Regarding the CFD computations, which provided the inputs for the following optimization process, a time-accurate solver was used to predict the pressure fluctuations due to the rotor/stator interactions on the rotor surface. An URANS approach was preferred, given that high-fidelity large eddy simulations (LES) are still too computationally expensive and not yet compatible with industrial design timing. In this regard, the URANS TRAF code, developed at the University of Florence, was used to solve the classical Reynolds averaged Navier–Stokes equations in order to extract the aerodynamic forcing.
The compact form of fluid motion equations can be written for a fixed blade in the following conservative form
where
U includes the conservative variables for mass, momentum, energy, and for turbulence quantities, when needed,
Q is the source term column, while
describes the three flux functions. These vectors are given by
In the previous expressions, turbulence quantities have been omitted for brevity.
x,
y, and
z are the orthogonal Cartesian coordinates, while
,
, and
are the velocity components. Moreover,
t is the time,
the density, and
p the pressure. The specific enthalpy is defined as
and
u is the internal energy. Quantity
is defined as
According to the Stokes hypothesis the viscous stress tensor and the thermal flux density are defined by
To mathematically close the problem, the state equation and the expressions of
and
k are needed. The ideal gas state equation can be added:
where
T is the temperature, and
represents the gas constant. Regarding the expressions of
and
k, the turbulence phenomenon must be considered:
where
is the dynamic viscosity,
is the turbulent viscosity and
is the constant-pressure specific heat.
and
are the Prandtl and the turbulent Prandtl numbers, respectively. Dynamic viscosity is a function of the thermodynamic state, whereas turbulent viscosity depends on the chosen turbulence model. The TRAF code solves the unsteady, three-dimensional, Reynolds averaged Navier–Stokes equations for internal flow, with a finite volume formulation, on H-type and O-type structured grids. The discretization of convective fluxes is based on a 2nd order TVD-MUSCL strategy built on Roe’s upwind scheme [
32], while viscous fluxes are computed with a central difference scheme. The turbulence closure is based on different models ranging from one to four equations. Regarding the unsteady simulations, these are performed by means of a dual time-stepping technique with sliding interfaces between adjacent rows, using linear interpolations to couple flow variables on the block interfaces. An important feature of the code is the high level of parallelization, achieved by means of a hybrid openMP/MPI code architecture and GPU acceleration [
25]. A key feature for aeromechanical evaluations is the possibility to activate a runtime time discrete Fourier decomposition when the unsteady solution has reached the periodicity in order to extract the complex forcing only at the engine orders (EOs) of interest.
5.1. Numerical Setup
Starting from the same CAD models used for the FEM analyses, the CFD domain was easily defined and consisted of a full annulus, multi-row environment with IGV, rotor, and stator rows. An in-house mesh generator, suitable for both H-grids and O-grids, was used to build a viscous grid with wall refinement to better capture viscous effects in the boundary layer, ensuring values of
, adopted following the industrial practice for this kind of computation. In this case, an O-grid type was selected to obtain the overall multi-block grid shown in
Figure 6. First, steady-state CFD simulations were performed to match the compressor operating point and evaluate mean flow quantities. This type of analysis requires only a single tangential sector for each blade row, taking advantage of the mixing plane approach to reduce computational cost. All the simulations used the
turbulence model in Wilcox’s formulation for the Navier–Stokes equation closure.
Inlet, outlet, walls, and periodicity boundary conditions for internal flows were imposed at the computational domain boundaries: at the inlet, span-wise distributions of total pressure, total temperature, blade-to-blade flow angle, and meridian flow angle were imposed, while at the outlet a span-wise distribution of static pressure was assigned. In steady analyses, the mixing plane approach was used to average the flow quantities passing from static to rotating blocks and vice versa, allowing the evaluation of the overall increase in static (and dynamic) pressure. Because of the averaging process, in this case, blades could not experience the wakes of upstream rows and other unsteady effects; thus, an unsteady multi-row analysis was performed to investigate stator–rotor interactions.
Unsteady multi-row analyses can be performed with both reduced-sector or full-annulus approaches, with the main aim to obtain a converged solution at each physical time step and an overall periodicity in time. For that case, the blade rows had a prime number of blades with respect to each other (15, 21, and 29), so a reduced-sector approach could not be adopted, and a full-annulus domain consisting of a 159-block domain was used (
Figure 6). This allowed the evaluation of the unsteady pressure field experienced by the rotor, caused by the potential effect of downstream rows, together with the wakes generated by upstream rows. It must be noted that while other different mechanisms of unsteady excitation can occur, in this paper, only the interactions between IGV and stator with the rotor were considered. These unsteady phenomena resulted in a pressure field fluctuation experienced by the rotor blades through a periodic change in the blade loading. These perturbations were then processed by a DFT algorithm of the unsteady pressure field over the blade surface, extracting the pressure harmonics linked to the most relevant engine orders, and stored in terms of real and imaginary parts of the main pressure harmonics.
5.2. Unsteady Forcing Decomposition
As previously stated, the TRAF code already performs at runtime a first discrete Fourier decomposition in time of the unsteady pressure history on selected blades, given the desired engine orders (EOs), as follows:
where EO is the engine order,
is the discrete, equally spaced pressure signal in time, and
is the total number of samples that discretize a single rotor revolution. The unsteady forcing resulting at the end of each unsteady run was then further Fourier decomposed in space, in the circumferential direction, to obtain a single traveling wave exciting a corresponding mode shape in a cyclic environment, as predicted by the interference diagram (
Figure 7). To do so, the Fourier coefficients in time on the blade surface were circumferentially decomposed on homologous points on adjacent blades (one on each blade, for a row pitch) by means of the in-house tool named OutDFT [
33] with the following formula:
where nd is the circumferential order,
is the discrete-time Fourier coefficient of the tangential distribution on corresponding surface points of the blade, and
N is the blade count. The method also takes into account the aliasing phenomenon, as the resulting rotating forcing function is characterized by a nodal diameter ranging from
to
, or from
to
in the case of an odd blade count, as suggested by Nyquist’s theorem. The sign of nd defines the spinning direction: if positive, it defines a forward-rotating wave; if negative, a backward-rotating wave. The final outputs of the tool are complex pressure maps on the blade surface related to single rotating forcings. Such distributions account for the static (average pressure) and dynamic parts and can be directly loaded into the commercial FE software for the next topology optimization setup or used for a forced-response assessment inside the same environment.