqnarray}J_{s}=s.\end{eqnarray}$$

We introduce an adjoint state vector $\boldsymbol{q}^{+}=(P^{+},u_{i}^{+},T^{+})$ containing the adjoint pressure, velocity and temperature variables. Taking the inner product of the direct equations and the corresponding adjoint variables, we construct a Lagrangian of the system (Gunzburger Reference Gunzburger2002),

(3.11) $$\begin{eqnarray}{\mathcal{L}}=s-\langle \boldsymbol{q}^{+},sB\hat{\boldsymbol{q}}+A\hat{\boldsymbol{q}}\rangle .\end{eqnarray}$$

The optimality condition sets any first Lagrangian variation to zero. Variation with respect to the adjoint and direct variables gives the direct and the adjoint state equations, respectively. As discussed in § A.1, the adjoint and the direct states of the thermoviscous acoustic problem are related by $P^{+}=\hat{P}^{\ast },u_{i}^{+}=-\hat{u} _{i}^{\ast },T^{+}=\hat{T}^{\ast }$ , subject to the normalization condition:

(3.12) $$\begin{eqnarray}\displaystyle 1 & = & \displaystyle \langle u_{i}^{+},\hat{u} _{i}\rangle +\langle P^{+},\unicode[STIX]{x1D6FE}\hat{P}-\hat{T}\rangle +\left\langle T^{+},\frac{\hat{T}}{\unicode[STIX]{x1D6FE}-1}-\hat{P}\right\rangle +\left\{u_{i}^{+},\frac{\unicode[STIX]{x2202}Z}{\unicode[STIX]{x2202}s}\hat{u} _{i}\right\}\nonumber\\ \displaystyle & & \displaystyle -\left\{\frac{\unicode[STIX]{x2202}T^{+}}{\unicode[STIX]{x2202}n},\frac{(\unicode[STIX]{x2202}\unicode[STIX]{x1D6FC}_{w}/\unicode[STIX]{x2202}s)}{(\unicode[STIX]{x1D6FE}-1)\tilde{Pe}}\frac{\unicode[STIX]{x2202}\hat{T}}{\unicode[STIX]{x2202}n}\right\},\end{eqnarray}$$

where we define the volume inner product $\langle a^{+},b\rangle =\int _{\unicode[STIX]{x1D6FA}}(a^{+})^{\ast }b\,\text{d}\boldsymbol{x}$ (A 3a ) and the surface inner product $\{a^{+},b\}=\int _{\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}}(a^{+})^{\ast }b\,\text{d}s$ (A 3b ).

For a shape deformation normal to a boundary, the oscillating flow eigenvalue sensitivity $G_{s}$ consists of the surface stress and the thermal terms, $G_{s}=G_{s}^{str}+G_{s}^{th}$ (derived in § A.2). Given that the direct and adjoint states are identical up to the sign of the velocity term, the sensitivity functionals are:

(3.13) $$\begin{eqnarray}\left.\begin{array}{@{}c@{}}G_{s}^{str}=-\left(2{\displaystyle \frac{\unicode[STIX]{x2202}\hat{u} _{i}}{\unicode[STIX]{x2202}n}}n_{j}\hat{\unicode[STIX]{x1D70E}}_{ij}+\unicode[STIX]{x1D705}\hat{u} _{i}\hat{\unicode[STIX]{x1D70E}}_{ij}n_{j}-\unicode[STIX]{x1D735}_{j}(\hat{u} _{i}\hat{\unicode[STIX]{x1D70E}}_{ij})\right),\\ G_{s}^{th}=2{\displaystyle \frac{\unicode[STIX]{x2202}\hat{T}}{\unicode[STIX]{x2202}n}}\hat{q}_{n}+\unicode[STIX]{x1D705}\hat{T}\hat{q}_{n}-\unicode[STIX]{x1D735}_{j}(\hat{T}\hat{q}_{j}),\end{array}\right\}\end{eqnarray}$$

where $\hat{q}_{i}\equiv ((\unicode[STIX]{x1D6FE}-1)\tilde{Pe})^{-1}\unicode[STIX]{x1D735}_{i}\hat{T}$ is the boundary heat flux, and $\hat{q}_{n}=(\hat{q}\boldsymbol{\cdot }\boldsymbol{n})$ is its normal component. The viscous and the thermal sensitivity functionals have equivalent structure in terms of the $(\hat{u} _{i},\hat{\unicode[STIX]{x1D70E}}_{ij})$ and $(\hat{T},\hat{q}_{i})$ pairs.

On the no-slip and stress-free boundaries, the viscous sensitivity functional simplifies to

(3.14a ) $$\begin{eqnarray}\displaystyle G_{s,nsl}^{str} & = & \displaystyle -\frac{\unicode[STIX]{x2202}\hat{u} _{i}}{\unicode[STIX]{x2202}n}n_{j}\hat{\unicode[STIX]{x1D70E}}_{ij},\end{eqnarray}$$

(3.14b ) $$\begin{eqnarray}\displaystyle G_{s,free}^{str} & = & \displaystyle \unicode[STIX]{x1D735}_{j}(\hat{u} _{i}\hat{\unicode[STIX]{x1D70E}}_{ij}),\end{eqnarray}$$

(3.15a ) $$\begin{eqnarray}\displaystyle G_{s,iso}^{th} & = & \displaystyle \frac{\unicode[STIX]{x2202}\hat{T}}{\unicode[STIX]{x2202}n}\hat{q}_{n},\end{eqnarray}$$

(3.15b ) $$\begin{eqnarray}\displaystyle G_{s,ad}^{th} & = & \displaystyle -\unicode[STIX]{x1D735}_{j}(\hat{T}\hat{q}_{j}).\end{eqnarray}$$

4.1 Numerical methods

4.1.1 Optimization domain

We start with a flow in a two-dimensional uniform-width channel, defined as

(4.1) $$\begin{eqnarray}\unicode[STIX]{x1D6FA}_{0}=\{(x,y)\in \mathbb{R}^{2}\mid [0,1]\times [0,0.1]\},\end{eqnarray}$$

with inlet and outlet boundaries

(4.2a ) $$\begin{eqnarray}\displaystyle \unicode[STIX]{x1D6E4}_{in,0} & = & \displaystyle \{(x,y)\in \mathbb{R}^{2}\mid x=0\},\end{eqnarray}$$

(4.2b ) $$\begin{eqnarray}\displaystyle \unicode[STIX]{x1D6E4}_{out,0} & = & \displaystyle \{(x,y)\in \mathbb{R}^{2}\mid x=1\},\end{eqnarray}$$

and no-slip boundaries $\unicode[STIX]{x1D6E4}_{w,0}=\unicode[STIX]{x2202}\unicode[STIX]{x1D6FA}_{0}\backslash (\unicode[STIX]{x1D6E4}_{in,0}\cup \unicode[STIX]{x1D6E4}_{out,0})$ . For the oscillating flow, the stress-free boundary is $\unicode[STIX]{x1D6E4}_{free,0}=\unicode[STIX]{x1D6E4}_{in,0}\cup \unicode[STIX]{x1D6E4}_{out,0}$ .

The boundary $\unicode[STIX]{x1D6E4}_{w,0}$ is to be optimized by modifying the no-slip boundaries, while fixing the inlet and the outlet. If equivalent boundary displacement fields are applied to the top and the bottom no-slip surfaces then the steady flow and the oscillating flow boundary sensitivities and the shape gradients remain symmetric. Therefore we may consider deformation of only the top boundary.

In this study, we parametrize the boundary with a set of $N$ control points $\{a^{k}\in \mathbb{R}^{2}\mid k=1\ldots N\}$ defining the third-order rational uniform B-spline curve. This provides a smooth surface of class $C^{2}$ for which parametric sensitivities can be calculated (Samareh Reference Samareh2001). Boundary displacement fields, $V^{k}$ , are, by definition, the boundary shape sensitivities to the control point positions,

(4.3) $$\begin{eqnarray}V^{k}=\frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D6E4}_{w}}{\unicode[STIX]{x2202}a^{k}},\end{eqnarray}$$

which implies that the objective function gradient with respect to the displacement field $J^{\prime }[V^{k}]$ transforms to the sensitivity with respect to the control parameters, $J^{\prime }[a^{k}]$ . As the positions of the control points are moved in the gradient direction, the domain is updated and the computational mesh is rebuilt.

We parametrize the top boundary of the initial rectangular domain with 11 control points, $a_{i}=(i/10,0.1)$ , $i=0\ldots 10$ , spaced uniformly at intervals of $0.1$ . The first and last points are kept at their initial positions so that the inlet and outlet boundaries are fixed and the channel’s length remains equal to 1.

4.1.3 Constrained gradient optimization

Our goals are to decrease the steady flow viscous dissipation and make the oscillation’s decay rate, $-\unicode[STIX]{x1D70E}$ , more negative. There is a trade-off between these goals, so here we set the steady flow dissipation as an inequality constraint and minimize $\text{Re}(s)$ .

(4.4a ) $$\begin{eqnarray}\displaystyle \min _{\unicode[STIX]{x1D6FA}} & & \displaystyle \quad \text{Re}(s)\end{eqnarray}$$

(4.4b ) $$\begin{eqnarray}\displaystyle \text{subject to} & & \displaystyle \quad J_{vd}\leqslant J_{vd}^{0}\end{eqnarray}$$

(4.4c ) $$\begin{eqnarray}\displaystyle \text{and} & & \displaystyle \quad \text{state equations }(2.13),(2.20).\end{eqnarray}$$

In § A.1 we derive shape gradients in Hadamard form of both the objective function and the constraint with respect to an arbitrary boundary displacement $V$ . The parameter-based approach restricts the number of possible boundary deformations to the number of control parameters. Essentially, a parametrization projects the function space of the admissible shape deformations onto a lower-dimensional subspace, which allows us to operate with the vector representation of shape gradients instead of the continuous boundary sensitivities.

We can estimate the optimality of a shape (but not the parametrization) by considering the scalar product of the objective shape sensitivity, $G$ , with the constraint shape sensitivity, $G^{\prime }$ . The surface inner product $\{G,G^{\prime }\}$ (A 3) and the surface norm $\Vert G\Vert _{\unicode[STIX]{x1D6E4}}^{2}=\{G,G\}$ form the optimality coefficient $\unicode[STIX]{x1D6FC}$ :

(4.5) $$\begin{eqnarray}\unicode[STIX]{x1D6FC}=\frac{\{G,G^{\prime }\}}{\Vert G\Vert _{\unicode[STIX]{x1D6E4}}\Vert G^{\prime }\Vert _{\unicode[STIX]{x1D6E4}}}\geqslant -1.\end{eqnarray}$$

The optimality coefficient indicates the cosine of the angle between the objective function shape sensitivity and the constraint shape sensitivity, such that $\unicode[STIX]{x1D6FC}=-1$ implies that they point in opposite directions and the system has reached its local optimum. In the case of $N$ optimization parameters, a sensitivity functional is realized on a shape deformation subspace, spanned by $V^{k}$ , $k=1\ldots N$ . The discrete gradient vector $\boldsymbol{g}\in \mathbb{R}^{N}$ is defined as $\boldsymbol{g}_{k}=\{V^{k},G\}$ , and the parametric optimality coefficient $\unicode[STIX]{x1D6FC}_{p}$ is:

(4.6) $$\begin{eqnarray}\unicode[STIX]{x1D6FC}_{p}=\frac{\boldsymbol{g}^{\text{T}}\boldsymbol{g}^{\prime }}{\Vert \boldsymbol{g}\Vert \boldsymbol{\cdot }\Vert \boldsymbol{g}^{\prime }\Vert }\leqslant \unicode[STIX]{x1D6FC}.\end{eqnarray}$$

In the parameter-based optimization, $\unicode[STIX]{x1D6FC}_{p}=-1$ implies that the local optimum has been reached within the choice of parametrization.

The optimization algorithm for the problem (4.4) is based on the method of moving asymptotes (Svanberg Reference Svanberg1987). The objective and the constraint values and their gradients are calculated by (i) solving the steady flow (2.13) and the oscillating eigenvalue (2.20) problems, (ii) finding the adjoint steady flow (3.9) and oscillating flow (3.12) states, (iii) calculating the boundary sensitivities $G_{vd},G_{s}$ using (3.8) and (3.13), and computing the objective and constraint gradients with respect to the boundary control points $s^{\prime }[a_{k}]=\{V^{k},G_{s}\},J_{vd}^{\prime }[a_{k}]=\{V^{k},G_{vd}\}$ . We use $\unicode[STIX]{x1D700}_{p}=\unicode[STIX]{x1D6FC}_{p}+1$ as a tolerance criteria for the optimization process.

4.2 Optimization results

4.2.1 Initial domain

The steady flow is computed in the initially flat channel (4.1) at $\overline{Re}=0.1$ with a parabolic inflow velocity profile. In the unaltered domain this results in the Poiseuille flow solution, and the corresponding viscous dissipation value $J_{vd}^{0}$ is taken as a reference. For the oscillating flow at $\tilde{Re}=1000$ , we choose the smallest non-zero-frequency natural mode as the target mode, with $s_{0}=-0.555+2.81\text{i}$ . Figure 1 shows the real and imaginary parts of the mode shape, normalized by (3.12). The pressure gradient $\unicode[STIX]{x2202}_{x}\hat{P}$ and the longitudinal velocity $\hat{u} _{x}$ are highest on the stress-free open end boundaries at $x=0,x=1$ . As indicated in figure 2, the regions with the highest contribution to the decay rate $\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}}$ are the no-slip wall regions close to the open ends, where the velocity magnitude isolines converge and therefore the transverse velocity gradient is the largest. For the initial channel configuration, $\unicode[STIX]{x1D6FC}_{p}=-0.7$ .

Figure 1.

First natural mode of the oscillating flow in a flat channel at $\tilde{Re}=1000$ . (a–h) Pressure $\hat{P}$ , longitudinal $\hat{u} _{x}$ and transverse $\hat{u} _{y}$ velocity components, and temperature $\hat{T}$ mode shapes; real (a,c,e,g) and imaginary (b,d,f,h) parts.

Figure 2.

Spatial distribution of the decay rate production $\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}}$ in a flat channel. Black lines correspond to the oscillating flow velocity magnitude isolines $\hat{u} =\text{const}.$

The steady flow state in the unperturbed channel is independent of the longitudinal coordinate $x$ so the viscous dissipation shape sensitivity $G_{vd}(\bar{u}_{i},\unicode[STIX]{x1D706}_{i})$ is constant along the no-slip walls (figure 3). Here, and later, $G_{vd}$ is normalized by the viscous dissipation value $J_{vd}^{0}$ in the starting geometry configuration. The shape sensitivity $G_{vd}$ is always negative, so any boundary displacement resulting in contraction of the channel’s width leads to growth of viscous dissipation. The complex eigenvalue shape sensitivity $G_{s}(\hat{P},\hat{u} _{i},\hat{T})$ is not uniform; the real part $\text{Re}(G_{s})$ is almost zero in the middle part of the boundary and grows towards the channel’s open ends where the decay rate production is highest, as shown previously. As for viscous dissipation, any shape deformation directed inwards $(V\boldsymbol{\cdot }\boldsymbol{n})<0$ leads to an increase in the decay rate magnitude.

Figure 3.

Shape sensitivity distribution along the flat channel top boundary for the decay rate $\text{Re}(G_{s})$ (solid line) and frequency $\text{Im}(G_{s})$ (dashed line) of the first oscillating mode, and the steady flow viscous dissipation shape sensitivity $G_{vd}/J_{0}$ normalized by viscous dissipation inside the channel (dotted line).

Figure 4.

Steady flow in the optimized channel at $\overline{Re}=0.1$ . (a) The direct steady flow $|\bar{u}|$ velocity magnitude, with the streamlines indicated (solid lines). (b) The adjoint steady flow $|\unicode[STIX]{x1D706}|$ velocity magnitude.

As indicated in figure 3, the decay rate is less sensitive to shape modifications in the middle region of the channel at $0.3\leqslant x\leqslant 0.7$ , and has higher sensitivity on the outer region. We expect therefore, that the channel will expand in the middle and shrink around the free boundaries to increase the decay rate while keeping the steady flow viscous dissipation constant. This is also what we expect on physical grounds: the channels will constrict where the acoustic velocity is greater.

4.2.2 Optimized domain

The optimized configuration is found in 20 iterations and has the optimality coefficient (4.5) of $\unicode[STIX]{x1D6FC}_{p}=-0.98$ . The first eigenmode in the optimized channel is $s=-1.31+1.68i$ . In comparison to the initial solution, the decay rate objective function changes by almost 140 %, and the frequency (which is unconstrained) decreases by 40 %. The total area almost doubles and the channel’s shape loses symmetry around the $x=0.5$ vertical plane, while remaining symmetric in the horizontal plane. The channel constricts near $x=0.07$ and $x=0.99$ , and the middle part of the channel expands, as expected.

Figure 4 shows the steady flow (a) and the adjoint flow (b) velocity magnitude $\bar{u}$ and $\unicode[STIX]{x1D706}$ for the optimized channel. The lines correspond to the steady flow streamlines. The no-slip boundaries are smooth and the flow remains attached to the walls with no recirculation zones. Viscous dissipation in the optimized channel is the same as in the initial channel.

The steady flow velocity amplitude and velocity gradients as well as the adjoint velocity are highest in the constricted areas. This makes the constricted regions much more sensitive to shape changes than the expanded part, where the adjoint velocity magnitude is almost zero.

The decay rate production is initially located in the corner regions of the uniform-width channel. When the boundaries shift, this region shifts inside the channel towards the constrictions, as indicated in figure 5. The decay rate production strongly concentrates in the narrow parts of the channel, with the maximum at $x=0.99$ more than 10 000 times higher than the average value. It is almost zero between the constrictions.

Figure 5.

Spatial distribution of the absolute value of the decay rate production in the optimized channel (on a logarithmic scale), $\ln (-\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}})$ . Black lines correspond to the oscillating flow velocity magnitude isolines $\hat{u} =\text{const.}$

Figure 6.

The decay rate shape sensitivity distribution (solid line) of the first oscillating mode in the optimized channel, and the steady flow viscous dissipation sensitivity (dotted line) normalized by viscous dissipation inside the channel.

Figure 6 illustrates the viscous dissipation $G_{vd}/J_{vd}^{0}$ (dash-dotted line) and the decay rate $\text{Re}(G_{s})/\unicode[STIX]{x1D714}_{0}$ (solid line) boundary sensitivities as functions of the longitudinal coordinate along the top boundary in the optimized channel. Both sensitivities reach their extreme values in the constricted areas and are much smaller in the intermediate region. They are almost equal and opposite to each other, showing that the design is almost optimal. Further improvements can still be made, for instance, by boundary re-parametrization or by introducing additional control points. However, this simple problem has achieved its purpose by showing that the optimization procedure can indeed increase acoustic dissipation while keeping the steady flow dissipation constant.

5.1 Generic geometry and shape parametrization

Figure 7.

A 2-D generic printhead geometry with a piezo-electric actuator. The channel is connected to the ink supply manifolds via the inlet and outlet boundaries. Circles denote the B-spline boundary control points. All sizes are in $\unicode[STIX]{x03BC}\text{m}$ .

Figure 7 shows a 2-D generic inkjet printhead chamber, which consists of a vertical inlet and outlet, connected to ink manifolds, and a horizontal main channel. The manifold cross-sections are much larger than the printhead cross-section. A $30~\unicode[STIX]{x03BC}\text{m}$ long conical printing nozzle, which has a $20~\unicode[STIX]{x03BC}\text{m}$ outer diameter and a taper of $8^{\circ }$ , is located in the middle of the printhead. A flat piezo-electric membrane is located on the top boundary opposite the nozzle. In three dimensions, the channel has a depth of $60~\unicode[STIX]{x03BC}\text{m}$ into the page. For this study, we approximate the channel to be uniform in that direction and examine only 2-D deformations, as in § 4.2. We aim to increase the decay rate of the oscillating flow while keeping the steady flow viscous dissipation constant.

We parametrize the printhead walls by third-order B-splines with the control points indicated in figure 7. The inlet and the outlet points are fixed. The nozzle shape cannot change but it can move in the vertical direction. The bottom wall cannot extend below the nozzle tip.

We choose a characteristic length $L=100~\unicode[STIX]{x03BC}\text{m}$ . The steady flow Reynolds number is $\overline{Re}=0.066$ and the Reynolds number based on the speed of sound is $\tilde{Re}=6000$ . The steady flow Mach number is $\unicode[STIX]{x1D707}=10^{-4}$ and the oscillating flow Mach number is $\unicode[STIX]{x1D716}=10^{-5}$ . For the steady flow, the inlet has a fixed parabolic velocity profile, the outlet is an open end with stress-free boundary condition (2.14c ) and the walls are no-slip boundaries and the nozzle exit is modelled as a no-slip boundary because there is no flow through it. For the acoustic flow, the walls are adiabatic no slip, and the open boundaries, including the nozzle exit, are stress free and isothermal. In this study we neglect the surface tension at the nozzle exit.

Figure 8 shows the steady flow direct $\bar{u}$ and adjoint $\unicode[STIX]{x1D706}$ velocity magnitude in the initial geometry. The largest direct velocity magnitude is in the narrow horizontal channel. The adjoint velocity has highest value near the sharp corners at the channel entrance and the nozzle. These regions have the greatest influence on the steady flow viscous dissipation.

Figure 8.

The direct velocity magnitude $\bar{u}$ (a) and the adjoint velocity magnitude $\unicode[STIX]{x1D706}$ (b) of the steady flow in the initial printhead channel at $\overline{Re}=0.066$ .

Figure 9.

The first natural mode of the oscillating flow in the initial printhead geometry at $\tilde{Re}=6000$ . (a,b) The magnitude of the velocity mode real part $\text{Re}(\hat{u} )$ in the entire domain (a) and near the nozzle (b); (c,d) the magnitude of the velocity mode imaginary part $\text{Im}(\hat{u} )$ ; (e) pressure mode $\hat{P}$ ; and (f) temperature mode $\hat{T}$ .

The frequency of the first natural mode is $\text{Im}(s_{1}/2\unicode[STIX]{x03C0})=\unicode[STIX]{x1D714}/2\unicode[STIX]{x03C0}=0.342~\text{MHz}$ , and the decay rate is $\text{Re}(s_{1}/2\unicode[STIX]{x03C0})=\unicode[STIX]{x1D70E}/2\unicode[STIX]{x03C0}=0.0171~\text{MHz}$ . Figure 9 shows the mode shape, normalized by (3.12). The pressure and the temperature modes are zero on the stress-free boundaries and have antinodes in the middle of the channel. Since the walls are adiabatic, the thermal boundary layer is absent and the pressure and temperature gradients are tangential to the boundaries. The velocity magnitude is highest near the nozzle, as shown in figure 9, where the viscous boundary layers overlap. The decay rate production is shown in figure 10. It is concentrated in the nozzle region of the initial printhead configuration, in the viscous boundary layers along the no-slip walls, and around the corners.

Figure 10.

Spatial distribution of the absolute value of the decay rate production in the initial printhead channel (on a logarithmic scale), $\ln (-\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}})$ .

The parametric optimality coefficient for the initial printhead design is $\unicode[STIX]{x1D6FC}_{p}=0.017$ , which implies that the decay rate and the viscous dissipation gradient vectors are almost orthogonal. Therefore we expect to be able to obtain a noticeable improvement in the objective function.

5.2 Optimization

Figure 11.

The initial (solid lines) and the optimized (dashed lines) printhead geometry. The inlet and the outlet boundaries remain fixed. The piezo-electric membrane and nozzle parts can move up and down but cannot change shape.

In this section we use the optimization algorithm in § 4.1.3 to update the control points until the relative improvement falls below the tolerance level of $\unicode[STIX]{x1D700}_{p}\leqslant 0.1$ . The optimized domain is shown in figure 11. The channel constricts near the corners of the top boundary, where the decay rate production had a local maximum. These constrictions increase the steady flow viscous dissipation there, but the central part of the channel expands to compensate. The new frequency of the first natural mode is $\text{Im}(s_{1}/2\unicode[STIX]{x03C0})=\unicode[STIX]{x1D714}/2\unicode[STIX]{x03C0}=0.264~\text{MHz}$ , and the decay rate increases to $\text{Re}(s_{1}/2\unicode[STIX]{x03C0})=\unicode[STIX]{x1D70E}/2\unicode[STIX]{x03C0}=0.0257~\text{MHz}$ , which is over 50 % higher than before. The steady flow viscous dissipation is the same as in the initial printhead. The parametric optimality coefficient of the optimized shape is $\unicode[STIX]{x1D6FC}_{p}=-0.94$ , showing that it is nearly optimal.

Figure 12 shows the spatial distribution of the decay rate production $\ln (-\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}})$ inside the optimized domain. The viscous boundary layers overlap in the narrow parts of the channel, resulting in higher acoustic energy dissipation there. The highest amplitudes of decay rate production are around the nozzle. This shows that changes to the nozzle geometry are particularly influential.

Figure 12.

Spatial distribution of the decay rate production absolute value in the optimized printhead channel (on a logarithmic scale), $\ln (-\unicode[STIX]{x1D70E}_{\unicode[STIX]{x1D6FA}})$ .