2018-01-0974

Adaptive Turbo Matching: Radial Turbine Design Optimization through 1D Engine

Simulations with Meanline Model in-the-Loop

P.Kapoor and A.Costall - Imperial College London

N.Sakellaridis, R.Lammers and J.Hooijer - Mitsubishi Turbocharger and Engine Europe B.V.
S. Guilain and H.Tartoussi – Groupe Renault

Abstract by superimposing the target engine operation on characteristic

compressor and turbine performance maps. The compressor options
that provide adequate surge and choke margins while achieving high
Turbocharging has become the favored approach for downsizing
efficiency are selected first. Then the turbine options that meet the
internal combustion engines to reduce fuel consumption and CO2
compressor power requirement with high efficiency and low inertia
emissions, without sacrificing performance. Matching a turbocharger
are chosen. Once this initial selection of potential compressor and
to an engine requires a balance of various design variables in order to
turbine options has been made, the corresponding maps are evaluated
meet the desired performance. Once an initial selection of potential
in 1D engine cycle simulations (e.g., GT-POWER) to further down-
compressor and turbine options is made, corresponding performance
select the turbocharger match that is predicted to best meet customer
maps are evaluated in 1D engine cycle simulations to down-select the
requirements.
best combination. This is the conventional matching procedure used
in industry and is ‘passive’ since it relies on measured maps, thus
only existing designs may be evaluated. In other words, turbine This is the typical matching procedure used in the automotive
characteristics cannot be changed during matching so as to explore industry, which takes place in the beginning of a turbocharger
the effect of design adjustments. Instead, this paper presents an project. It may be reviewed and updated as more information
‘adaptive’ matching methodology for the turbocharger turbine. By becomes available (e.g., customer engine test data for model
coupling an engine cycle simulation to a turbine meanline model (‘in- validation), but nonetheless sets the initial design direction and thus
the-loop’), adjustments in turbine geometry are reflected in both the has important consequences for all that follows. The authors describe
exhaust boundary conditions and overall engine performance. it as a ‘passive’ approach in the sense that it relies on knowing the
Running the coupled engine-turbine model within an optimization aforementioned maps that have been previously measured on the hot
framework, the optimal turbine design evolves. The methodology is gas bench, i.e., only existing turbocharger designs may be evaluated
applied to a Renault 1.2L turbocharged gasoline engine, to minimize against engine project requirements. This means the performance
fuel consumption over given full- and part-load operating points, characteristics of the selected turbine, for example, cannot be
while meeting performance constraints. Despite the current series changed during the matching simulation in order to explore whether
production turbine being a very good match already, and with small geometrical changes would better match the particular engine
optimization restricted to a few turbine geometric parameters, the flow conditions. On the other hand, while 3D CFD can be used to
full-load case predicted a significant cycle-averaged BSFC reduction optimize a turbine aerodynamic design for given engine flow
of 3.5 g/kWh, while the part-load optimized design improved BSFC boundary conditions, this approach would not be consistent due to the
by 0.9 g/kWh. No engine design parameters were changed, so further highly-coupled nature of the engine-turbocharger system, i.e., any
efficiency gains would be possible through simultaneous engine- turbine design change would require a new set of boundary
turbocharger optimization. The proposed methodology is not only conditions. What’s required instead is a framework that models both
useful for improving existing designs; it can also develop a bespoke engine and turbine with acceptable accuracy and accounts for their
turbine geometry in new engine projects where there is no previously coupling. This paper presents an ‘active’ or ‘adaptive’ matching
available match. For these reasons, ‘adaptive’ turbo matching will methodology for the turbocharger turbine. By coupling an engine
become the standard approach in the automotive industry. cycle simulation to a meanline model of the turbine’s aerodynamic
performance (i.e., ‘in-the-loop’), the effect of turbine geometry
changes will be reflected in the exhaust boundary conditions, as well
Introduction as the overall engine performance prediction. Then, by running this
coupled engine-turbine model within an optimization framework, the
Turbocharging the gasoline passenger car engine is now optimal turbine design will evolve.
commonplace in industry as part of the technology mix alongside
gasoline direct injection (GDI) and downsizing, in the effort to lower A meanline model is a quasi-1D approach to predict the aerodynamic
fuel consumption and CO2 emissions from personal transportation. performance of, in this case, a radial turbine, using limited geometric
The process of selecting a turbocharger for a particular customer parameters. It assumes there is a mean streamline through the turbine
engine application, known as turbocharger matching, is a critical step stage along which a number of calculation stations may be identified
in being able to meet the desired performance characteristics. There and which are representative of average flow conditions. This permits
is typically a trade-off between key end user requirements such as straightforward thermodynamic calculations of turbine performance
fuel economy versus driveability (the latter being a subjective using a minimum of geometric data. The main limitation, however, is
combination of high power, fast transient response and especially the reliance on empirical loss correlations, which are necessary to
‘good low-end torque’) and this is strongly affected by turbocharger estimate the magnitude of various sources of pressure loss and flow
choice. The most suitable turbocharger configurations are identified blockage [1]. Much of the classical work involving meanline
Page 1 of 14

01/30/2018
modeling was performed at NASA [2,3], establishing the basic set of
loss models to predict radial turbine performance. Baines [4]
developed a meanline loss modeling system capable of predicting
both the on- and off-design performance of radial turbines, as a
refinement of the NASA approach. Qiu and Baines [1] extended the
meanline calculation method into the high pressure ratio region of the
turbine map and provided a method to obtain consistent predictions
under subsonic, transonic or supersonic flow conditions. Abidat [5]
used a meanline model to predict radial turbine performance in steady
and pulsating flow conditions. Romagnoli and Martinez-Botas [6]
developed a meanline model to predict performance of nozzled and
nozzleless mixed flow turbines and validated their predictions against
experimental data. Sakellaridis and Hountalas [7] developed a
meanline model for simulation of turbochargers to support diagnostic
investigations in diesel engines.

Following description of the methodology, the paper applies the Figure 1. Stations for turbine meanline model calculation.
proposed coupled engine-turbine optimization framework to a
Renault 1.2L turbocharged gasoline passenger car engine. Having
Swirl loss coefficient (𝑺)
been developed over a number of generations, the current series
production turbine supplied for this engine is already a very good
match. Nonetheless, the turbine aerodynamic design is optimized In an ideal volute, angular momentum of is conserved. In reality,
over a number of steady-state engine operating points under both full- some is lost due to wall friction between the flow and volute [9]. To
and part-load conditions while meeting certain constraints, starting account for this, the swirl loss coefficient 𝑆 is introduced to the
with the current production geometry as the baseline. The conservation of angular momentum in Equation 1, where 𝐶 is the
corresponding engine model has been previously validated against absolute velocity, 𝐶θ the tangential velocity, and 𝑟 the radius. Typical
measured engine dynamometer data, whilst the turbine meanline values range between 0.85–0.95 [10].
model is also shown to correctly predict performance of the baseline
production turbine when compared against flow bench data. 𝐶1 𝑟1 𝑆 = 𝐶θ2 𝑟2 (1)

Turbine Meanline Methodology Pressure loss (𝑲𝐏 )

The meanline model in this work is based on the quasi-one The pressure loss coefficient, 𝐾P (Equation 2), models the pressure
dimensional procedure initially used by the authors in Ref. [8], which (𝑃) losses due to volute wall friction [11]; typical values lie in the
determines the flow state at (in this case) three different stations range 0.1–0.3 [10].
through the nozzleless radial turbine stage: 1) volute, 2) rotor inlet,
and 3) rotor outlet, as shown in Figure 1. (N.B. a stage employing
𝑃01 −𝑃02
nozzles will require more stations [6].) 𝐾P = (2)
𝑃02 −𝑃2

The meanline model accounts for energy dissipation along the flow
path through the turbine stage by deploying a set of empirical Rotor loss modeling
correlations (or loss models) that describe loss generation. A common
feature of most loss modeling systems is a conceptual division of the Energy losses in the rotor are modeled according to the NASA
overall loss into separate components, corresponding to different approach [2,3], which attributes losses due to incidence effects,
physical loss mechanisms [4]. The first classification distinguishes friction in the flow passage, clearance between rotor and shroud, and
volute and rotor losses – which are described next. disk friction on the wheel.

Volute loss modelling Incidence loss (Li)

As per Ref. [6], two major volute loss mechanisms are considered, Ideal flow conditions at rotor inlet do not actually correspond to
embodied in the swirl and pressure loss coefficients. These account perfect alignment between the flow and the blade. This phenomenon
for irreversibilities due to mixing, secondary flow and recirculation. has been demonstrated experimentally by Yeo and Baines [12] and is
due to the pressure difference between the blades’ pressure and
suction surfaces. In a radial turbine, this results in an optimum
relative inlet flow angle, 𝛽2,opt , of somewhere between -20 to -30
[10]. The incidence angle, 𝑖2 , is defined in the meanline model as the
difference between the actual and optimum relative inlet flow angles
(Equation 3), i.e., the ideal situation is when 𝑖2 = 0. (N.B. other
works refer the incidence angle to the blade angle; both definitions
are in general use and so care must be taken to be consistent.)

Page 2 of 14

01/30/2018
 𝑖2 = 𝛽2 − 𝛽2,opt (3) 0.02125 𝑈23 𝜌22
𝐿df = 𝜌2 𝑈2 𝑟2 0.2
(9)
𝑚̇( )
𝜇
As the relative inlet flow angle departs from the optimum (i.e., 𝑖2
becomes non-zero), flow separation becomes more likely and mixing The meanline model has been programmed in FORTRAN, and
losses in the rotor increase [11]. The enthalpy loss due to incidence, requires the turbine rotational speed, total inlet conditions, static
𝐿i (Equation 4), is modeled assuming that the change in relative outlet pressure, the thermodynamic properties of the working
tangential kinetic energy manifests as an increase in internal energy medium, and the basic turbine geometric parameters as inputs. The
of the gas (and a consequent increase in entropy). Here 𝐾i is the model returns the mass flow rate and total-to-static efficiency as
incidence loss coefficient and 𝑊2 is the relative inlet velocity. outputs.

𝐿i = 0.5 𝐾i 𝑊22 sin2 𝑖2 (4)

Turbine Meanline Model Validation
Clearance loss (𝑳𝐜 )
Experimental validation
There must exist a clearance between the blade tip and the shroud,
the latter provided by the inside of the turbine housing (in a typical As a validation exercise, the meanline model was used to predict the
turbocharger). The pressure difference between the pressure and performance of an off-the-shelf, ~36mm diameter radial turbine for a
suction blade surfaces drives a tip clearance flow through this gap, mass-produced passenger car engine turbocharger, manufactured by
resulting in an enthalpy loss. This clearance loss, 𝐿c , is based on that Mitsubishi Turbocharger and Engine Europe BV (MTEE). First of
in Ref. [4], shown here in Equation 5 all, the model is calibrated against a single speed line from the
measured turbine map, using a genetic algorithm (GA) to obtain the
𝑈23 𝑁
coefficients for the different loss mechanisms, with the objective of
𝐿c = (𝐾a 𝑒a 𝐶a + 𝐾r 𝑒r 𝐶r + 𝐾ar √𝑒a 𝐶a 𝑒r 𝐶r ) (5) minimizing the sum of squares between the model prediction and
8𝜋
experimental data. Once calibration was attained for this single speed
where line, the model was exercised to predict turbine performance for the
remaining four speed lines in the measured map. Figure 2 compares
the measured and predicted swallowing capacity, while Figure 3
 𝑈2 is the inlet blade speed;
compares the efficiency, for the five speed lines. Here, rotational
 𝑁 is the blade number; speed has been normalized against the maximum measured speed; the
 𝐾a and 𝐾r are resp. axial and radial clearance loss coefficients; speed line used for calibration is labelled ‘100% Speed’.
 𝑒a and 𝑒r are axial and radial tip clearances, resp.;
 𝐾ar is the cross-coupling coefficient; and

where the axial and radial absolute velocities, resp. 𝐶aand 𝐶r , are
𝑟3,tip
1−( )
𝑟2
𝐶a = (6)
𝐶r,2 𝑏2

and

𝑟3,tip 𝑧−𝑏2
𝐶r = ( ) (7)
𝑟2 𝐶a,3 𝑟3 𝑏3

Passage loss (Lp)

Passage loss accounts for pipe friction and blade loading losses in the Figure 2. Comparison of measured and predicted turbine swallowing capacity.
blade passage. The meanline model uses the treatment from Ref.
[10], described here in Equation 8, where 𝐾p is the passage loss
coefficient and 𝑊3 is the relative outlet velocity.

𝐿p = 0.5 𝐾p (𝑊22 cos2 𝑖2 + 𝑊32 ) (8)

Disk friction loss (Ldf)

The disk friction or windage loss, 𝐿df , accounts for friction on the
backface of the turbine wheel. The meanline model employs the
expression in Ref. [3], shown here in Equation 9, where 𝜌2 is the
rotor inlet gas density, 𝑚̇ its mass flow rate, and 𝜇 its dynamic
viscosity.

Figure 3. Comparison of measured and predicted turbine efficiency.

Page 3 of 14

01/30/2018
It can be seen that the calibrated model predicts the turbine behavior
very well both in terms of mass flow and efficiency within the
measured data range. However, since the model predicts turbine
performance over a considerably wider range of pressure ratio
compared to that available from gas stand measurements, it does
mean the meanline model cannot be validated against experimental
data in these regions – this issue will be addressed in the next section.

When employing the turbocharger compressor as the loading device

during map measurement, the measured turbine speed lines will be
quite limited in width due to compressor surge and choke. This width
does not sufficiently cover the range of turbine operation on-engine,
Figure 5. Computational mesh of turbine volute and rotor.
and so when used in 1D engine cycle simulations, such measured
maps must be extrapolated, using a mathematical or physics-based A total pressure of 150 kPa and total temperature of 373.15K were
technique, or a combination thereof. Nevertheless, that approach will applied as the inlet boundary conditions at the plane identified in
likely have a weaker physical basis than using the meanline model Figure 4, together with a medium turbulent intensity of 5%, for all
presented here, resulting in poorer turbine performance prediction simulations. Total-to-static efficiency and mass flow rate were
when the engine operates away from the measured map. monitored to track convergence, with a convergence criterion of
1x10-5 used for all residuals. Static pressure was specified as the
CFD validation outlet boundary condition, and was varied to obtain different
operating conditions. The rotor shroud was defined as a counter
To ascertain the predictive accuracy of the meanline model outside of rotating wall to simulate the relative motion between the turbine rotor
the map measured map range, a 3D CFD analysis of the same subject and volute. As shown in Figure 4, frozen rotor interfaces were
turbine was performed. The meanline model predictions were specified between the stationary volute and rotor, and between the
compared against corresponding predictions obtained from CFD. rotor and outlet duct. All walls were defined as adiabatic.

CFD setup and pre-processing Comparison of CFD and meanline predictions

The 3D geometry of the turbine volute and rotor was provided by Figure 6 compares the predictive capability of the meanline model
MTEE for the CFD analysis. A preliminary step was to obtain a clean against the CFD simulations, for three speed lines.
geometry prior to meshing, i.e., removal of feature details that are
irrelevant for CFD analysis and which would potentially impose
unnecessarily high local mesh resolution. Altair HyperMesh was then
used to mesh the entire fluid domain: turbine volute, rotor, plus inlet
and outlet extrusions, as shown in Figure 4.

Figure 4. Turbine stage CFD domain and boundary conditions.

An unstructured mesh (Figure 5) was created for all components,

with 10 elements in the near wall boundary layer mesh. An average
y+ value of 5 was achieved. The grid consisted of 2.1 million
elements in the volute and 3.2 million elements in the rotor. 3D
RANS steady-state simulations were carried out using the ANSYS
CFX 18.1 finite volume solver. The two-equation 𝑘-𝜔 SST
turbulence model was selected; this combines the robust formulation
of the 𝑘-𝜔 model in the near wall region with the far stream
independence of the 𝑘-𝜀 model [13]. It provides more accurate
predictions when there is flow separation under adverse pressure
gradients and in cases of wall bounded flows [14].
Figure 6. Comparison of 3D CFD and meanline model predictions.
Page 4 of 14

01/30/2018
The meanline model swallowing capacity and efficiency predictions
agree well with CFD (within ± 3%-age points), with trends captured
for all three speed lines (not only at the 100% speed used for
calibration), even at very low and high pressure ratios (where test
data isn’t available). Combined with the earlier experimental
validation, this gives confidence that the meanline model will provide
accurate turbine performance prediction across the full operating
range experienced in 1D engine simulations.

Engine Model Validation

A Renault 1.2L turbocharged GDI engine (Table 1) was used as the
subject engine in this study. The engine air system employs a fixed
geometry, wastegated turbocharger supplied by MTEE. The
wastegate is used to control the delivered boost pressure by
increasing the effective exhaust flow area, bypassing some exhaust
gas around the turbine, restricting the developed turbine power.
Table 1. Renault 1.2L turbocharged GDI engine specifications.

Capacity (cc) 1198

Combustion
4-stroke GDI
system
No. of cylinders 4

Bore x stroke (mm) 72.2 x 73.1 Single, fixed

Air system geometry, wastegated
Compression ratio 10:1 turbocharger

A GT-POWER 1D engine cycle simulation model was supplied by

the engine manufacturer for this study. Measured compressor and
turbine performance maps, provided by MTEE, were used for
modeling the turbocharger at this stage. The engine model targets the
desired torque at each speed using the in-built wastegate controller.
Validation was carried out for 14 full-load steady-state operating Figure 7. Comparison of measured and simulated engine performance.
points (Table 2) against engine dynamometer test data.
Table 2. Full-load engine operating points.

Engine speed Normalized target Engine speed Normalized target

(rpm) engine torque (rpm) engine torque

1000 0.65 3000 1.00

1250 0.83 3500 1.00

1500 1.00 4000 1.00

1750 1.00 4500 0.95

2000 1.00 5000 0.85

2250 1.00 5500 0.78

2500 1.00 6000 0.71

All results in the paper have been normalized by the maximum value
of the corresponding parameter. Figure 7 presents the comparison of
simulated engine performance against the engine test data. It can be
seen in Figure 7 (a) that the engine model predicts brake power well Figure 8. Comparison of measured and simulated turbocharger performance.
across the speed range; however, when translated into brake torque
any small differences, particularly at low engine speeds, are Figure 8 compares the predicted cycle-averaged turbocharger
amplified. Indeed, Figure 7 (b) shows that the model over predicts the performance against test data. In Figure 8 (a), the engine model
torque at the third and fourth engine speeds. Similarly, Figure 7 (c) predicts turbocharger speed reasonably well over the entire measured
shows the trend of brake specific fuel consumption (BSFC), which is range, again except for engine speeds between 1500–2000 rpm.
the preferred indicator of overall system efficiency in this paper, is However, in Figure 8 (b), the compressor pressure ratio (thus boost)
well-captured except at the lowest engine speeds.
Page 5 of 14

01/30/2018
is well-predicted only for the mid-speed range (2000–4500 rpm). At
high engine speeds (5000–6000 rpm), the model predicts too high a
boost, in order to meet the required engine torque. Despite the
deviations between simulated and measured engine performance
(mainly at low and high engine speeds), cycle-averaged predictions
can be used to evaluate relative engine performance.

Turbine Modeling: Map versus Meanline

As mentioned, 1D engine cycle simulation tools conventionally use
experimentally-measured turbine and compressor maps in order to
simulate turbocharger performance. Due to measured width
limitations, turbine maps especially must usually be extrapolated
before they can be used in such simulations, thereby introducing a
certain amount of prediction inaccuracy [7], especially when
operating far outside the measured range [15]. The long-term
objective of this work is to reduce this modeling uncertainty and
reliance on the measured turbocharger turbine maps for engine
simulations. Thus the radial turbine meanline model was integrated
with the engine model to predict turbine behavior, supplanting the
turbine map component. (Compressor performance prediction suffers
less from extrapolation, and continued to be modeled using maps.) In
an engine simulation, the instantaneous turbocharger rotational speed
is known, while turbine expansion ratio and thermodynamic fluid
properties are defined by the instantaneous pressure and temperature
conditions in adjoining ducts. At every time step, these are provided
as boundary conditions to the meanline model, which in turn supplies
its prediction of instantaneous turbine mass flow and efficiency.

To assess how the choice of map-based or meanline-based turbine

modeling affects performance prediction, 1D engine simulations were
performed for both approaches, for the 14 full-load engine operating
points in Table 2, and 8 additional part-load points listed in Table 3.

Table 3. Part-load engine operating points.

Engine speed Normalized target Engine speed Normalized target

(rpm) engine torque (rpm) engine torque

1500 0.85 2500 0.80

1500 0.75 2500 0.85

2000 0.85 3000 0.75

2000 0.75 3500 0.60

Figure 9 compares the engine model full-load predictions using the

two turbine simulation methods: map-based and meanline model. It
can be seen in Figure 9 (a) and Figure 9 (b) that the meanline model- Figure 9. Map vs meanline turbine models: full-load engine performance.
based simulation predicts the same engine torque and power as the
map-based simulation. The BSFC predictions in Figure 9 (c) match
well between methods at low and mid-engine speeds, however at the
highest engine speed the meanline model-based simulation shows a
1.2% lower BSFC than the map-based prediction. This can be
explained by looking at the PMEP predictions for the two cases in
Figure 9 (d), where the turbine meanline model results in improved
(less negative) pumping work and hence lower BSFC. Turbocharger
speed and compressor operation were identical between methods.

Page 6 of 14

01/30/2018
Figure 10. Map vs meanline turbine models: turbine operation at full load. Figure 11. Map vs meanline turbine models: part-load engine performance.

Figure 10 continues the comparison of map- and meanline-based Next, predictions at the part-load engine operating points in Table 3
turbine modeling options, at the turbine level. Both estimate very are compared. Figure 11 shows both methods again predict similar
similar turbine mass flow rate and blade speed ratio (BSR), in Figure engine performance, in terms of brake power, torque and BSFC. It
10 (a) and Figure 10 (d) resp., at all engine speeds. Pressure ratio and can be seen from Figure 12 (a) and Figure 12 (c) that the meanline
efficiency are also similarly predicted except at the highest engine model-based simulation predicts slightly higher mass flow through
speeds where there is a ~2.5% difference. This is likely due to the the turbine and lower efficiency at the same engine speed, compared
extrapolation required in this operating region. to the map-based approach. Since these parameters have an opposite
effect on pumping work, they balance each other out, and so the
predicted BSFC remains similar (Figure 11 (c)). The deviation in
mass flow and efficiency predictions may again be attributed to
extrapolation effects in the map-based approach.

Page 7 of 14

01/30/2018
 In this work, the GA optimizer available in the 1D engine cycle
simulation software GT-POWER was used. Figure 13 shows the
basic turbine geometric parameters. Table 4 lists the parameters to be
varied by the optimizer – a maximum perturbation of ±8% to the
nominal values was allowed (to limit the change in turbine wheel
inertia). A population size of 30 was specified, which is more than
twice the number of design variables, as recommended by Ref. [18].
A crossover rate of 1, a crossover rate distribution index of 15, and a
mutation rate of 0.14 were specified to create diversity in the
population. The scope of optimization was purposely limited to the
turbine, i.e., no compressor or engine model parameters were
changed. The objective was to minimize overall system BSFC over a
set of (1) full- and (2) part-load steady-state engine operating points.

Figure 13. Basic turbine geometric parameters.

Table 4. Turbine geometric parameters varied by the optimizer.

Parameter Description

𝐴1 Volute inlet area

𝑟1 Volute inlet radius

𝐵2 Rotor inducer height

𝑟2 Rotor inlet radius

𝑧 Blade length

𝑟3,tip Rotor outlet tip radius

𝛽3b Rotor outlet blade angle

Case 1: Full-load turbine design optimization

This case considered an imagined steady-state drive cycle,

Figure 12. Map vs meanline turbine models: turbine operation at part load. comprising a sub-set of 6 full-load engine operating points (Table 5),
over which turbine aerodynamic optimization was performed. Here,
Turbine Optimization to Reduce System BSFC these were given equal importance, but it is straightforward to assign
weightings to distinguish different drive cycles.
The coupled engine-turbine meanline model permits 1D turbocharged Table 5. Full-load engine operating points for turbine optimization (Case 1).
engine simulations without recourse to measured turbine maps,
instead relying on a description of the turbine geometry (plus, of Engine speed Normalized target Engine speed Normalized target
(rpm) engine torque (rpm) engine torque
course, calibrated loss correlations). And since the geometry may be
adjusted during simulation, it becomes a suitable platform for 1250 0.83 3500 1.00
performing aerodynamic optimization. A genetic algorithm [16] was
selected for this purpose, a type of evolutionary algorithm [17] 1500 1.00 4500 0.95
inspired by biological processes (e.g., mutation, crossover, natural
selection), variously applied in turbomachinery design to iteratively 2250 1.00 6000 0.71
improve a set of solution candidates [e.g., 18, 19].
Page 8 of 14

01/30/2018
 Figure 15, where the shaded bars represent different engine speeds.
Sensitivity is computed by a linear least squares approach using all
iteration data. The slopes determined by least squares fitting are
normalized by the sum of all slopes [20].

Straightaway it may be inferred that, in this case, blade length 𝑧 has

little influence on BSFC, while the rotor outlet radius 𝑟3,tip shows the
greatest influence, irrespective of engine speed. This can be related to
the variation of turbine throat area with rotor outlet radius, which
dictates swallowing capacity. The rotor inlet radius 𝑟2 also has a
significant impact on engine BSFC. The rotor outlet blade angle 𝛽3b
is the next most important design parameter. The volute parameters
𝐴1 and 𝑟1 also influence BSFC as they determine the turbine ‘A/R’, a
parameter used in the industry to denote relative turbine housing size.
Figure 14. Optimization convergence at full load (Case 1). For example, a smaller A/R will tend to raise exhaust back pressure,
consequently increasing pumping work and overall BSFC. Finally,
Figure 14 shows the convergence plot for the optimization process, while there is some variance in sensitivity to the same parameter
which is halted once the change in objective function (i.e., BSFC) is between engine speeds, this is much smaller than sensitivity
less than < 0.05 g/kWh. In Case 1, after ~300 iterations (~6 days on a differences between parameters. Although brief, this analysis helps
workstation using two cores of an i7-2600 processor, clock speed 3.4 identify influential turbine design parameters, translating adjustments
GHz), the optimizer achieved an improvement in the cycle-averaged at the turbine component level to overall engine system level BSFC.
BSFC of ~3.5 g/kWh, a significant reduction.

Investigating further, Table 6 presents the relative change in the

optimized design parameters compared to the baseline design. It can
be seen that Case 1 optimization resulted in an increase in the volute
inlet area 𝐴1 , the rotor inducer height 𝐵2 , the rotor inlet radius 𝑟2 , and
the rotor outlet tip radius 𝑟3,tip , but a decrease in the volute inlet
radius 𝑟1 , the rotor outlet blade angle 𝛽3b , and the blade length 𝑧.
Table 6. Relative change in design parameters optimized at full load (Case 1).
Parameter Description Change relative to baseline (%)

𝐴1 Volute inlet area 7.12

𝑟1 Volute inlet radius -6.18

𝐵2 Rotor inducer height 4.04

𝑟2 Rotor inlet radius 7.38

𝑧 Blade length -2.44

𝑟3,tip Rotor outlet tip radius 6.73

𝛽3b Rotor outlet blade angle -6.71

Figure 15. Influence of turbine geometric parameters on engine BSFC at

different full-load engine speeds (Case 1).

Next, optimization data was used to explore the influence of

individual turbine geometric parameters on engine BSFC, shown in Figure 16. Comparison of full-load engine performance for baseline and
optimized turbine designs (Case 1).
Page 9 of 14

01/30/2018
 Figure 16 compares predicted engine performance for the baseline
and optimized turbine designs. Figure 16 (a) and Figure 16 (b) simply
confirm that the optimized design meets the torque and power targets.
Figure 16 (c) shows the magnitude of BSFC reduction is larger at
higher engine speeds, but there is an improvement at all speeds. This
stems from improved (less negative) pumping work, as seen in Figure
16 (d), which can itself be explained by comparing the turbine
operation in Figure 17.

Figure 17 (a) and Figure 17 (c) indicate that design optimization

resulted in a turbine with improved flow capacity and efficiency at all
engine operating points, while imposing a lower expansion ratio
(Figure 17 (b)). This reduces exhaust back pressure and hence
pumping work, reflected in improved BSFC. At the turbine stage
level, improved efficiency may be attributed to the optimized design
operating closer to the optimum BSR of ~0.7 for an ideal radial
turbine [10], at least at the three highest engine speeds. Indeed,
Figure 17 (d) shows that the optimized design operates at a slightly
higher BSR at all engine speeds. Compressor operation for both
designs remained the same; this is to be expected since the same
compressor map is used, and, ignoring any slight differences in
cylinder scavenging, the same nominal boost level will be required to
meet the target torque. It follows then that, as shown in Figure 17 (e),
turbocharger rotational speed remains almost identical for the two
designs. Figure 17 (f) shows wastegate mass flow rate reduces
compared to the baseline design, in order for the turbine to meet the
power demanded by the compressor for the required boost pressure.

So, despite the current series production turbine already being a good
match, and with optimization restricted to a handful of parameters,
simulation results suggest there remains a worthwhile margin of 3.5
g/kWh for system efficiency gains at full load.

Case 2: Part-load turbine design optimization

Table 7 lists the 5 part-load engine operating points for which the
second optimization case was performed. Again, each point was
assigned equal importance in the absence of detailed drive cycle data.
Table 7. Part-load engine operating points for turbine optimization (Case 2).

Engine speed Normalized target Engine speed Normalized target

(rpm) engine torque (rpm) engine torque

2000 0.75 3000 0.75

2250 0.75 3500 0.60

2500 0.80

Figure 17. Comparison of turbine performance at full-load for baseline and

optimized turbine designs (Case 1).
Figure 18. Optimization convergence at part load (Case 2).
Page 10 of 14

01/30/2018
The same GA configuration and parameter list as described for full- some room to improve part-load engine efficiency through turbine
load optimization was again used, with convergence achieved after optimization, albeit to a lesser extent that at full load.
~350 iterations (Figure 18). Table 8 shows the relative change in the
design parameters as a result of optimization at part load, resulting in
an increase in the volute area 𝐴1 , the rotor inlet radius 𝑟2 , and the
rotor outlet tip radius 𝑟3,tip . A decrease is seen in the volute inlet
radius 𝑟1 , the rotor inducer height 𝐵2 , the rotor outlet blade angle 𝛽3b ,
and the blade length 𝑧. The direction of these changes is the same as
for the full-load case (except for rotor inducer height 𝐵2 , which
increased by ~4% under optimization at full load). Nonetheless the
small changes in magnitude go to show that the optimal turbine
design does differ slightly between full- and part-load operation.
Table 8. Relative change in design parameters optimized at part load (Case 2)
compared to those at full load (Case 1).
Change relative to baseline (%)
Parameter
Full load (Case 1) for comparison Part load (Case 2)

𝐴1 7.12 7.66

𝑟1 -6.18 -5.30

𝐵2 4.04 -0.17

𝑟2 7.38 7.08

𝑧 -2.44 -3.77

𝑟3,tip 6.73 7.32

𝛽3b -6.71 -5.82

Figure 19 presents the sensitivity of BSFC to each design parameter,

for the part-load optimization. Compared with Case 1 (full load), now
both the rotor outlet blade angle 𝛽3b and the rotor outlet tip radius
𝑟3,tip show considerable influence on engine BSFC, at all engine
speeds. Part-load BSFC is also sensitive to volute inlet area and
radius. Rotor inducer height has a small influence on BSFC, as for
the full-load case. Lastly, the blade length 𝑧 continues to play a very
minor role in determining engine BSFC.

Figure 20. Comparison of part-load engine performance for baseline and

optimized turbine designs (Case 2).

Figure 21 compares optimized and baseline turbine performance at

part load. As for full load, both mass flow and efficiency increase
(Figure 21 (a) & (c) resp.), while expansion ratio (Figure 21 (b))
decreases for the optimized design. Optimization again results in a
design that operates at higher BSR (Figure 21 (d)), turbocharger
rotational speed (Figure 21 (e)) remains constant, and wastegate flow
Figure 19. Influence of turbine geometric parameters on engine BSFC at (Figure 21 (f)) reduces to match the required compressor power.
different part-load engine speeds (Case 2).

Figure 20 compares predicted engine performance for the baseline Conclusions

and part-load optimized turbines. Figure 20 (a) and Figure 20 (b)
again confirm that the optimized design meets the desired part-load This paper presents an ‘adaptive’ turbine matching methodology that
torque and power. In Figure 20 (c), the optimized design improves couples a turbine meanline model to a 1D engine model. The turbine
the BSFC at all points, resulting in a slightly better cycle-averaged meanline model is based on existing loss correlations in the literature,
fuel consumption of 0.9 g/kWh overall. Again there appears to be and is first calibrated against a single speed line from a measured
turbine map. Using the baseline series production turbine, validation

Page 11 of 14

01/30/2018
of the meanline model against both experimental data and CFD The meanline model was coupled to an engine model for a Renault
provides confidence that the meanline model accurately predicts 1.2L turbocharged GDI passenger car engine, which was generally a
turbine performance across the full on-engine operation range. good match to the engine dynamometer test data. Validation of the
coupled engine-turbine model results in the same engine performance
given the same baseline turbine design. A comparison of engine
performance predictions using both map-based and meanline turbine
models highlighted some differences, particularly in regions expected
to suffer from turbine map extrapolation.

The main objective of the current paper is to demonstrate the use of a

meanline model to optimize turbine design, in a scenario where the
objective is to minimize fuel consumption over a given set of engine
operating points. Two cases were considered: full- and part load, with
all points given equal importance in each. In the full-load case, the
GA-optimized turbine geometry predicted a significant reduction in
cycle-averaged BSFC of 3.5 g/kWh. However, it should be noted that
real passenger car drive cycles spend the clear majority of the time
under part load conditions. In this case, the part-load optimized
design showed just a 0.9 g/kWh cycle-averaged BSFC improvement
– this is nonetheless worth having. Though it must also be noted that
this was achieved with optimization restricted to a handful of turbine
geometric parameters, and without any modification to the engine
design or breathing and combustion control parameters (e.g., valve
timing and spark advance). This suggests greater efficiency gains
may be possible if engine and turbocharger optimization is performed
simultaneously. Overall then, there appears to be attainable system-
level efficiency benefits through turbine aerodynamic optimization.

The paper also briefly examined the influence of turbine design

parameters on engine BSFC, using the data generated by the
optimization process. This highlighted that, for the turbine design
parameters under investigation:
 The most influential were the rotor outlet tip radius 𝑟3,tip at full
load, and the rotor outlet blade angle 𝛽3b (closely followed by
𝑟3,tip ) at part load;

 The least influential was the blade length 𝑧 in either case; and
 There is a noticeable but less critical variation in the influence of
each parameter across different engine speeds.

In sum, the benefits of ‘adaptive’ turbine matching by employing a

coupled engine-turbine meanline model are that:
 It removes reliance on measured turbine maps (and the associated
poor predictive accuracy incurred by map extrapolation);
 It enables aerodynamic optimization of an existing turbine
geometry, or development of a bespoke turbine geometry in new
engine projects where there is no previously available match, for
given engine-level customer objective(s), e.g., BSFC; and
 It permits sensitivity of engine-level performance (including
BSFC) to component-level design parameters (the turbine, in this
case) to be studied.

The cost of these benefits is the not insignificant computational time

required for an optimized turbine design to evolve. However, this
must be weighed against the total time taken by the standard
matching approach in which any number of turbine maps may need
to be evaluated before the most suitable (though non-optimal) turbine
design reveals itself. Based on their experience, the authors suggest
that an ‘adaptive’ turbine matching process will in reality take no
longer, yet will likely result in better designs. For these reasons, it is
Figure 21. Comparison of turbine performance at part load for baseline and
optimized turbine designs (Case 2).
Page 12 of 14

01/30/2018
expected that some form of adaptive turbocharger matching will 6. Romagnoli, A. and Martinez-Botas, R., “Performance
eventually become the standard approach. prediction of a nozzled and nozzleless mixed-flow turbine in
steady conditions,” Int. J. Mech. Sci. 53(8):557–574, 2011,
Future work doi:10.1016/j.ijmecsci.2011.05.003.

7. Sakellaridis, N. and Hountalas, D., “Meanline Modeling of

The following items remain to be addressed:
Radial Turbine Performance for Turbocharger Simulation
 The current geometry optimization process is constrained to and Diagnostic Applications,” SAE Technical Paper 2013-
generate purely radial designs. This will be relaxed in future work 01-0924, 2013, doi:10.4271/2013-01-0924.
to allow mixed flow designs to be evolved, in order to explore
further efficiency improvements that are potentially on offer. 8. Kapoor, P., Javed, A., and Costall, A., “Optimization of a
Radial Turbine Design for Improved Engine-Turbocharger
 The current work used an imagined set of full- and part-load Matching,” In Proc. THIESEL Conference on Thermo-and
engine operating points, and considered them separately, Fluid Dynamic Processes in Direct Injection Engines, pp.1–
resulting in slightly different designs. This goes to show that the 12, 2014.
optimal turbine design differs between full- and part-load
operation, and in general between any non-identical set of engine
9. Romagnoli, A., “Aerodynamic and thermal characterization
operating points. Hence future work will move towards turbine
of turbocharger turbines: experimental and computational
optimization over more realistic drive cycles, which will
evaluation,” PhD Thesis, Imperial College London, 2010.
inherently comprise a mixture of full- and part- load operation.
 In the current work, optimization relies on the accuracy of the 10. Japikse, D. and Baines, N.C., “Introduction to
meanline model. While it has been validated against both test Turbomachinery,” Concepts ETI, 1997.
data and CFD for an existing design, future work must consider
the corresponding optimized 3D geometry and its simulation in 11. Whitfield, A. and Baines, N.C., “Design of Radial
3D CFD, alongside experimental testing, to fully validate the Turbomachines,” Longman Scientific & Technical, UK,
meanline optimization process. 1990.
 So far optimization has been aerodynamic-only; no inertia or
structural constraints are currently imposed. Future work must 12. Yeo, J. H. and Baines, N. C., “Pulsating Flow Behaviour in a
consider implications of design changes on turbine inertia, since Twin-Entry Vaneless Radial-Flow Turbine,” In Proc.
this affects engine transient response (a critical customer IMechE 4th Int. Conf. on Turbocharging & Turbochargers,
requirement), and blade shape optimization must accommodate Paper C405/004, 1990.
mechanical stress constraints, if the methodology is to be
commercially useful. 13. Menter, F.R., “Two-equation eddy-viscosity turbulence
models for engineering applications,” AIAA J. 32(8):1598–
 As mentioned, no engine design parameters have so far been 1605, 1994, doi:10.2514/3.12149.
adjusted, but this could reveal further efficiency gains. This will
require closer collaboration with the engine manufacturer to 14. Javed, A. and Kamphues, E., “Evaluation of the Influence of
perform simultaneous engine and turbocharger optimization. Volute Roughness on Turbocharger Compressor
Performance From a Manufacturing Perspective,” In Proc.
References ASME Turbo Expo, Paper No. GT2014-26949, 2014,
doi:10.1115/GT2014-26949.
1. Qiu, X. and Baines, N., “Performance Prediction for High
Pressure-Ratio Radial Inflow Turbines,” In Proc. ASME 15. Mason, A., Costall, A., and McDonald, J., "The Sensitivity
Turbo Expo, Paper No. GT2007-27057, 2007, of Transient Response Prediction of a Turbocharged Diesel
doi:10.1115/GT2007-27057. Engine to Turbine Map Extrapolation," SAE Technical
Paper 2017-24-0019, 2017, doi:10.4271/2017-24-0019.
2. Wasserbauer, C.A. and Glassman, A.J., “FORTRAN
Program for Predicting Off-Design Performance of Radial- 16. Goldberg, D.E., “Genetic Algorithms in Search,
Inflow Turbines,” NASA Tech. Note D-8063, 1975. Optimization, and Machine Learning,” 13th Ed., Addison
Wesley Publishing Co., USA, 1989.
3. Meitner, P.L. and Glassman, A.J., “Computer code for off-
design performance analysis of radial-inflow turbines with 17. Petrowski, A. and Hamida Ben, S., “Evolutionary
rotor blade sweep,” AVRADCOM Tech. Rep. 83-C-4, 1983. Algorithms,” 1st Ed., John Wiley & Sons, Inc., USA, 2017.

4. Baines, N.C., “A meanline prediction method for radial 18. Javed, A., “Developing Generic Design Expertise for Gas
turbine efficiency,” In Proc. IMechE 6th Int. Conf. on Turbine Engines: Robust Design of a Micro Centrifugal
Turbocharging & Air Management Systems, Paper No. Compressor,” PhD Thesis, TU Delft, 2014.
C554/006, 1998.
19. Khairuddin, U. and Costall, A.W., “Aerodynamic
5. Abidat, M., “Design and testing of a highly loaded mixed Optimization of the High Pressure Turbine and Interstage
flow turbine,” PhD Thesis, University of London, 1991. Duct in a Two-Stage Air System for a Heavy-Duty Diesel
Engine,” ASME J. Eng. Gas Turbines Power
140(5):052801–052801-10, 2017, doi:10.1115/1.4038024.
Page 13 of 14

01/30/2018
20. Gamma Technologies Inc., “GT-SUITE Optimization 3 rotor outlet
Manual,” 2016.
a axial

Contact Information b blade

c clearance
Dr Aaron Costall
Lecturer in Thermofluids df disk friction
Imperial College London
Department of Mechanical Engineering i incidence
South Kensington Campus
London SW7 2AZ max maximum
United Kingdom
min minimum
a.costall@imperial.ac.uk
opt optimal
Acknowledgments
p passage
The authors would like to express their thanks to Mitsubishi r radial
Turbocharger and Engine Europe BV for funding this project and
providing turbine geometrical and performance data, and Groupe tip blade tip
Renault for providing the engine model and test data.
𝜃 tangential

Definitions/Abbreviations Abbreviations

Roman symbols BMEP Brake Mean Effective Pressure

B blade height BSFC Brake Specific Fuel Consumption

C absolute velocity BSR Blade:Speed Ratio

e tip clearance CFD Computational Fluid Dynamics

i incidence GA Genetic Algorithm

K calibration coefficient GDI Gasoline Direct Injection

L enthalpy loss MTEE Mitsubishi Turbocharger and Engine Europe BV

RANS Reynolds-averaged Navier-Stokes

N blade number
PMEP Pumping Mean Effective Pressure
P pressure

r radius

S swirl loss coefficient

U blade speed

W relative velocity

z blade length

Greek symbols

𝛽 relative flow angle

𝜇 dynamic viscosity

𝜌 density

Subscripts

0 total condition

1 volute inlet

2 rotor inlet

Page 14 of 14

01/30/2018