Computers and Fluids 284 (2024) 106438

Contents lists available at ScienceDirect

Computers and Fluids

journal homepage: www.elsevier.com/locate/compfluid

Application of the iteratively regularized Gauss–Newton method to

parameter identification problems in Computational Fluid Dynamics
Stefan Langer
German Aerospace Center - Institute for Aerodynamics and Flow Technology, Center for Computer Applications in Aerospace Science and
Engineering, Lilienthalplatz 7, Braunschweig, 38108, Germany

ARTICLE INFO ABSTRACT

Keywords: Field Inversion and Machine Learning is an active field of research in Computational Fluid Dynamics (CFD).
RANS equations This approach can be leveraged to obtain a closed-form correction for a given turbulence model to improve
Iteratively regularized Gauss–Newton method the predictions. The fundamental approach is to insert a parameter into the system of RANS equations and
Field Inversion and parameter identification
determine it in a way such that, for example, a given pressure distribution is better approximated compared to
the one obtained with the original set of equations. The goal of this article is twofold. Numerical arguments are
presented that these kinds of problems can be severely ill-posed. In the second part, an approach is presented
to directly reconstruct the turbulent viscosity field along with an example. The Iteratively Regularized Gauss-
Newton Method (IRGNM) is used for a realization. The construction of a problem-adapted norm for a finite
volume method is presented. Finally, an outlook is presented on how this approach can be used to possibly
modify or improve turbulence models such that not only one, but a larger number of test cases are considered.

1. Introduction (b) this steady-state solution represents important features compared

to a solution of the original Navier–Stokes equations,
In the Reynolds averaged Navier-Stokes (RANS) framework it is (c) it is possible to compute numerically an approximation to this
the goal of turbulence modeling to determine the Reynolds stress ten- steady-state solution with acceptable effort based on the idea that
sor 𝜌𝑢′′ ′′
𝑖 𝑢𝑗 . This can be done by introducing closures entailing different
the use of a turbulence model allows a significant reduction in the
levels of complexity. For standard CFD applications, the Boussinesq degrees of freedom compared to a Large Eddy Simulation (LES)
approximation is generally invoked, which relates the Reynolds stresses or a Direct Numerical Simulation (DNS).
directly to the trace-free rate of strain 𝑆𝑖𝑗 . As a consequence the In many situations the turbulence models serve their purpose, and there
turbulent viscosity, typically denoted by 𝜇𝑡 in the literature, is in- are many examples where, for example, drag and lift coefficients as
corporated into the Navier–Stokes equations. Prominent examples of well as other aerodynamic quantities agree well with measurements.
such an approach are the Spalart–Allmaras turbulence model [1,2], On the other hand, there are various examples in which solutions to
the Wilcox 𝑘-𝜔 model [3,4], the Menter SST 𝑘-𝜔 model [5–7], and 𝑘- the RANS equations are not unique or for which no steady-state solution
𝜀 models (see [8,9]). The definition of these models is typically not can be computed, at least on a discrete level (see e.g. [11]). All of these
unique, but there exist different variants. A rather complete overview are indications that boundary-value problems for the RANS equations
of the number of possible models as well as their relationships can be are generally not well-posed in the sense of Hadamard [12]. For this
found in the report by Bredberg [10]. reason, it is still of great interest to understand how the turbulence
When applying a linear turbulence model, then formally, the RANS
equations and the resulting turbulent viscosity influence the RANS
equations differ from the Navier–Stokes equations by the additional
equations and how existing turbulence models can be modified to
unknown turbulent viscosity 𝜇𝑡 representing an addition to the lami-
obtain a well-posed problem. Therefore, trying to improve and adapt
nar viscosity 𝜇𝑙 , typically given by Sutherland’s law. For closure, the
these models for certain flows and in general is still an active field of
applied turbulence model, given by an additional set of equations, is
research.
constructed such that the resulting effective viscosity 𝜇ef f = 𝜇𝑙 + 𝜇𝑡 may
In addition to the classic approach attempts are now being made
satisfy that
to increasingly use machine learning methods to exploit the potential
(a) a steady-state solution of the RANS equations exists and it is of existing turbulence models. To adapt a given turbulence model for
unique,

E-mail address: [email protected].

https://doi.org/10.1016/j.compfluid.2024.106438
Received 12 October 2023; Received in revised form 9 August 2024; Accepted 16 September 2024
Available online 19 September 2024
0045-7930/© 2024 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
S. Langer Computers and Fluids 284 (2024) 106438

one or several flow cases, a Field Inversion and Machine Learning for approximating the parameter to be identified sufficiently well for
(FIML) approach can be leveraged to obtain a closed-form correction aerodynamically relevant data, even under noisy data.
for the given turbulence model to improve the predictions [13–16]. The outline of the article is as follows: The governing equations of
For example, for the Spalart–Allmaras turbulence model the main idea, fluid flow as well as the turbulence models are discussed in Section 2.
which can be found in the literature [13,14], is, to manipulate a field The forward and the inverse problem are presented in Section 3.
parameter 𝛽 weighting the production term 𝑃 𝑟 in the source term 𝑄 of Though not a proof, but sustainable and understandable arguments are
the turbulence model, that is given that the inverse problem for the considered example is severely
ill-posed and it is at least ill-conditioned on the dicrete level. The
𝑄 = 𝛽𝑃 𝑟 − 𝐷𝑒 + 𝐷𝑖, (1)
iteratively regularized Gauss–Newton method for solving the inverse
where 𝛽 ≡ 1 yields the original unmodified model of Spalart and problem is discussed in Section 4. A numerical example of transonic
Allmaras, and 𝐷𝑒 and 𝐷𝑖 denote the destruction and diffusion source flow over an RAE2822 for the successful reconstruction of a turbulent
term, respectively. Then, one determines 𝛽 such that the corresponding viscosity 𝜇𝑡 is presented in Section 5. The article concludes with a
solution of the RANS equations fits, for example, some 𝑐𝑝 -distribution perspective on how the approach proposed can be used to modify
‘‘better’’ with respect to some measure (norm). Such an approach has existing turbulence models so that they may show more suitable results
been applied successfully for several examples [17]. for a large number of test cases.
To discuss the fundamental mathematical challenges of this ap-
proach, in this article, we neglect knowledge of a turbulence model, 2. Governing equations and turbulence models
but attempt to directly reconstruct a turbulent viscosity field 𝜇𝑡 from
given data. To do so, we consider the turbulent viscosity field as a 2.1. Governing equations of fluid flow
parameter in the RANS equations, which needs to be identified. Then,
the reconstruction of 𝜇𝑡 , or 𝛽 for the problem above, are classical The fluid dynamics equations of interest take the form
parameter identification problems well-known in the field of Inverse
𝜕𝜌
Problems (IP). Here it is the goal to identify two types of unknowns, the + ∇ ⋅ (𝜌𝐮) = 0, (3a)
𝜕𝑡
parameter 𝜇𝑡 and the state 𝑊 linked via an equation formally written
𝜕(𝜌𝐮)
as + ∇ ⋅ (𝜌𝐮 ⊗ 𝐮) + ∇𝑝 − ∇ ⋅ 𝜏 = 0, (3b)
𝜕𝑡
𝑅(𝑊 ; 𝜇𝑡 ) = 𝑅𝜇𝑡 (𝑊 ) = 0, 𝑅 ∶  ×  → . (2) 𝜕(𝜌𝐸)
+ ∇ ⋅ (𝜌𝐻𝐮) + ∇ ⋅ 𝐪 − ∇𝐮 ∶ 𝜏 = 0, (3c)
𝜕𝑡
Considering the problem formulated above one needs to replace 𝜇𝑡 by 𝛽 where 𝜌, 𝐮 = (𝑢1 , … , 𝑢𝑚 ), 𝑝, 𝐸 and 𝐻 are density, velocity, pressure,
such that 𝑅(𝑊 ; 𝛽) = 𝑅𝛽 (𝑊 ) = 0. Here ,  and  are Hilbert spaces. specific total energy and total enthalpy, respectively. Considering a
Finding a solution of the state Eq. (2) is the direct problem.
Newtonian fluid the viscous stress tensor is given by
For many parameter identification problems it is reasonable to [ ]
𝜕𝑅 2
assume that 𝑅 is Fréchet differentiable and 𝜕𝑊 (𝑊 ; 𝜇𝑡 ) ∶  →  is a 𝜏 = 𝜇ef f ∇𝐮 + (∇𝐮)𝑇 − (∇ ⋅ 𝐮) 𝐈 , (4)
continuous operator with continuous inverse. Under these assumptions, 3
due to the implicit function theorem, we can conclude that (2) has where 𝜇ef f ∶= 𝜇𝑙 +𝜇𝑡 is an effective viscosity determined by the laminar
( )
a unique solution 𝑊 = 𝑊 𝜇𝑡 and we can introduce a well-defined viscosity and an eddy or turbulent viscosity 𝜇𝑡 ≥ 0 defined by some
( )
parameter-to-solution operator 𝛹 ∶  → , 𝜇𝑡 ↦ 𝑊 𝜇𝑡 solving (2). turbulence model. The laminar viscosity is given by Sutherland’s law
An important and well investigated class of applications for such in- ( )3∕2
𝑇 𝑇∞ + 𝑇̄ 𝜌 𝑢 𝐿
verse problems are inverse scattering problems, both acoustic scattering 𝜇𝑙 ∶= 𝜇𝑙,∞ , 𝜇𝑙,∞ ∶= ∞ ∞ , (5)
𝑇∞ 𝑇 + 𝑇̄ 𝑅𝑒
and electromagnetic scattering. In inhomogeneous medium scattering
problems one tries to reconstruct a refractive index with compact sup- whereby 𝜌∞ > 0 and 𝑢∞ > 0 denote some constant reference density
port from farfield patterns [18–23]. These inverse problems arise from and velocity, 𝐿 > 0 is some constant reference length scale, 𝑅𝑒 > 0 is
linear partial differential equations, such as the Helmholtz equation the corresponding Reynolds number, and 𝑇̄ ∶= 110.4K is Sutherland’s
or the Maxwell equations, which are well understood in the sense constant. Furthermore, to determine 𝑝 the fluid is presumed to act as an
that Fréchet differentiability of the operators is established and the ideal gas, that is 𝑝 = 𝜌ℜ𝑇 , where ℜ is the universal gas constant. The
underlying choice of the Hilbert spaces ,  and  is well investigated. conductive heat flux is supposedly described by the Fourier law 𝐪 =
For the RANS equations, where the operator 𝑅 corresponds to −𝜅ef f ∇𝑇 , where 𝜅ef f ∶= 𝜅𝑙 + 𝜅𝑡 is the sum of the laminar conductivity
a boundary-value problem for the governing equations of compress- 𝑐𝑝 𝜇𝑙 𝛾
ible fluid flow, there does not exist a sustainable solution theory. 𝜅𝑙 (𝑊 ) ∶= , 𝑐𝑝 ∶= ℜ , (6)
𝑃 𝑟𝑙 𝛾 −1
In addition, unlike the scattering problems the governing equations
withe the laminar Prandtl number given by 𝑃 𝑟𝑙 ∶= 0.72, 𝛾 is the gas
are nonlinear and of transport-type, making the inverse problems for-
dependent ratio of specific heats, which is given by 1.4 for air, and
mulated above inherently more complex. Furthermore, though it is
𝜇
heuristically understandable that these problems are ill-posed, there 𝜅𝑡 ∶= 𝑐𝑝 𝑡 , 𝑃 𝑟𝑡 ∶= 0.92. (7)
is no proof of this in the strict mathematical sense. Moreover, the 𝑃 𝑟𝑡
question of whether the problems are mildly or severely ill-posed is Throughout the rest of the article the vector of conserved variables for
also not sufficiently answered yet. Using an example, we will show an open and simply-connected domain 𝐷 is summarized as
that there is significant evidence that these problems are possibly
even severely ill-posed. One additional focus of this article is on the 𝑊 ∶ 𝐷 × [0, 𝑇 ) → R𝑚+2 ,
application of the iteratively regularized Gauss–Newton method to 𝑊 (𝑥, 𝑡) ∶= (𝜌(𝑥, 𝑡), 𝜌(𝑥, 𝑡)𝐮(𝑥, 𝑡), 𝜌(𝑥, 𝑡)𝐸(𝑥, 𝑡))𝑇 .
reconstruct both the state 𝑊 and 𝜇𝑡 . In order to stabilize this method,
it was necessary to develop a suitable Sobolev norm for the presented 2.2. Turbulence models
example of a compressible flow around an airfoil. This Sobolev norm is
derived on the basis of a finite volume discretization. Furthermore, by When employing a turbulence model, the equations of (3) are
systematically adding noise to the given data, we will investigate how frequently supplemented by one ore more differential or integral equa-
the method behaves when we couple it with Morozov’s discrepancy tions. For example, the two equation 𝑘-𝜔 models have the form
principle as a truncation criterion. This approach is intended to address (( ) )
𝜕 (𝜌𝑘)
the question of the extent to which the iterative method is suitable + ∇ ⋅ (𝜌𝑘𝐮) = ∇ ⋅ 𝜇𝑙 + 𝜎𝑘 𝜇𝑡 ∇𝑘 + 𝜌𝑄𝑘,(𝑘,𝜔) , (8a)
𝜕𝑡

2
S. Langer Computers and Fluids 284 (2024) 106438

𝜕 (𝜌𝜔) (( ) )
+ ∇ ⋅ (𝜌𝜔𝐮) = ∇ ⋅ 𝜇𝑙 + 𝜎𝜔 𝜇𝑡 ∇𝜔 + 𝜌𝑄𝜔,(𝑘,𝜔) . (8b) 3.2. Data for the inverse problem
𝜕𝑡
where the unknowns 𝑘 and 𝜔 are the turbulence kinetic energy per For the inverse problem, it is our interest to determine the field
unit mass and the specific dissipation rate of the turbulence kinetic parameter 𝜇𝑡 with respect to observed data related to a given state 𝑊 =
energy, respectively. The source terms for the 𝑘 and 𝜔 equations include 𝑊 (𝜇𝑡 ). This observation operator could be, for example, either the iden-
a production (𝑃 𝑟), a destruction (𝐷𝑒) and a diffusion (𝐷𝑖) term, such tity (distributed measurement), or a trace operator to boundary values
that of the solution (boundary measurements). In CFD, these are often
the 𝑐𝑝 -distribution and/or the skin friction coefficient 𝑐𝑓 on the surface
𝑄𝑘,(𝑘,𝜔) ∶= 𝑃 𝑟𝑘,(𝑘,𝜔) − 𝐷𝑒𝑘,(𝑘,𝜔) , 𝑄𝜔,(𝑘,𝜔) ∶= 𝑃 𝑟𝜔,(𝑘,𝜔) − 𝐷𝑒𝜔,(𝑘,𝜔) + 𝐷𝑖𝜔,(𝑘,𝜔) .
of the geometry. Such viewpoint is, for example, discussed in [14,17].
The symbols 𝜎𝑘 and 𝜎𝜔 are either fixed constants or functions describing Typically the data, that is for example these surface distributions,
the transition between a 𝑘-𝜔 and a 𝑘-𝜀 model. For details we refer
(a) come from measurement,
to [3–6,24,25]. The additional variables required for the turbulence
(b) or are artificially generated for test purposes.
model are then used to determine the turbulent viscosity 𝜇𝑡 = 𝜌𝑘∕𝜔.
An example of a one-equation turbulence model is given by the one of In both cases, the data is typically noisy with an unknown noise level 𝛿,
Spalart and Allmaras, for example
𝜕 𝜈̃ 1[ ( ) ] ‖𝑐𝑝𝛿 − 𝑐𝑝† ‖ ≤ 𝛿,
+ ∇ ⋅ (𝜈𝐮)
̃ = ∇ ⋅ (𝜈𝑙 + 𝜈)∇ ̃ 2 + 𝑃 𝑟 − 𝑇 + 𝐷𝑒,
̃ 𝜈̃ + 𝑐𝑏2 (∇𝜈) (9)
𝜕𝑡 𝜎
where the superscript 𝛿 indicates the noisy and † the exact data, respec-
which represents an equation for the auxiliary variable 𝜈,
̃ related to
tively. In this article, to have a sufficient amount of data, we consider
the turbulent viscosity 𝜇𝑡 = 𝜌𝑓𝑣1 𝜈.
̃ In summary, formally, these linear
a distributed measurement artificially generated for test purposes. An
turbulence models represented by differential or integral equations
inverse crime is omitted by generating the data from a finer mesh.
reveal additional quantities in the considered fluid flow equations.
The occurring variables extend the degrees of freedom given by the
3.3. Inverse problem: Parameter-to-solution operator
conservative variables 𝑊 by further unknown quantities

𝑊𝑡 ∶ 𝐷 × [0, 𝑇 ) → R𝑁𝑡 . To formulate the inverse problem as an operator equation, we use

the parameter-to-solution operator 𝐹 ∶  → ,
Here 𝑁𝑡 ∈ N depends on the turbulence model. For the models
𝐹 (𝜇𝑡 ) = 𝑊 (𝜇𝑡 ) − 𝑊 † . (11)
considered we have
†
That is, a complete field state 𝑊 is given, and the corresponding
𝑁𝑡 = 1 for the Spalart–Allmaras model, 𝑊𝑡 = 𝜈̃𝑡
turbulent viscosity 𝜇𝑡† is the field parameter to be reconstructed. Then,
𝑁𝑡 = 2 for the Wilcox 𝑘-𝜔 model and the SST model, 𝑊𝑡 = (𝑘, 𝜔). the inverse problem is to find a root of Eq. (11). When the given data
are a surface distribution, then (11) is replaced by
Note, for example, when using an algebraic model, such as a Baldwin-
( ) ( )
Lomax model [26], 𝑁𝑡 = 0 and no additional unknowns need to be 𝐹 (𝜇𝑡 ) = 𝑃 𝑊 (𝜇𝑡 ) − 𝑃 𝑊 † , (12)
considered.
and 𝑃 denotes the trace operator corresponding to the given surface
data. For example, if a 𝑐𝑝 - or a 𝑐𝑓 -distribution are the given data,
2.3. Boundary-value problem then (12) reads as
( ) ( )
𝐹 (𝜇𝑡 ) = 𝑃𝑐𝑝 𝑊 (𝜇𝑡 ) − 𝑐𝑝† or 𝐹 (𝜇𝑡 ) = 𝑃𝑐𝑓 𝑊 (𝜇𝑡 ) − 𝑐𝑓† .
For a complete boundary-value problem, the Eqs. (3) and (8) or (9)
Even both, as a system of equations, can be considered,
need to be supplemented by boundary conditions. Here, for the no-slip ( ( )) ( † )
boundary we postulate an adiabatic wall boundary condition 𝑃𝑐𝑝 𝑊 (𝜇𝑡 ) 𝑐
𝐹 (𝜇𝑡 ) = ( ) − 𝑝† .
𝜕𝑇 𝑃𝑐𝑓 𝑊 (𝜇𝑡 ) 𝑐
𝐮=0 and =0 𝑓
𝜕𝑛 Note that the state 𝑊 = 𝑊 (𝜇𝑡 ) satisfies (2), so Eq. (11) involves the
and solution of the RANS equations given the parameter 𝜇𝑡 , that is
6𝜇𝑙
𝑘 = 0, lim 𝜔(𝑥 + ℎ𝑛(𝑥))ℎ2 = or 𝜈̃𝑡 = 0. 𝐹 (𝜇𝑡 ) = 𝑅−1 †
𝜇 (0) − 𝑊 . (13)
ℎ→0+ 𝛽𝜌 𝑡

Here 𝑛 = 𝑛(𝑥) denotes the outer unit normal vector. For the outer Here the symbol 𝑅−1 𝜇𝑡 (0) means in practice to approximate a solution
boundary we consider characteristic farfield boundary conditions. of the discretized RANS equations for the given turbulent viscosity 𝜇𝑡 .
Considering the FIML approach (1), then we obtain in a completely
analogous way the equation 𝐹 (𝛽) = 𝑅−1𝛽
(0) − 𝑊 † .
3. An inverse view on data driven turbulence modeling
3.4. Ill-posed character of the field reconstruction
3.1. Forward problem
First of all, let us clarify what kind of classification of problem and
degree of ill-posedness we are dealing with. Since a profound theory is
Formally, the forward problem is formulated as follows: We con- missing, we at least want to formulate heuristic arguments. Using the
sider 𝑅 to be the residual operator corresponding to the set of equations implicit function theorem, we compute the derivative
for a boundary-value problem for the RANS equations, 𝜇𝑡 is the field
𝑑𝑅−1 ( )
𝑑𝑊 𝜇𝑡 (0) 𝜕𝑅(𝜇𝑡 , 𝑊 ) −1 𝜕𝑅
parameter of interest and 𝑊 a state vector. Then these are linked via 𝐹 ′ [𝜇𝑡 ] = = =− . (14)
the equation 𝑑𝜇𝑡 𝑑𝜇𝑡 𝜕𝑊 𝜕𝜇𝑡
𝜕𝑅(𝜇 ,𝑊 ) 𝜕𝑅
𝑅(𝑊 ; 𝜇𝑡 ) = 𝑅𝜇𝑡 (𝑊 ) = 0. (10) Both required derivatives 𝜕𝑊𝑡 and 𝜕𝜇 can be determined explicitly.
𝑡
To understand the character of problem (13) we recall the following
The solution of this state Eq. (10) for a given fixed parameter 𝜇𝑡 can be classification of ill-posed problems [21,27,28]. An ill-posed problem is
interpreted as the direct problem. called

3
S. Langer Computers and Fluids 284 (2024) 106438

𝑑𝑊
Fig. 1. Used mesh and decay behavior of the singular values of 𝐹 ′ [𝜇𝑡 ] = 𝑑𝜇𝑡
.

(a) mildly ill-posed, if the singular values 𝜎𝑗 of 𝐹 ′ [𝜇𝑡 ] decay at a To approximately solve the operator Eq. (13) we substitute 𝐹 by its
polynomial rate, i.e. linear approximation. Given an initial guess 𝜇𝑡,0 , in each Newton step
the least squares problem
𝑐𝑚 𝑗 −𝛼 ≤ 𝜎𝑗 ≤ 𝐶𝑚 𝑗 −𝛼 , 𝛼, 𝑐𝑚 , 𝐶𝑚 > 0
‖𝐹 ′ [𝜇𝑡,𝑛 ]ℎ𝑛 + 𝐹 (𝜇𝑡,𝑛 )‖2 = min! (15)
(b) severely ill-posed, if the singular values 𝜎𝑗 of 𝐹 ′ [𝜇𝑡 ] decay at an ℎ𝑛

exponential rate, i.e. needs to be solved and the update is computed by 𝜇𝑡,𝑛+1 = 𝜇𝑡,𝑛 + ℎ𝑛 .
𝛽 𝛽 This approach is well-known as Gauss–Newton method. In the case
𝑐𝑒 𝑒−𝑐𝑠 𝑗 ≤ 𝜎𝑗 ≤ 𝐶𝑒 𝑒−𝑐𝑠 𝑗 𝛽, 𝑐𝑠 , 𝑐𝑒 , 𝐶𝑒 > 0
where (13) is ill-posed, it is important to consider the situation where
The degree of ill-posedness depends on the differential operator of the the given data 𝑊 † is replaced by noisy data 𝑊 𝛿 satisfying
underlying differential equation. For the example presented in Sec-
‖𝑊 † − 𝑊 𝛿 ‖ ≤ 𝛿.
tion 5, Eq. (14) is used to compute the singular values 𝜎𝑗 . Fig. 1 (right)
( )
shows the decay behavior of 𝜎𝑗 𝐹 ′ [𝜇𝑡 ] 𝑗=1,2,… for an iterate in a neigh- In this case a straightforward implementation of the Gauss–Newton
borhood of the solution. The decay at an exponential rate of these method usually fails and does not lead to a good reconstruction of
singular values is obvious. This observation serves as an indicator that the solution after several Newton steps. One reason for the failure
solving (11) is, at least for the example considered, a severely ill- of the Gauss–Newton method in this situation is the ill-posedness of
posed problem and that the stability condition for a well-posed problem the least squares problem (15), which is inherited from the original
is violated. This is of course not a mathematical proof in the strict operator Eq. (13). Thus, to perform one Newton step some sort of
sense, but in practice it means that we have to solve systems of linear regularization has to be employed when solving (15). In this article we
equations with very large condition numbers. Hence, regularization consider a regularized Gauss–Newton method, where instead of (15),
needs to be employed. the regularized least squares problem

4. Discretization and solution algorithm ‖𝐹 ′ [𝜇𝑡,𝑛

𝛿 𝛿
]ℎ𝑛 + 𝐹 (𝜇𝑡,𝑛 )‖2 + 𝛾𝑛 ‖ℎ𝑛 ‖2 = min! (16)
ℎ𝑛

is solved in each Newton step. For the iteratively regularized Gauss–

4.1. Discretization strategy
Newton method we have

For a detailed description of the discretization strategy and the ‖𝐹 ′ [𝜇𝑡,𝑛

𝛿 𝛿
]ℎ𝑛 + 𝐹 (𝜇𝑡,𝑛 )‖2 + 𝛾𝑛 ‖ℎ𝑛 + 𝜇𝑡,𝑛
𝛿
− 𝜇𝑡,0 ‖2 = min! (17)
ℎ𝑛
solution algorithm for the RANS equations we refer to [29–31]. It is
based on a finite-volume formulation where a median dual grid forms Here
the control volumes with the unknowns at vertices of the primary grid.
𝛾𝑛 = 𝛾0 𝛾 −𝑛 > 0, 0 ≤ 𝛾 ≤ 1, 𝛾0 > 0 (18)
The dual grid is created in a preprocessing step. Throughout this paper,
the set of points in the dual grid where the unknowns are located is denotes a regularization parameter. This iterative regularization
denoted by 𝐺 ∶= {1, … , 𝑁} and the unknowns are given by 𝑊1 , … , 𝑊𝑁 . method can be interpreted as a common Newton method where in
The neighbors of some point 𝑖 ∈ 𝐺 are described by  (𝑖), and the each step Tikhonov regularization is applied to the linearized equation.
corresponding face between the cells 𝑖 and 𝑗 is denoted by 𝑒𝑖𝑗 . The problems (16) and (17) are well-posed, in particular there exists a
The set of Eqs. (3) are integrated and the resulting set of integral uniquely defined minimizer ℎ† . In the numerical experiments it turned
equations are discretized for each control volume. The convective part out that solving (17) had better stability compared to (16), in particular
of the equations is discretized using a matrix-valued artificial viscosity the initial regularization parameter 𝛾0 could be significantly reduced.
scheme, which is similar and close to some kind of Roe-scheme. On Due to the severe ill-posedness and resulting instability of the
a sufficiently regular mesh this discretization can be shown to be of problem, we still need to include into our process a stopping criterion.
second order. A pressure switch is included into the artificial viscosity Here, for simplicity, we use the discrepancy principle of Morozov [32],
part reducing the method to first order in a neighborhood of a shock. that is we stop the iteration at the first index 𝑀 satisfying
Gradients are reconstructed using the method of Green-Gauss.
‖ ( 𝛿 )‖ ‖
‖ −1
‖
‖
‖𝐹 𝜇 ‖ 𝑅 𝛿 (0) − 𝑊 𝛿 ‖ ≤ 𝜏𝛿,
‖ 𝑡,𝑀 ‖ = ‖
‖ ‖
𝜏 > 1. (19)
‖ ‖ ‖ 𝑡,𝑀
𝜇
‖
4.2. An iterative regularized Gauss–Newton method
Typically, a violation of such stopping criterion may lead to oscillations
The derivation and realization of the regularized Gauss–Newton and instabilities in the solution. This is also confirmed for the example
method follows [19,21–23]. discussed in Section 5 and shown in Figs. 5 and 6.

4
S. Langer Computers and Fluids 284 (2024) 106438

3 ⟨
∑ ⟩
4.3. Choice of Hilbert scales
= ⟨𝐵ℎ, 𝐵ℎ⟩𝐿2 (𝐷) + 𝐵𝑥𝑗 ℎ, 𝐵𝑥𝑗 ℎ , (21)
𝐿2 (𝐷)
𝑗=1
So far, nothing has been said about the evaluation of the norms in
the image space and in the pre-image space, that is ‖ . ‖ and ‖ . ‖ . where
Such distinction is important and in general the expectation is that it ⎛√ ( )⎞
1
has an effect on both the accuracy of the approximate solution and the 𝐵 ∶= diag ⎜ ( ) + vol 𝐷𝑖 ⎟
⎜ vol 𝐷 2 ⎟
success of the algorithm. ⎝ 𝑖 ⎠
It is well-known from other applications that for a successful imple- and we have used the representation of the function ℎ in the finite-
mentation of iterative regularization methods that it is rather important volume context,
to work in problem-suited function spaces. If available, these should of
∑
𝑁
course be based on a well-founded theory. For example, for scattering ℎ(𝑥) = ℎ𝑗 𝟏𝐷𝑗 (𝑥).
problems [19,23] Fréchet differentiability of the occurring operators 𝑗=1
is proven with respect to the ‖ . ‖𝐶 1 −norm. Then, for a numerical To implement the Levenberg–Marquardt algorithm or the IRGNM we
implementation one goes over to a suitable Sobolev norm, since one use the formulations (16) and (17), respectively. Using  = 𝐿2 (𝐷)
needs a Hilbert space structure. and  = 𝐻 1 (𝐷), the least squares problems are given by
In our implementation we observed in general unstable or less [ ] ( )
accurate solution for the choice  = 𝐿2 (𝐷) and the algorithm failed. ⎛𝐹 ′ 𝝁𝛿𝑡,𝑛 ⎞ ⎛−𝐹 𝝁𝛿𝑡,𝑛 ⎞
Unfortunately, there is no reliable theory for the RANS equations. Since ⎜ √ ⎟ ⎜ √ ⎟
⎜ 𝛾𝑛 𝐵 ⎟ ⎜ 𝛾𝑛 𝐵𝑏𝑛 ⎟
gradients and derivatives play an essential role, we have endeavored to ⎜√ ⎟ ⎜√ ⎟
⎜ 𝛾𝑛 𝐵𝑥1 ⎟ ℎ𝑛 = ⎜ 𝛾𝑛 𝐵𝑥1 𝑏𝑏 ⎟ . (22)
incorporate them into the implementation of the algorithm. To do this,
⎜√ ⎟ ⎜√ ⎟
we implement an 𝐻 1 (𝐷)-norm in the sense of a finite volume method, ⎜ 𝛾𝑛 𝑥2 ⎟
𝐵 ⎜ 𝑛 𝑥2 𝑛 ⎟
𝛾 𝐵 𝑏
that is a discrete realization of ⎜√ ⎟ ⎜√ ⎟
⎝ 𝛾𝑛 𝐵𝑥 ⎠ ⎝ 𝛾𝑛 𝐵𝑥 𝑏𝑛 ⎠
3 3
‖𝑔‖2 = ‖𝑔‖2 2 + ‖grad 𝑔‖2 2 = |𝑔(𝑥)|2 𝑑𝑥 + |grad 𝑔(𝑥)|2 𝑑𝑥. Choosing 𝑏𝑛 = 0 corresponds to (16), which is the Levenberg–Marquardt
𝐻 1 (𝐷) 𝐿 (𝐷) 𝐿 (𝐷) ∫𝐷 ∫𝐷
𝛿 is the IRGNM, which is preferred here and
algorithm, 𝑏𝑛 = 𝜇𝑡,0 − 𝜇𝑡,𝑛
In a finite-volume context the function 𝑔 is approximated by
used for the numerical example in Section 5. To reduce the complexity
{
∑𝑁
1, 𝑥 ∈ 𝐷𝑗 , of the algorithm in our implementation we have replaced Eq. (22) by
𝑔(𝑥) = 𝑔𝑗 𝟏𝐷𝑗 (𝑥), where 𝟏𝐷𝑗 (𝑥) = [ ] ( )
𝑗=1
0, 𝑒𝑙𝑠𝑒. ⎛ 𝐹 ′ 𝝁𝛿 ⎞ ⎛ −𝐹 𝝁𝛿 ⎞
⎜ 𝑡,𝑛 ⎟
ℎ = ⎜ 𝑡,𝑛 ⎟
. (23)
Then, in our implementation the derivative is approximated using a ⎜√𝛾 𝐵 total ⎟ 𝑛 ⎜√𝛾 𝐵 total 𝑏 ⎟
⎝ 𝑛 ⎠ ⎝ 𝑛 𝑛⎠
Green-Gauss method, that is for 𝑘 = 1, 2, 3 we have
( )GG where 𝐵 total = 𝐵 + 𝐵𝑥1 + 𝐵𝑥2 + 𝐵𝑥3 . This corresponds to exchanging the
𝜕𝑔 1 ∑ 𝑛𝑘,𝑒𝑖𝑗 ( )
= svol(𝑒𝑖𝑗 ) 𝑔𝑖 + 𝑔𝑗 , norm in (21) for an expression of the form
𝜕𝑥𝑘 𝐷 vol(𝐷𝑖 ) 2 ⟨ ⟩
𝑖 𝑗∈ (𝑖)
‖ℎ‖2,mod = 𝐵 total ℎ, 𝐵 total ℎ .
and hence
( )GG ⎧∑ svol(𝑒𝑖𝑗 )𝑛𝑘,𝑒𝑖𝑗 , 𝑚 = 𝑖, 4.4. Linear solvers
𝜕 𝜕𝑔 1 ⎪ 𝑗∈ (𝑖)
= ⎨ svol(𝑒𝑖𝑚 ), 𝑚 ∈  (𝑖), (20)
𝜕𝑔𝑚 𝜕𝑥𝑘 𝐷𝑖 2vol(𝐷𝑖 ) ⎪ To solve the linear least-squares problems two different approaches
⎩ 0, else.
are used. The first is to reformulate (23) by the normal equation
Eqs. (20) represent the sparse matrices ( [ ]𝑇 [ ])
( )𝑇
(( )GG ) ( ( )GG ) 𝛾𝑛 𝐵 total 𝐵 total + 𝐹 ′ 𝝁𝛿𝑡,𝑛 𝐹 ′ 𝝁𝛿𝑡,𝑛 ℎ𝑛
𝑑 𝜕𝑔 𝜕 𝜕𝑔
𝐵𝑥𝑘 ∶= = , 𝑘 = 1, 2, 3, [ ]𝑇 ( ) ( )𝑇
𝑑𝑔 𝜕𝑥𝑘 𝜕𝑔𝑚 𝜕𝑥𝑘
𝐷𝑖 𝑖=1,…,𝑁 𝐷𝑖 𝑖,𝑚=1,…,𝑁 = −𝐹 ′ 𝝁𝛿𝑡,𝑛 𝐹 𝝁𝛿𝑡,𝑛 + 𝛾𝑛 𝐵 total 𝐵 total 𝑏𝑛 . (24)
satisfying
Then this linear system is directly solved using an 𝐿𝑈 -decomposition.
( )GG
(( ⎛ 𝜕𝑔 ⎞ The second method we used is to solve (22) using an 𝑄𝑈 -
)GG ) ⎛ 1 ⎞ ⎜ 𝜕𝑥𝑘 𝐷1 ⎟
𝑔 decomposition. Of course, both methods are equivalent in the sense that
𝑑 𝜕𝑔 ⎜ ⎟ ⎜ ⎟,
𝐵𝑥𝑘 𝑔 = ⋮ = ⋮ 𝑘 = 1, 2, 3. they determine the unique minimizer of (23). Solving (24) involves the
𝑑𝑔 𝜕𝑥𝑘 𝐷 ⎜ ⎟ ⎜( )GG ⎟
𝑖 𝑖=1,…,𝑁 ⎝𝑔𝑁 ⎠ ⎜ 𝜕𝑔 ⎟ evaluations of
⎝ 𝜕𝑥𝑘 𝐷𝑁 ⎠ ( )
[ ] 𝛿 , 𝑊 ) −1
𝜕𝑅(𝜇𝑡,𝑛
Application of 𝐵𝑥𝑘 introduces a discrete possibility to evaluate the norm ′ 𝛿 𝜕𝑅
𝐹 𝝁𝑡,𝑛 = , (25)
of the derivative of 𝑔, that is 𝜕𝑊 𝜕𝜇𝑡
‖ (( )GG ) ‖2 [ ]𝑇 ( )𝑇 ( 𝜕𝑅(𝜇 𝛿 , 𝑊 ) )−𝑇
‖ 𝑑𝑔 ‖2 ‖ 𝜕𝑔 ‖ ⟨ ⟩ 𝜕𝑅 𝑡,𝑛
‖ ‖ ‖
‖ 𝑑𝑥 ‖ 2 ≈ ‖
‖
‖ = 𝐵𝑥𝑘 𝑔, 𝐵𝑥𝑘 𝑔 𝐿 (𝐷) , 𝑘 = 1, 2, 3. 𝐹 ′ 𝝁𝛿𝑡,𝑛 = . (26)
‖ 𝑘 ‖𝐿 (𝐷) ‖ 𝜕𝑥𝑘 𝐷𝑖 ‖ 2 𝜕𝜇𝑡 𝜕𝑊
‖ 𝑖=1,…,𝑁 ‖𝐿2 (𝐷)

And as a consequence we obtain This means that we need to determine the inverse of the matrix
( 3 ( ) ) 3 ⟨
𝛿 ,𝑊 )
𝜕𝑅(𝜇𝑡,𝑛
∑ 𝜕𝑔 2 ∑ ⟩ ∈ R5𝑁×5𝑁 ,
‖grad 𝑔‖2 2 = 𝑑𝑥 = 𝐵𝑥𝑗 𝑔, 𝐵𝑥𝑗 𝑔 . 𝜕𝑊
𝐿 (𝐷) ∫𝐷 𝜕𝑥𝑗 𝐿2 (𝐷)
𝑗=1 𝑗=1 which is rather time consuming. Using
Furthermore, to penalize for the influence of the strongly varying mesh ( )−𝑇 ( ) 𝑇
𝛿 ,𝑊 )
𝜕𝑅(𝜇𝑡,𝑛 ⎛ 𝜕𝑅(𝜇𝛿 , 𝑊 ) −1 ⎞
cell sizes, which exist due to the strong anisotropies needed in meshes
=⎜ ⎟
𝑡,𝑛
(27)
for simulating turbulent flows, we finally define the discrete norm for 𝜕𝑊 ⎜ 𝜕𝑊 ⎟
⎝ ⎠
the pre-image space by,
only the inverse of the original matrix needs to be determined. When
∑𝑁 ( )
∑3 ⟨ ⟩
1 applying the 𝑄𝑈 -decomposition only (25) is required. Since the de-
‖ℎ‖2 = + vol(𝐷𝑖 ) ℎ2𝑖 + 𝐵𝑥𝑗 ℎ, 𝐵𝑥𝑗 ℎ
vol(𝐷𝑖 )2 𝐿2 (𝐷) termination of this inverse represents the main complexity of the
𝑖=1 𝑗=1

5
S. Langer Computers and Fluids 284 (2024) 106438

algorithm, both application of the normal Eq. (24) or solving the that 𝑇 sets of data 𝑔1 , … , 𝑔𝑇 are given. Similar as (12) we can write
least-squares problem (23) directly using a QU-decomposition were these as
comparable in required time. ( ( )−1 ) ⎞
⎛ 1
In summary, the main complexity of the proposed implementation ⎜ 1𝑃 𝑅 (0) ⎟
𝜇1 ,…,𝜇𝑠 ⎛𝑔 ⎞
⎜ ⎟ ⎜ 1⎟
of algorithm (15) is given by: 𝐹 (𝜇1 , … , 𝜇𝑠 ) = ⎜( ⋮ ⎟ − ⋮ , (28)
⎜ ( )−1 )⎟ ⎜ ⎟
𝑇 ⎝𝑔𝑇 ⎠
𝛿 ) in (23) we need to solve the RANS equations
(a) To determine 𝐹 (𝜇𝑡,𝑛 ⎜ 𝑃𝑇 𝑅𝜇1 ,…,𝜇𝑠 (0) ⎟
𝛿 . ⎝ ⎠
for the given iterate 𝜇𝑡,𝑛
𝛿 ,𝑊 )
𝜕𝑅(𝜇𝑡,𝑛 which represents in general a large-scale, possibly ill-posed nonlinear
(b) For each step we need to determine the inverse of 𝜕𝑊
. system of equations. For example, following either the FIML approach
(c) For each step we need to solve the linear least-squares prob- or the reconstruction of a turbulent viscosity field, then one parameter
lem (23). either 𝛽 or 𝜇𝑡 needs to be determined such that 𝑇 equations must be
satisfied simultaneously, that is (28) simplifies to
( ( ) )
5. Numerical example ⎛ 1
−1 ⎞
⎜ 1𝑃 𝑅 𝜇𝑡 (0) ⎟ ⎛𝑔 ⎞
⎜ ⎟ ⎜ 1⎟
As an example we consider turbulent flow over the RAE 2822 airfoil 𝐹 (𝜇𝑡 ) = ⎜( ⋮ )⎟ − ⎜ ⋮ ⎟ , (29)
⎜ ( ) −1 ⎟ ⎝𝑔𝑇 ⎠
at 𝑀 = 0.73, angle of attack 𝛼 = 2.79◦ and 𝑅𝑒 = 6.5 ⋅ 106 . This case ⎜ 𝑃𝑇 𝑅𝜇𝑡 𝑇 (0) ⎟
features a shock on the upper part of the airfoil. At this early stage ⎝ ⎠
of these investigations the computations have been done on a rather or
( ( ) )
coarse mesh of size 160 × 32 shown in Fig. 1 (left). As initial guess we ⎛ 1
−1 ⎞
use the turbulent viscosity 𝜇𝑡,0 computed by the same geometry and the ⎜ 𝑃1 𝑅𝛽 (0) ⎟
⎛𝑔 ⎞
⎜ ⎟ ⎜ 1⎟
flow conditions 𝑀 = 0.75, angle of attack 𝛼 = 2.81◦ and 𝑅𝑒 = 6.2 ⋅ 106 . 𝐹 (𝛽) = ⎜( ⋮ ⎟ − ⎜ ⋮ ⎟. (30)
( )−1 )
The regularization parameter (18) is chosen by 𝛾0 = 10, 𝛾 = 21 . ⎜ 𝑇
⎟ ⎝𝑔𝑇 ⎠
⎜ 𝑇𝑃 𝑅 𝛽
(0) ⎟
To show the applicability of the method we first have used exact ⎝ ⎠
data. Fig. 2 shows the convergence histories, on the left of For both (29) and (30) we look for a scalar function 𝜇𝑡 or 𝛽, such
‖𝐹 (𝜇𝑡,𝑛
𝛿 )‖
,𝑛=0,1,2,… and on the right the corresponding aerodynamic that 𝑇 equations are satisfied simultaneously, at least in the sense of
coefficients of lift and drag. The plotted residual is normalized and minimizing the least-squares problem ‖𝐹 (𝜇𝑡 )‖ → min! or ‖𝐹 (𝛽)‖ →
given by min!, that is
( )
⎛ ⎞ ⎛ 𝛿 ⎞ 𝑇 ‖( ( )−1 ) ‖ 𝑇 ‖( ( )−1 ) ‖
∑ ∑
⎜ ‖𝑊 𝜇𝑡,𝑛 − 𝑊 ‖ ⎟
𝛿
⎜ ‖𝐹 (𝜇𝑡,𝑛
𝛿 )‖
⎟ ‖ ‖ ‖ ‖
min ‖ 𝑃𝑘 𝑅𝑘𝜇𝑡 (0) − 𝑔𝑘 ‖ or min ‖ 𝑃𝑘 𝑅𝑘𝛽 (0) − 𝑔𝑘 ‖ .
log ⎜ ( ) ⎟ = log ⎜ ( ) ⎟. 𝜇𝑡 ‖ ‖ 𝛽 ‖ ‖
𝑘=1 ‖ ‖ 𝑘=1 ‖ ‖
⎜ ‖𝑊 𝜇𝑡,0𝛿 − 𝑊 †‖ ⎟ ⎜ ‖𝑊 𝜇𝑡,0
𝛿 − 𝑊 †‖ ⎟
⎝ ⎠ ⎝ ⎠ This can also be interpreted the other way around. For example,
The convergence follows an expected behavior of the IRGNM [33]. This considering the original Spalart–Allmaras model (9), then we obtain
example shows that it is in principle possible to directly reconstruct the 𝑇 ‖( ( )−1 ) ‖ 𝑇 ‖( ( )−1 ) ‖
∑ ‖ ‖ ∑ ‖ ‖
parameter 𝜇𝑡 given a sufficient number of prescribed data. Additionally, min ‖ 𝑃𝑘 𝑅𝑘𝛽 (0) − 𝑔𝑘 ‖ ≤ min ‖ 𝑃𝑘 𝑅𝑘𝛽 (0) − 𝑔𝑘 ‖
𝛽 ‖ ‖ 𝛽≡1 ‖ ‖
𝑘=1 ‖ ‖ 𝑘=1 ‖ ‖
Fig. 3 demonstrates that for these exact data the surface pressure (left) 𝑇 ‖(
∑‖ ( )−1 ) ‖
and skin-friction (right) distributions are perfectly recovered. ‖
= ‖ 𝑃𝑘 𝑅𝑘𝜇 =𝜇 (SA model) (0) − 𝑔𝑘 ‖ .
‖ 𝑡 𝑡 ‖
In a second computation we have added about 10% noise to the 𝑘=1 ‖ ‖
data. The final iterate was selected by (19) with 𝜏 = 1.5. Fig. 5 shows (31)
the reconstructed pressure (left) and skin-friction (right) as well as
In this sense, the original Spalart–Allmaras model is only one instan-
these distributions perturbed by noise. As expected, the regularization
tiation of a more general representation of the model including one
has an effective smoothing effect while the shock position is still
or even more parameters. On the other hand, it is also clear that in
perfectly resolved. Moreover, Figs. 5 and 6 also show the pressure and
practice one will only ever find one optimizer for the given data and
skin-friction distributions when violating the stopping criterion (19),
conditions, and it is unclear whether this minimizer is unique and also
that is, we have plotted the 15th and 20th iterate. One clearly notices
applies to other data sets. The argument can be extended to other
the influence of the noisy data, the pressure distribution is perturbed
turbulence models. For example, we can simply estimate
and the skin-friction distribution starts to oscillate. In particular, as
∑𝑇 ‖( ( )−1 ) ‖
expected, the skin-friction in a neighborhood of the shock is signif- ‖ ‖
min ‖ 𝑃𝑘 𝑅𝑘𝜇 (0) − 𝑔𝑘 ‖
icantly amplified. As expected from theory, for the 15th iterate the 𝜇𝑡 ‖ 𝑡 ‖
𝑘=1 ‖ ‖
regularization parameter 𝛾𝑛 is too small to balance the data noise error.
∑ ‖( ( )−1 ) ‖
⎧ 𝑇 ‖ ‖⎫
After 20 iterations the obtained solution is completely worthless. The ⎪ 𝑘=1 ‖‖ 𝑃 𝑅 𝑘 (0) − 𝑔𝑘 ‖ ⎪
𝑘 𝜇𝑡 =𝜇𝑡 (SA model) ‖
corresponding convergence history for noisy data is shown in Fig. 4. ⎪ ‖ ‖⎪
≤ min ⎨ ⎬. (32)
⎪∑𝑇 ‖ ( ( )−1 ) ‖⎪
‖ ‖
⎪ 𝑘=1 ‖ 𝑃𝑘 𝑅𝜇 =𝜇 (𝑘-𝜔 model)
𝑘 (0) − 𝑔𝑘 ‖⎪
6. Machine learning of a turbulent viscosity ⎩ ‖ 𝑡 𝑡 ‖⎭
‖ ‖
That is, among all possible turbulent viscosities 𝜇𝑡 that minimize the
6.1. Turbulence modeling as an inverse problem initial minimization problem ((29) (left)), the SA model and the 𝑘-𝜔
model represent only two possible variants. Then, a given turbulence
From a general point of view, one is interested in learning a tur- model is only one way of determining an approximation to this min-
bulent viscosity or improving an existing turbulence model for a large imizer, which, depending on the given test case, sometimes reflects
number of different test cases. For example, this could be either for the the available data better and sometimes worse. Interpreting the topic
same geometry and a number of different flow conditions or even a of turbulence modeling in this way, it corresponds mathematically to
change in the geometry can be considered. Then we can generalize the the inverse problem formulated in Sections 3.2 and 3.3. Though not
considered idea in this article in the following way. Assume that we given a proof, due to the nature of the problem it can be assumed to
want to reconstruct (learn) the field parameters 𝜇1 , … , 𝜇𝑠 , 𝑠 ∈ N and be ill-posed, in general. Because of (32) and (31) we can also assume

6
S. Langer Computers and Fluids 284 (2024) 106438

Fig. 2. Convergence history for the IRGNM using exact data and convergence of lift and drag coefficient.

Fig. 3. Reconstructed pressure (left) and skin-friction (right) distribution for exact data.

Then we have not only one but six free parameters to minimize (28),
that is we can minimize ‖ ‖
‖𝐹 (𝑟11 , 𝑟12 , 𝑟13 , 𝑟21 , 𝑟23 , 𝑟33 )‖.

6.2. Computational complexity

Note that (28), (29) and (30) represent in general large-scale, non-
linear ill-posed problems, which, in addition to the algorithmic chal-
lenges of finding approximate solutions, also contain many technical
challenges. Even on today’s High Performance Cluster infrastructure, it
could be difficult to provide the necessary storage space with a naive
approach, if 𝑇 is large. On the other hand, only if 𝑇 is large enough,
that is sufficiently data is available, we may expect a representative
determined parameter which can be applied to a large number of test
cases.
If implementation and solving the problem does not appear pos-
Fig. 4. Convergence history for the IRGNM using noisy data. sible due to technical requirements, another option is to examine,
weight and summarize the data in advance. If we express this process
mathematically by another projection operator 𝑃 , we can solve a
projected equation instead of solving (28), (29) or (30), that is for
example ‖𝑃 𝐹 ‖ → min𝜇1 ,…,𝜇𝑠 !. Such an approach can possibly reduce the
number of unknowns, such that technical feasibility becomes possible.
that minimizing a parameter in a given turbulence model is probably
generally an easier task than directly minimizing for the turbulent
viscosity, since the basic turbulence model serves already as some kind 6.3. Generalization to a variety of fluid flow test cases
of a priori knowledge.
In order to find a better minimizer, one obvious way is to allow for a So far the proposed methodology depends on the mesh in the sense
vector-valued function instead of the scalar function 𝜇𝑡 . For example, a that the specific parameter is determined at the grid nodes. If we want
natural generalization, instead of using a linear turbulence model, is to to carry over the determined parameters 𝜇1 , … , 𝜇𝑠 to other geometries,
determine the six components (𝑟𝑖𝑗 )1≤𝑖<𝑗≤3 of the Reynolds stress tensor. then we have to get away from considering this as a field parameter for

7
S. Langer Computers and Fluids 284 (2024) 106438

Fig. 5. Reconstructed pressure (left) and skin-friction (right) distribution for noisy data.

Fig. 6. Reconstructed pressure (left) and skin-friction (right) distribution for noisy data.

a given mesh. Instead we need to find a suitable ansatz function. That as a parameter in the RANS equations and to determine it on the basis
is we consider of the given data. The feasibility of this approach was demonstrated
for a well-known example in CFD, namely a turbulent flow over the
𝜇𝑗 = 𝜇𝑗 (𝜔𝑠1 , … 𝜔𝑠𝑁 ), 𝑗 = 1, … , 𝑠,
RAE2822 airfoil at transonic conditions where the flow on the upper
and 𝜔𝑠1 , … , 𝜔𝑠𝑁 represent the coefficients in the ansatz function. For side of the airfoil has a shock. To solve the problem the iteratively
example, the approach presented here can be adapted for this directly. regularized Gauss–Newton method (17) was applied together with the
Then the derivative needs to be adjusted to the ansatz coefficients, that discrepancy principle (19) as a stopping criterion. It was shown that for
𝜕𝑅 noisy data this stopping criterion leads to reasonable reconstructions of
is 𝜕𝜇 is replaced by
𝑡
the turbulent viscosity such that aerodynamic data such as 𝑐𝑝 and 𝑐𝑓
𝜕𝑅 𝜕𝑅 𝜕𝜇𝑡 distributions are well represented. Moreover, it was also demonstrated
= , 𝜇𝑡 = 𝜇𝑡 (𝜔1 , … , 𝜔𝑑 ), 𝑑 ∈ N.
𝜕(𝜔1 , … , 𝜔𝑑 ) 𝜕𝜇𝑡 𝜕(𝜔1 , … , 𝜔𝑑 ) that in particular for noisy data a violation of the stopping criterion
Hence, additionally the derivative of the ansatz function with respect blurrs the turbulent viscosity 𝜇𝑡 such that for the given test case strong
to its coefficients is required. Possible ansatz functions which can oscillations in particular in a neighborhood of the shock in the 𝑐𝑓
be considered in future work are for example convolutional neural distribution occur.
networks. To solve such problem based on a finite-volume discretization, a
methodology was suggested to implement on a discrete level suited
7. Conclusion and future work function spaces approximating a Sobolev 𝐻 1 -norm, which in our im-
plementation was crucial for the success of the iteratively regularized
In this article, it was shown that the relatively new approach Gauss–Newton method. Neglecting such gradient information in the
of modifying existing turbulence models by the introduction of field formulation and regularization of the problem typically gave instabil-
parameters can be interpreted as an application of parameter identifi- ities in our implementation of the algorithm and the reconstructions
cation problems studied for a long time in the field of Inverse Problems. showed up strong oscillations.
For the examples shown, it was demonstrated at least heuristically, that Beyond that, based on the interpretation of the FIML approach as a
these problems can be severely ill-posed, on the discrete level at least parameter identification problem for the RANS equations, in Section 6
ill-conditioned, in the sense that the singular values of the linearized it was discussed in which way this technique can possibly be exploited
operator tend at an exponential decay to zero. if one aims for a general improvement of a given turbulence model
Moreover, we have demonstrated that it is possible, provided suffi- or if one wants to represent a turbulent viscosity 𝜇𝑡 directly with the
cient data are available, to interpret the turbulent viscosity 𝜇𝑡 directly help of ansatz functions for a large number of test cases. Due to the

8
S. Langer Computers and Fluids 284 (2024) 106438

possibly ill-posed character of the problem and the large number of References
required data, the determination of the unknown parameter seems
very challenging, since many large-scale problems need to be solved [1] Spalart PR, Allmaras SR. A one-equation turbulence model for aerodynamic
flows. In: AIAA computational fluid dynamics conference. Conference proceeding
at the same time to approximate a minimizer of a corresponding least-
series, AIAA; 1992, number 1992-439.
squares problem. This might even be infeasible due to the numerical [2] Allmaras Steven R, Johnson Forrester T, Spalart Philippe R. Modifications
complexity and memory requirements. A first way out would therefore and clarifications for the implementation of the Spalart-Allmaras turbulence
possibly be to use the approach only for an improvement of a certain model. In: International conference on computational fluid dynamics 7, hawaii.
class of problems in order to try to improve the underlying model for Conference proceeding series, 2012, number ICCFD7-1902.
[3] Wilcox David C. Reassessment of the scale-determining equation for advanced
these applications. This seems under the current technical possibilities, turbulence models. AIAA J 1988;26(11):1299–310.
from the author’s point of view, more promising than trying to achieve [4] Wilcox David C. Formulation of the 𝑘-𝜔 turbulence model revisited. AIAA J
general improvements. 2008;46(11):2823–38.
As a further way to reduce the numerical complexity of the method, [5] Menter Florian R. Two-equation eddy-viscosity turbulence models for engineering
applications. AIAA J 1994;32:1598–605.
it was discussed to possibly summarize data and to solve a projected
[6] Menter Florian R, Kuntz M, Langtry R. Ten years of industrial experience with
problem. Therefore, future work may focus on the possibility of data the SST turbulence model. Turbul Heat Mass Transf 2003;4.
reduction (i.e., instead of using distributed measurements one assumes [7] Menter Florian R. Review of the shear-stress transport turbulence model
that only surface data are available). Theoretically this is also interest- experience from an industrial perspective. Int J Comput Fluid Dyn
ing in the sense to clarify how much data is required to determine the 2009;23(4):305–16.
[8] Rodi W, Spalding DB. A two-parameter model of turbulence, and its application
considered parameter, for example 𝜇𝑡 or 𝛽, uniquely. Furthermore, to to free jets. Wärme Stoffübertragung 1970;3(2):85–95.
improve results a stochastic modeling of the noise level can be included, [9] Jones WP, Launder BE. The prediction of laminarization with a two-equation
requiring often a deep understanding of the underlying problem as well model of turbulence. Int J Heat Mass Transfer 1972;15(2):301–14.
as the generation of measurements. Then, one may also replace (19) by [10] Bredberg Jonas. On two-equation eddy-viscosity models. Internal report, Göte-
borg, Sweden: Department of Thermo and Fluid Dynamics, Chalmers University
some other, possibly better suited stopping criterion.
of Technology; 2001.
Furthermore, as discussed in Section 6, when using only one scalar [11] Kamenetskii Dmitry S, Bussoletti John E, Hilmes Craig L, Venkatakrish-
function 𝜇𝑡 or the modification of a linear turbulence model, which nan Venkat, Wigton Laurence B, Johnson Forrester T. Numerical evidence of
ultimately only affects this turbulent viscosity, we gave arguments multiple solutions for the Reynolds-averaged Navier-Stokes equations. AIAA J
that the possibilities to improve in general solutions of the RANS 2014;52(8).
[12] Hadamard Jacques. Jacques Hadamard lectures on Cauchy’s problem in linear
equations appear limited. As shown in (32), one such solution, just
partial differential equations. New York: Dover publications; 1952.
like any turbulence model, only represents a certain approximation [13] Parish Eric J, Duraisamy Karthik. A paradigm for data-driven predictive modeling
to a possibly non-unique minimizer. Major improvements may only using field inversion and machine learning. J Comput Phys 2016;305:758–74.
be expected when one allows for more degrees of freedom such as in [14] Jäckel Florian. Sensitivity analysis of discrepancy terms introduced in turbulence
models using field inversion. In: Dillmann Andreas, Heller Gerd, Krämer Ewald,
Reynolds stress models. Therefore, it is certainly an interesting question
Wagner Claus, editors. 22. STAB/DGLR symposium 2020. Notes on numerical
for future work, whether the components of the Reynolds stress tensor fluid mechanics and multidisciplinary design, vol. 151, Springer Nature; 2021, p.
can directly be reconstructed, instead of the turbulent viscosity 𝜇𝑡 or a 625–34, Hardcover ISBN 978-3-030-79560-3 Softcover ISBN 978-3-030-79563-4
parameter within a given turbulence model. eBook ISBN 978-3-030-79561-0 Series ISSN 1612-2909 Series E-ISSN 1860-0824.
A further important issue for future work is the replacement of the [15] Duraisamy Karthik. Perspectives on machine learning-augmented reynolds-
averaged and large eddy simulation models of turbulence. Phys Rev Fluids
direct linear solvers. Instead one may consider to solve (23) by the
2021;6:050504.
conjugate gradient method for the normal equations (CGNE). Then a [16] Grabe Cornelia, Jäckel Florian, Khurana Parv, Dwight Richard P. Data-driven
direct evaluation of (25) and (26) can be avoided since only the matrix– augmentation of a RANS turbulence model for transonic flow prediction. Internat
vector products are required. These can be realized, for example, by the J Numer Methods Heat Fluid Flow 2023;33:1544–61.
[17] Jäckel Florian. A closed-form correction for the Spalart-Allmaras turbulence
application of a possibly preconditioned GMRES method.
model for separated flows. AIAA J 2023;61(6):2319–30.
[18] Colton David, Kreß Rainer. Inverse acoustic and electromagnetic scattering
CRediT authorship contribution statement theory. 2nd ed.. Berlin, Heidelberg, New York: Springer Verlag; 1997.
[19] Hohage Thorsten. On the numerical solution of a three-dimensional inverse
Stefan Langer: Data curation, Formal analysis, Funding acquisition, medium scattering problem. Inverse Problems 2001;17:1743–63.
Investigation, Methodology, Project administration, Resources, Soft- [20] Hohage Thorsten. Fast numerical solution of the electromagnetic medium
scattering problem and applications to the inverse problem. J Comput Phys
ware, Supervision, Validation, Visualization, Writing – original draft,
2006;214:224–38.
Writing – review & editing. [21] Langer Stefan. Preconditioned Newton methods for ill-posed problems (Ph.D.
thesis), University of Göttingen; 2007.
Declaration of competing interest [22] Langer S. Investigation of preconditioning techniques for the iteratively regular-
ized Gauss—Newton method for exponentially ill-posed problems. SIAM J Sci
The authors declare that they have no known competing finan- Comput 2010;32(5):2543–59.
[23] Hohage Thorsten, Langer Stefan. Acceleration techniques for regularized Newton
cial interests or personal relationships that could have appeared to
methods applied to electromagnetic inverse medium scattering problems. Inverse
influence the work reported in this paper. Problems 2010;26(7):074011.
[24] Wilcox David C. Turbulence modeling for CFD. DCW Industries, Incorporated;
Data availability 1994.
[25] Menter Florian R. Improved two-equation 𝑘-𝜔 turbulence models for aerodynamic
Data will be made available on request. flows. Technical report, NASA; 1992.
[26] Baldwin BS, Lomax H. Thin layer approximation and algebraic model for
Acknowledgments separated turbulent flows. AIAA J 1978;78–257.
[27] Engl Heinz W, Hanke Martin, Neubauer Andreas. Regularization of inverse
problems. Dordrecht, Boston, London: Kluwer Academic Publisher; 1996.
The author gratefully acknowledges the scientific support and HPC [28] Langer Stefan. Complexity analysis of the iteratively regularized Gauss-Newton
resources provided by the German Aerospace Center (DLR). The HPC method with inner CG-iteration. J Inverse Ill-posed Probl 2009;17:871–90.
system CARA is partially funded by ‘‘Saxon State Ministry for Economic [29] Langer Stefan. Application of a line implicit method to fully coupled sys-
Affairs, Labour and Transport’’ and ‘‘Federal Ministry for Economic tem of equations for turbulent flow problems. Int J Comput Fluid Dyn
2013;27(3):131–50.
Affairs and Climate Action’’. This work has been done within the [30] Langer Stefan. Agglomeration multigrid methods with implicit Runge-Kutta
framework of the DLR internal project ‘Machine Learning and Quantum smoothers applied to aerodynamic simulations on unstructured grids. J Comput
Computing - Digitalization of Aircraft Development 2.0’. Phys 2014;277:72–100.

9
S. Langer Computers and Fluids 284 (2024) 106438

[31] Langer Stefan. Preconditioned Newton methods to approximate solutions of [33] Langer Stefan, Hohage Thorsten. Convergence analysis of an inexact iteratively
the Reynolds averaged Navier-Stokes equations. Technical report, Institut für regularized Gauss-newton method under general source conditions. J Inverse and
Aerodynamik und Strömungstechnik; 2018. Ill-posed Prob 2007;15:311–27.
[32] Morozov VA. On the solution of functional equations by the method of
regularization. Sov Math Dokl 1966;7:414–7.

10