Computers and Fluids 284 (2024) 106438

Contents lists available at ScienceDirect

ComputersandFluids
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 Field Inversion and Machine Learning is an active field of research in Computational Fluid
Dynamics (CFD). This approach can be leveraged to obtain a closed-form correction for a given
turbulence model to improve the predictions. The fundamental approach is to insert a
Keywords: parameter into the system of RANS equations and determine it in a way such that, for
RANS equations
example, a given pressure distribution is better approximated compared to the one obtained
Iteratively regularized Gauss–
with the original set of equations. The goal of this article is twofold. Numerical arguments are
Newton method
presented that these kinds of problems can be severely ill-posed. In the second part, an
Field Inversion and parameter
identification approach is presented to directly reconstruct the turbulent viscosity field along with an
example. The Iteratively Regularized GaussNewton Method (IRGNM) is used for a realization.
ABSTRACT
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
In the Reynolds averaged Navier-Stokes (RANS) Navier–Stokes equations,
framework it is the goal of turbulence modeling to (c) it is possible to compute numerically an
determine the Reynolds stress ten- approximation to this steady-state solution with
acceptable effort based on the idea that
sor 𝜌𝑢′′𝑖 𝑢′′𝑗 . This can be done by introducing closures the use of a turbulence model allows a significant
entailing different levels of complexity. For standard CFD reduction in the degrees of freedom compared to a
applications, the Boussinesq approximation is generally Large Eddy Simulation (LES) or a Direct Numerical
invoked, which relates the Reynolds stresses directly to the Simulation (DNS).
trace-free rate of strain 𝑆𝑖𝑗. As a consequence the
In many situations the turbulence models serve their
turbulent viscosity, typically denoted by 𝜇𝑡 in the literature,
purpose, and there are many examples where, for
is incorporated into the Navier–Stokes equations.
example, drag and lift coefficients as well as other
Prominent examples of such an approach are the Spalart–
aerodynamic quantities agree well with measurements.
Allmaras turbulence model [1,2], the Wilcox 𝑘-𝜔 model
On the other hand, there are various examples in which
[3,4], the Menter SST 𝑘-𝜔 model [5–7], and 𝑘𝜀 models
solutions to the RANS equations are not unique or for
(see [8,9]). The definition of these models is typically not
which no steady-state solution can be computed, at least
unique, but there exist different variants. A rather
on a discrete level (see e.g. [11]). All of these are
complete overview of the number of possible models as
indications that boundary-value problems for the RANS
well as their relationships can be found in the report by
equations are generally not well-posed in the sense of
Bredberg [10].
Hadamard [12]. For this reason, it is still of great interest to
When applying a linear turbulence model, then
understand how the turbulence equations and the
formally, the RANS equations differ from the Navier–Stokes
resulting turbulent viscosity influence the RANS equations
equations by the additional unknown turbulent viscosity
and how existing turbulence models can be modified to
𝜇𝑡 representing an addition to the laminar viscosity 𝜇𝑙,
obtain a well-posed problem. Therefore, trying to improve
typically given by Sutherland’s law. For closure, the applied
and adapt these models for certain flows and in general is
turbulence model, given by an additional set of equations,
still an active field of research.
is constructed such that the resulting effective viscosity 𝜇eff
= 𝜇𝑙 +𝜇𝑡 may satisfy that

(a) a steady-state solution of the RANS equations exists

and it is unique,
S. Langer Computers and Fluids 284 (2024) 106438
In addition to the classic approach attempts are now An important and well investigated class of applications
being made to increasingly use machine learning methods for such inverse problems are inverse scattering problems,
to exploit the potential of existing turbulence models. To both acoustic scattering and electromagnetic scattering. In
adapt a given turbulence model for one or several flow inhomogeneous medium scattering problems one tries to
cases, a Field Inversion and Machine Learning (FIML) reconstruct a refractive index with compact support from
approach can be leveraged to obtain a closed-form farfield patterns [18–23]. These inverse problems arise
correction for the given turbulence model to improve the from linear partial differential equations, such as the
predictions [13–16]. For example, for the Spalart–Allmaras Helmholtz equation or the Maxwell equations, which are

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/ ).
turbulence model the main idea, which can be found in well understood in the sense that Fréchet differentiability
the literature [13,14], is, to manipulate a field parameter 𝛽 of the operators is established and the   underlying
weighting the production term 𝑃𝑟 in the source term 𝑄 of choice of the Hilbert spaces , and is well investigated.
the turbulence model, that is For the RANS equations, where the operator 𝑅
corresponds to a boundary-value problem for the
𝑄 = 𝛽𝑃𝑟 − 𝐷𝑒 + 𝐷𝑖,
governing equations of compress-
ible fluid flow, there does not exist a sustainable solution
where 𝛽 ≡ 1 yields the original unmodified model of theory. In addition, unlike the scattering problems the
Spalart and governing equations are nonlinear and of transport-type,
Allmaras, and 𝐷𝑒 and 𝐷𝑖 denote the destruction and making the inverse problems formulated above inherently
diffusion source term, respectively. Then, one determines more complex. Furthermore, though it is heuristically
𝛽 such that the corresponding solution of the RANS understandable that these problems are ill-posed, there is
equations fits, for example, some 𝑐𝑝-distribution ‘‘better’’ no proof of this in the strict mathematical sense.
with respect to some measure (norm). Such an approach Moreover, the question of whether the problems are
has been applied successfully for several examples [17]. mildly or severely ill-posed is also not sufficiently
To discuss the fundamental mathematical challenges of answered yet. Using an example, we will show that there is
this approach, in this article, we neglect knowledge of a significant evidence that these problems are possibly even
turbulence model, but attempt to directly reconstruct a severely ill-posed. One additional focus of this article is on
turbulent viscosity field 𝜇𝑡 from given data. To do so, we the application of the iteratively regularized Gauss–
consider the turbulent viscosity field as a parameter in the Newton method to reconstruct both the state 𝑊 and 𝜇𝑡.
RANS equations, which needs to be identified. Then, the In order to stabilize this method, it was necessary to
reconstruction of 𝜇𝑡, or 𝛽 for the problem above, are develop a suitable Sobolev norm for the presented
classical parameter identification problems well-known in example of a compressible flow around an airfoil. This
the field of Inverse Problems (IP). Here it is the goal to Sobolev norm is derived on the basis of a finite volume
identify two types of unknowns, the parameter 𝜇𝑡 and the discretization. Furthermore, by systematically adding noise
state 𝑊 linked via an equation formally written as to the given data, we will investigate how the method
behaves when we couple it with Morozov’s discrepancy
principle as a truncation criterion. This approach is
𝑅(𝑊 ;𝜇𝑡) = 𝑅𝜇𝑡 (𝑊 ) = 0, 𝑅 ∶  × →. (2) intended to address the question of the extent to which
the iterative method is suitable for approximating the
Considering the problem formulated above one needs to parameter to be identified sufficiently well for
aerodynamically relevant data, even under noisy data.
replace   𝜇𝑡 by 𝛽 such that 𝑅(𝑊 ;𝛽) = 𝑅𝛽(𝑊 ) = 0.
The outline of the article is as follows: The governing
Here , and are Hilbert spaces. Finding a solution of the equations of fluid flow as well as the turbulence models
state Eq. (2) is the direct problem. are discussed in Section 2. The forward and the inverse
problem are presented in Section 3. Though not a proof,
For many parameter identification problems it is but sustainable and understandable arguments are given
reasonable to
that the inverse problem for the considered example is
severely ill-posed and it is at least ill-conditioned on the
assume that 𝑅 is Fréchet differentiable and 𝜕𝑊
𝜕𝑅
(𝑊
dicrete level. The iteratively regularized Gauss–Newton
;𝜇𝑡) →  is a
method for solving the inverse problem is discussed in
continuous operator with continuous inverse. Under these Section 4. A numerical example of transonic flow over an
assumptions, due to the implicit function theorem, we can RAE2822 for the successful reconstruction of a turbulent
conclude that (2) has viscosity 𝜇𝑡 is presented in Section 5. The article concludes
() with a perspective on how the approach proposed can be
a unique solution 𝑊 = 𝑊 𝜇𝑡 and we can introduce a well- used to modify existing turbulence models so that they
may show more suitable results for a large number of test
defined  ( ) parameter-to-solution operator 𝛹 ∶ → , 𝜇𝑡 cases.

↦ 𝑊 𝜇𝑡 solving (2).
2
S. Langer Computers and Fluids 284 (2024) 106438
2. Governing equations and turbulence 2.2. Turbulence models

models 2.1. Governing equations of fluid When employing a turbulence model, the equations of
(3) are frequently supplemented by one ore more
flow differential or integral equations. For example, the two
equation 𝑘-𝜔 models have the form
The fluid dynamics equations of interest take the form
𝜕 (𝜌𝑘) (( ) )
𝜕𝜌 + ∇ ⋅ (𝜌𝑘𝐮) = ∇ ⋅ 𝜇𝑙 + 𝜎𝑘𝜇𝑡 ∇𝑘 + 𝜌𝑄𝑘,(𝑘,𝜔),
+ ∇ ⋅ (𝜌𝐮) = 0, (3a) 𝜕𝑡
𝜕𝑡 𝜕 (𝜌𝜔) (( ) )
𝜕(𝜌𝐮) + ∇ ⋅ (𝜌𝜔𝐮) = ∇ ⋅ 𝜇𝑙 + 𝜎𝜔𝜇𝑡 ∇𝜔 + 𝜌𝑄𝜔,(𝑘,𝜔).
+ ∇ ⋅ (𝜌𝐮 ⊗ 𝐮) + ∇𝑝 − ∇ ⋅ 𝜏 = 0, (3b) (8b)
𝜕𝑡 𝜕𝑡
𝜕(𝜌𝐸) where the unknowns 𝑘 and 𝜔 are the turbulence kinetic
+ ∇ ⋅ (𝜌𝐻𝐮) + ∇ ⋅ 𝐪 − ∇𝐮 ∶ 𝜏 = 0, (3c) energy per unit mass and the specific dissipation rate of
𝜕𝑡 the turbulence kinetic energy, respectively. The source
where 𝜌, 𝐮 = (𝑢1,…,𝑢𝑚), 𝑝, 𝐸 and 𝐻 are density, velocity, terms for the 𝑘 and 𝜔 equations include a production
pressure, specific total energy and total enthalpy, (𝑃𝑟), a destruction (𝐷𝑒) and a diffusion (𝐷𝑖) term, such
respectively. Considering a Newtonian fluid the viscous that
stress tensor is given by
𝑄𝑘,(𝑘,𝜔) ∶= 𝑃𝑟𝑘,(𝑘,𝜔) − 𝐷𝑒𝑘,(𝑘,𝜔),𝑄𝜔,(𝑘,𝜔) ∶= 𝑃𝑟𝜔,(𝑘,𝜔) − 𝐷𝑒𝜔,(𝑘,𝜔) +
𝐷𝑖𝜔,(𝑘,𝜔).
]
𝜏 = 𝜇eff𝐈 , The symbols 𝜎𝑘 and 𝜎𝜔 are either fixed constants or
functions describing the transition between a 𝑘-𝜔 and a
where 𝜇eff ∶= 𝜇𝑙 +𝜇𝑡 is an effective viscosity determined by 𝑘-𝜀 model. For details we refer to [3–6,24,25]. The
additional variables required for the turbulence model are
the laminar≥ viscosity and an eddy or turbulent viscosity then used to determine the turbulent viscosity 𝜇𝑡 = 𝜌𝑘∕𝜔.
𝜇𝑡 0 defined by some turbulence model. The laminar An example of a one-equation turbulence model is given
by the one of Spalart and Allmaras,
viscosity is given by Sutherland’s law
𝜕𝜈̃ 1 2
( 𝑇 )3∕2 𝑇∞ + 𝑇̄ 𝜌∞𝑢∞𝐿 +∇⋅ (𝜈̃𝐮) = [∇ ⋅ ((𝜈𝑙 + )
𝜇𝑙 ∶= 𝜇𝑙,∞ ̄, 𝜇𝑙,∞ ∶= 𝑅𝑒 , 𝑃𝑟 − 𝑇 + 𝐷𝑒, (9)
(5) 𝜈̃)∇𝜈̃ + 𝑐𝑏2 (∇𝜈̃) +
𝑇∞ 𝑇+𝑇 𝜕𝑡 𝜎
which represents an equation for the auxiliary variable 𝜈̃,
whereby 𝜌∞ > 0 and 𝑢∞ > 0 denote some constant
related to the turbulent viscosity 𝜇𝑡 = 𝜌𝑓𝑣1𝜈̃. In summary,
reference density and velocity, 𝐿 > 0 is some constant
formally, these linear turbulence models represented by
reference length scale, 𝑅𝑒 > 0 is the corresponding
differential or integral equations reveal additional
Reynolds number, and 𝑇̄ ∶= 110.4K is Sutherland’s
quantities in the considered fluid flow equations. The
constant. Furthermore, to determine 𝑝 the fluid is
occurring variables extend the degrees of freedom given
presumed to act as an ideal gas, that is 𝑝 = 𝜌ℜ𝑇, where ℜ
by the conservative variables 𝑊 by further unknown
is the universal gas constant. The conductive heat flux is
quantities 𝑊𝑡 ∶ 𝐷 × [0,𝑇) →R𝑁𝑡 .
𝜅
supposedly described by the Fourier law 𝐪 = − eff ∇𝑇,
𝜅 𝜅 𝜅 Here 𝑁𝑡 ∈ N depends on the turbulence model. For the
where eff ∶= 𝑙 + 𝑡 is the sum of the laminar conductivity
models considered we have
𝑐 𝑝𝜇𝑙 𝛾
𝜅𝑙 (𝑊 ) ∶= , 𝑐𝑝 ∶= ℜ , (6) 𝑃𝑟𝑙 𝛾−1 𝑁𝑡 = 1 for the Spalart–Allmaras model, 𝑊𝑡 = 𝜈̃𝑡

withe the laminar Prandtl number given by 𝑃𝑟𝑙 ∶= 0.72, 𝛾 𝑁𝑡 = 2 for the Wilcox 𝑘-𝜔 model and the SST model,
is the gas dependent ratio of specific heats, which is given 𝑊𝑡 = (𝑘,𝜔).
by 1.4 for air, and
Note, for example, when using an algebraic model, such as
𝜇𝑡 a BaldwinLomax model [26], 𝑁𝑡 = 0 and no additional
𝜅𝑡 ∶= 𝑐𝑝 , 𝑃𝑟𝑡 ∶= 0.92. (7) unknowns need to be considered.
𝑃𝑟𝑡
Throughout the rest of the article the vector of conserved
2.3. Boundary-value problem
variables for an open and simply-connected domain 𝐷 is
summarized as
For a complete boundary-value problem, the Eqs. (3)
𝑊 ∶ 𝐷 × [0,𝑇) →R𝑚+2,
and (8) or (9) need to be supplemented by boundary
𝑊 (𝑥,𝑡) ∶= (𝜌(𝑥,𝑡),𝜌(𝑥,𝑡)𝐮(𝑥,𝑡),𝜌(𝑥,𝑡)𝐸(𝑥,𝑡))𝑇 . conditions. Here, for the no-slip boundary we postulate an
adiabatic wall boundary condition
𝜕𝑇
3
S. Langer Computers and Fluids 284 (2024) 106438
𝐮=0 and =0 𝐹(𝜇𝑡) = 𝑊 (𝜇𝑡) − 𝑊 †.
𝜕𝑛
†
That is, a complete field state 𝑊 is given, and the
and
†
corresponding turbulent viscosity 𝜇𝑡 is the field parameter
6𝜇
𝑘 = 0, lim 𝜔(𝑥 + ℎ𝑛(𝑥))ℎ2 = 𝑙 to be reconstructed. Then, the inverse problem is to find a
or 𝜈̃𝑡 = 0. ℎ→0+ root of Eq. (11). When the given data are a surface
𝛽𝜌 distribution, then (11) is replaced by
Here 𝑛 = 𝑛(𝑥) denotes the outer unit normal vector. For
( ) ( †),
the outer boundary we consider characteristic farfield
𝐹(𝜇𝑡) = 𝑃 𝑊 (𝜇𝑡) − 𝑃 𝑊
boundary conditions.
and 𝑃 denotes the trace operator corresponding to the
given surface data. For example, if a 𝑐𝑝- or a 𝑐𝑓 -
3. An inverse view on data driven turbulence modeling
distribution are the given data, then (12) reads as

3.1. Forward problem ( ) †

or 𝐹(𝜇𝑡) = 𝑃𝑐𝑓 (𝑊 (𝜇𝑡)) − 𝑐𝑓†
𝐹(𝜇𝑡) = 𝑃𝑐𝑝 𝑊 (𝜇𝑡) − 𝑐𝑝
Formally, the forward problem is formulated as follows:
Even both, as a system of equations, can be considered,
We consider 𝑅 to be the residual operator corresponding
( ( )) ( †)
to the set of equations for a boundary-value problem for
the RANS equations, 𝜇𝑡 is the field parameter of interest 𝑃𝑐𝑝 (𝑊 (𝜇𝑡)) 𝑐𝑝
and 𝑊 a state vector. Then these are linked via the 𝐹(𝜇𝑡) = 𝑃𝑐𝑓 − † .
equation 𝑊 (𝜇𝑡) 𝑐𝑓
𝑅(𝑊 ;𝜇𝑡) = 𝑅𝜇𝑡 (𝑊 ) = 0.
Note that the state 𝑊 = 𝑊 (𝜇𝑡) satisfies (2), so Eq. (11)

The solution of this state Eq. (10) for a given fixed involves the solution of the RANS equations given the
parameter 𝜇𝑡 can be interpreted as the direct problem. parameter 𝜇𝑡, that is
3.2. Data for the inverse problem
𝐹(𝜇𝑡) = 𝑅−1𝜇𝑡 (0) − 𝑊 †.
For the inverse problem, it is our interest to determine
the field parameter 𝜇𝑡 with respect to observed data Here the symbol 𝑅−1𝜇𝑡 (0) means in practice to approximate
related to a given state 𝑊 = 𝑊 (𝜇𝑡). This observation a solution of the discretized RANS equations for the given
operator could be, for example, either the identity turbulent viscosity 𝜇𝑡. Considering the FIML approach (1),
(distributed measurement), or a trace operator to
then we obtain in a completely analogous way the
boundary values of the solution (boundary
equation 𝐹(𝛽) = 𝑅−1𝛽 (0) − 𝑊 †.
measurements). In CFD, these are often the 𝑐𝑝-distribution
and/or the skin friction coefficient 𝑐𝑓 on the surface of the
geometry. Such viewpoint is, for example, discussed in 3.4. Ill-posed character of the field reconstruction
[14,17].
First of all, let us clarify what kind of classification of
Typically the data, that is for example these surface
problem and degree of ill-posedness we are dealing with.
distributions,
Since a profound theory is missing, we at least want to
(a) come from measurement, formulate heuristic arguments. Using the implicit function
(b) or are artificially generated for test purposes. theorem, we compute the derivative

In both cases, the data is typically noisy with an unknown 𝑑𝑅−1(0) ( ) 𝐹′[𝜇𝑡] = 𝜇𝑡
𝑑 𝜕
noise level 𝛿, for example =𝑡,𝑊 ) −1 = − .
𝜕𝑅(𝜇
𝑑 𝑡 𝑑 𝑡 𝜕 (14) 𝜕𝜇𝑡
𝜕 (�
𝑡, )
‖𝑐𝑝 𝛿
𝛿,
Both required derivatives and 𝜕𝑅 can be determined
where the superscript 𝛿 indicates the noisy and † the exact explicitly. 𝜕𝑊 𝜕𝜇𝑡
data, respectively. In this article, to have a sufficient
To understand the character of problem (13) we recall the
amount of data, we consider a distributed measurement
following classification of ill-posed problems [21,27,28]. An
artificially generated for test purposes. An inverse crime is
ill-posed problem is called
Fig. 1. Used mesh and decay behavior of the singular values of 𝐹′[𝜇𝑡] = 𝑑𝑊 .
omitted by generating the data from a finer mesh. (a) mildly ill-posed, if the singular values 𝜎𝑗 of 𝐹′[𝜇𝑡]
decay at a polynomial rate, i.e.
3.3. Inverse problem: Parameter-to-solution operator

𝑐𝑚𝑗−𝛼 ≤ 𝜎𝑗 ≤ 𝐶𝑚𝑗−𝛼, 𝛼,𝑐𝑚,𝐶𝑚 > 0

To formulate the inverse problem as an operator
equation, we use  the parameter-to-solution operator (b) severely ill-posed, if the singular values 𝜎𝑗 of 𝐹′[𝜇𝑡]
decay at an exponential rate, i.e.
𝐹∶→,

4
S. Langer Computers and Fluids 284 (2024) 106438
𝑐𝑒𝑒−𝑐𝑠𝑗𝛽 ≤ 𝜎𝑗 ≤ 𝐶𝑒𝑒−𝑐𝑠𝑗𝛽 𝛽,𝑐𝑠,𝑐𝑒,𝐶𝑒 > 0 𝑑𝜇𝑡

The degree of ill-posedness depends on the differential To approximately solve the operator Eq. (13) we
operator of the underlying differential equation. For the substitute 𝐹 by its linear approximation. Given an initial
example presented in Sec- guess 𝜇𝑡,0, in each Newton step the least squares problem
)
tion 5, Eq. (14) is used to compute the singular values( ′[𝜇𝑡] ‖𝐹′[𝜇𝑡,𝑛]ℎ𝑛 + 𝐹(𝜇𝑡,𝑛)‖2 = minℎ𝑛 !
𝑗=1,2,… for an iterate in a neigh-𝜎𝑗. Fig. 1 (right)

shows the decay behavior of 𝜎𝑗 𝐹 needs to be solved and the update is computed by 𝜇𝑡,𝑛+1 =
borhood of the solution. The decay at an exponential rate 𝜇𝑡,𝑛 + ℎ𝑛. This approach is well-known as Gauss–Newton
of these singular values is obvious. This observation serves
as an indicator that solving (11) is, at least for the example method. In the case where (13) is ill-posed, it is important
considered, a severely illposed problem and that the to consider the situation where the given data 𝑊 †
is
stability condition for a well-posed problem is violated.
replaced by noisy data 𝑊 𝛿 satisfying ‖𝑊 † − 𝑊 𝛿‖ ≤ 𝛿.
This is of course not a mathematical proof in the strict
sense, but in practice it means that we have to solve In this case a straightforward implementation of the
systems of linear equations with very large condition Gauss–Newton method usually fails and does not lead to a
numbers. Hence, regularization needs to be employed. good reconstruction of the solution after several Newton
steps. One reason for the failure of the Gauss–Newton
4. Discretization and solution algorithm method in this situation is the ill-posedness of the least
squares problem (15), which is inherited from the original
4.1. Discretization strategy operator Eq. (13). Thus, to perform one Newton step some
sort of regularization has to be employed when solving
For a detailed description of the discretization strategy
(15). In this article we consider a regularized Gauss–
and the solution algorithm for the RANS equations we
Newton method, where instead of (15), the regularized
refer to [29–31]. It is based on a finite-volume formulation
least squares problem
where a median dual grid forms the control volumes with
the unknowns at vertices of the primary grid. The dual grid
‖𝐹′[𝜇𝑡,𝑛𝛿 ]ℎ𝑛 + 𝐹(𝜇𝑡,𝑛𝛿 )‖2 + 𝛾𝑛‖ℎ𝑛‖2 = minℎ𝑛 !
is created in a preprocessing step. Throughout this paper,
the set of points in the dual grid where the unknowns are
located is denoted by 𝐺 ∶= {1,…,𝑁} and the unknowns are is solved in each Newton step. For the iteratively
given by𝑊1,…,𝑊𝑁. The neighbors of some point 𝑖 ∈ 𝐺 regularized Gauss– Newton method we have
are described by (𝑖), and the corresponding face between

the cells 𝑖 and 𝑗 is denoted by 𝑒𝑖𝑗.

The set of Eqs. (3) are integrated and the resulting set ‖𝐹′[𝜇𝑡,𝑛𝛿 ]ℎ𝑛 + 𝐹(𝜇𝑡,𝑛𝛿 )‖2 + 𝛾𝑛‖ℎ𝑛 + 𝜇𝑡,𝑛𝛿 − 𝜇𝑡,0‖2 =
of integral equations are discretized for each control minℎ𝑛 !
volume. The convective part of the equations is discretized
using a matrix-valued artificial viscosity scheme, which is Here
similar and close to some kind of Roe-scheme. On a
sufficiently regular mesh this discretization can be shown
to be of second order. A pressure switch is included into 𝛾𝑛 = 𝛾0𝛾−𝑛 > 0, 0 ≤ 𝛾 ≤ 1, 𝛾0 > 0
the artificial viscosity part reducing the method to first
denotes a regularization parameter. This iterative
order in a neighborhood of a shock. Gradients are
regularization method can be interpreted as a common
reconstructed using the method of Green-Gauss.
Newton method where in each step Tikhonov
regularization is applied to the linearized equation. The
4.2. An iterative regularized Gauss–Newton method
problems (16) and (17) are well-posed, in particular there
The derivation and realization of the regularized exists a uniquely defined minimizer ℎ†. In the numerical
Gauss–Newton method follows [19,21–23]. experiments it turned out that solving (17) had better

5
S. Langer Computers and Fluids 284 (2024) 106438
stability compared to (16), in particular the initial = svol(𝑒𝑖𝑗) 𝑔𝑖 + 𝑔𝑗 ,
regularization parameter 𝛾0 could be significantly reduced. 𝜕𝑥𝑘 𝐷𝑖 vol(𝐷𝑖) 𝑗∈(𝑖) 2
Due to the severe ill-posedness and resulting instability
of the problem, we still need to include into our process a and hence
stopping criterion. Here, for simplicity, we use the
discrepancy principle of Morozov [32], that is we stop the 𝜕
iteration at the first index 𝑀 satisfying ( 𝜕𝑔 )GG 1 ⎧⎪∑𝑗∈(𝑖) svol(𝑒𝑖𝑗)𝑛𝑘,𝑒𝑖𝑗 ,
=𝑖, (20)
=
‖ 𝜕𝑔𝑚 𝜕𝑥𝑘 𝐷𝑖 2vol(𝐷𝑖) ⎨⎪⎩ svol0(𝑒, 𝑖𝑚),𝑚 ∈else.
‖‖‖‖𝐹 (𝜇𝑡,𝑀𝛿 )‖‖‖‖ = ‖‖‖‖‖𝑅−1𝜇𝑡,𝑀𝛿(0) − 𝑊 𝛿‖‖ ‖‖ ≤ 𝜏𝛿, 𝜏 >
(𝑖),
1. (19)

Typically, a violation of such stopping criterion may lead to Eqs. (20) represent the sparse matrices
oscillations and instabilities in the solution. This is also
𝑑 (( 𝜕𝑔 )GG) ( 𝜕 ( 𝜕𝑔 )GG)
confirmed for the example discussed in Section 5 and
𝐵𝑥𝑘 ∶= =
shown in Figs. 5 and 6.
𝑘 = 1,2,3,
4.3. Choice of Hilbert scales
𝑑𝑔 𝜕𝑥𝑘 𝐷𝑖 𝜕𝑔𝑚 𝜕𝑥𝑘 𝐷𝑖
𝑖=1,…,𝑁 𝑖,𝑚=1,…,𝑁
So far, nothing has been said about the evaluation of
satisfying
the norms in the image space and in the pre-image space, (
)
that is ‖ . ‖ and ‖ . ‖.
(( 𝜕𝑔 )GG) ⎛𝑔1 ⎞ ⎛⎜ 𝜕𝑥𝜕𝑔𝑘GG𝐷1 ⎞⎟
Such distinction is important and in general the
expectation is that it has an effect on both the accuracy of 𝐵 𝑥𝑘 𝑔 = 𝑑 ⎜⋮⎟=⎜ ⋮ ⎟,
the approximate solution and the success of the algorithm. = 1,2,3.
It is well-known from other applications that for a
successful implementation of iterative regularization ⎜ ⎟ ⎜⎜ (
𝑑𝑔 𝜕𝑥𝑘 𝐷𝑖 𝑖=1,…,𝑁 ⎝𝑔𝑁 ⎠ ⎝ 𝜕𝑥𝜕𝑔𝑘 𝐷GG𝑁 )
methods that it is rather important to work in problem-
suited function spaces. If available, these should of course ⎟⎟
⎠
be based on a well-founded theory. For example, for
scattering problems [19,23] Fréchet differentiability of the
occurring operators is proven with respect to the Application of 𝐵𝑥𝑘 introduces a discrete possibility to
‖.‖𝐶1−norm. Then, for a numerical implementation one evaluate the norm of the derivative of 𝑔, that is
goes over to a suitable Sobolev norm, since one needs a
Hilbert space structure.
In our implementation we observed in general unstable or
‖ 𝑑𝑔 ‖ ‖ ‖ 𝜕𝑔 ‖
less ‖ ‖𝑑𝑥 2
𝑘 ‖ ‖‖ 𝐿2(𝐷) ≈ ‖‖‖‖ ((𝜕𝑥 𝑘 )GG𝐷𝑖 )𝑖=1,…,𝑁 ‖‖‖‖ ‖𝐿22(𝐷)
=⟨𝐵𝑥𝑘 𝑔,𝐵𝑥𝑘 𝑔⟩𝐿2(𝐷) ,
accurate solution for the choice  = 𝐿2(𝐷) and the ‖
algorithm failed.
And as a consequence we obtain
Unfortunately, there is no reliable theory for the RANS
equations. Since gradients and derivatives play an
essential role, we have endeavored to incorporate them
into the implementation of the algorithm. To do this, we ‖2 ( 3 ( 𝜕𝑔 )2)𝑑𝑥⟨𝐵𝑥𝑗 𝑔,𝐵𝑥𝑗 𝑔⟩ 2(𝐷) . grad 𝑔‖𝐿𝐿
implement an 𝐻1(𝐷)-norm in the sense of a finite volume
method, that is a discrete realization of
Furthermore, to penalize for the influence of the strongly
varying mesh cell sizes, which exist due to the strong
‖𝑔‖2𝐻1(𝐷) = ‖𝑔‖2𝐿2(𝐷) +‖grad 𝑔‖2𝐿2(𝐷) = ∫ |𝑔(𝑥)|2 𝑑𝑥+∫ |grad anisotropies needed in meshes
𝑔(𝑥)|2 𝑑𝑥. for simulating turbulent flows, we finally define the
𝐷 𝐷 discrete norm for the pre-image space by,
In a finite-volume context the function 𝑔 is approximated
by
2 = ∑𝑁 ( 1 ) 3 ⟨𝐵𝑥𝑗 ℎ,𝐵𝑥𝑗 ℎ⟩

‖ℎ‖ 2𝐿
∑𝑁 {1, 𝑥 ∈ 𝐷𝑗,
vol(𝐷
𝑔𝑗𝟏𝐷𝑗 (𝑥), where 𝟏𝐷𝑗 (𝑥) = 0, 𝑒𝑙𝑠𝑒.
𝑖=1 𝑖) 𝑗=1
𝑗=1
∑3 ⟨ ⟩
Then, in our implementation the derivative is
approximated using a = ⟨𝐵ℎ,𝐵ℎ⟩𝐿2(𝐷) + 𝐵𝑥𝑗 ℎ,𝐵𝑥𝑗 ℎ 2(𝐷) ,
𝐿 𝑗=1
Green-Gauss method, that is for 𝑘 = 1,2,3 we have
where
( 𝜕𝑔 )GG 1 ∑ 𝑛𝑘,𝑒𝑖𝑗 ( )
⎛√ 1 ( )⎞

6
S. Langer Computers and Fluids 284 (2024) 106438
𝐵 ∶= diag⎜⎜⎝ vol(𝐷𝑖)2 + vol 𝐷𝑖 ⎟⎟⎠ 𝐹

and we have used the representation of the function ℎ in ′ [𝝁𝛿𝑡,𝑛]𝑇 = (𝜕𝜇𝜕𝑅𝑡 )𝑇 (𝜕𝑅(𝜇𝜕𝑊𝑡,𝑛𝛿 ,𝑊 ))−𝑇 .
the finitevolume context, ∑𝑁 ℎ(𝑥) = ℎ𝑗𝟏𝐷𝑗 (𝑥).
𝑗=1 𝐹
use the formulations (16) and (17), respectively. Using  =
This means that we need to determine the inverse of the
𝐿2(𝐷) To implement the Levenberg–Marquardt algorithm
matrix
or the IRGNM we and  = 𝐻1(𝐷), the least squares
𝜕𝑅(𝜇𝑡,𝑛𝛿 ,𝑊 ) ∈ R5𝑁×5𝑁,
problems are given by

[ ] ( ) 𝜕𝑊
⎛𝐹′ 𝝁𝛿𝑡,𝑛 ⎞ ⎛−𝐹 𝝁𝛿𝑡,𝑛 ⎞ which is rather time consuming. Using

= (𝜕𝑅(𝜇𝛿 ,𝑊 ))−𝑇 ⎛(𝜕𝑅(𝜇𝑡,𝑛𝛿 ,𝑊 ))−1⎞𝑇

⎜ √𝛾𝑛𝐵 ⎟⎟ ⎜⎜ √𝛾𝑛𝐵𝑏𝑛 ⎟⎟ ⎜ ⎜√𝛾𝑛𝐵𝑥1 ⎟⎟ℎ𝑛 ⎜⎜√𝛾𝑛𝐵𝑥1𝑏𝑏⎟⎟.
(22) =⎜ ⎟
⎜ (27) 𝑡,𝑛
𝜕𝑊 ⎜⎝ 𝜕𝑊 ⎟⎠

⎜⎜√𝛾𝑛𝐵𝑥2 ⎟⎟ ⎜⎜√𝛾𝑛𝐵𝑥2𝑏𝑛⎟⎟
only the inverse of the original matrix needs to be
determined. When applying the 𝑄𝑈-decomposition only
⎜⎝√𝛾𝑛𝐵𝑥3 ⎟⎠ ⎜⎝√𝛾𝑛𝐵𝑥3𝑏𝑛⎟⎠
(25) is required. Since the de-
Choosing 𝑏𝑛 = 0 corresponds to (16), which is the termination of this inverse represents the main complexity
Levenberg–Marquardt algorithm, 𝑏𝑛 = 𝜇𝑡,0 − 𝜇𝑡,𝑛𝛿 is the of the algorithm, both application of the normal Eq. (24) or
solving the least-squares problem (23) directly using a QU-
IRGNM, which is preferred here and used for the
decomposition were comparable in required time.
numerical example in Section 5. To reduce the complexity In summary, the main complexity of the proposed
of the algorithm in our implementation we have replaced implementation of algorithm (15) is given by:
Eq. (22) by
[ ] ( ) (a) To determine 𝐹(𝜇𝑡,𝑛𝛿 ) in (23) we need to solve the
RANS equations for the given iterate 𝜇𝑡,𝑛𝛿 .
⎛⎜⎝⎜ 𝐹′ 𝝁𝛿𝑡,𝑛 ⎞⎟ℎ𝑛 = ⎛⎜√−𝛾𝐹𝑛𝐵𝝁total𝛿𝑡,𝑛𝑏𝑛⎞⎟⎟⎠. 𝜕𝑅(𝜇𝑡,𝑛𝛿 ,𝑊 )
(b) For each step we need to determine the inverse of
.
(23) √𝛾𝑛𝐵total⎠⎟ ⎜⎝ 𝜕𝑊
(c) For each step we need to solve the linear least-
where 𝐵total = 𝐵 + 𝐵𝑥1 + 𝐵𝑥2 + 𝐵𝑥3. This corresponds to squares problem (23).

exchanging the norm in (21) for an expression of the form

5. Numerical example
2 = ⟨𝐵totalℎ,𝐵totalℎ⟩.
‖ℎ‖ ,mod As an example we consider turbulent flow over the RAE
2822 airfoil at 𝑀 = 0.73, angle of attack 𝛼 = 2.79◦ and 𝑅𝑒 =
4.4. Linear solvers 6.5 ⋅ 106. This case features a shock on the upper part of
the airfoil. At this early stage of these investigations the
To solve the linear least-squares problems two
computations have been done on a rather coarse mesh of
different approaches are used. The first is to reformulate
size 160 × 32 shown in Fig. 1 (left). As initial guess we use
(23) by the normal equation
the turbulent viscosity 𝜇𝑡,0 computed by the same
geometry and the flow conditions 𝑀 = 0.75, angle of
( ( total)𝑇 𝐵total + 𝐹′ [𝝁𝛿𝑡,𝑛]𝑇 𝐹′ [𝝁𝛿𝑡,𝑛])ℎ𝑛 attack 𝛼 = 2.81◦ and 𝑅𝑒 = 6.2 ⋅ 106. The regularization
𝛾𝑛 𝐵
parameter (18) is chosen by 𝛾 .
To show the applicability of the method we first have
′ [𝝁𝛿𝑡,𝑛]𝑇 𝐹 (𝝁𝛿𝑡,𝑛) + 𝛾𝑛 (𝐵total)𝑇 𝐵total𝑏𝑛. (24) = −𝐹 used exact data. Fig. 2 shows the convergence histories, on
the left of
Then this linear system is directly solved using an 𝐿𝑈-
decomposition. ‖𝐹(𝜇𝑡,𝑛𝛿 )‖,𝑛=0,1,2,… and on the right the corresponding
The second method we used is to solve (22) using an aerodynamic coefficients of lift and drag. The plotted
𝑄𝑈decomposition. Of course, both methods are residual is normalized and given by
equivalent in the sense that they determine the unique
minimizer of (23). Solving (24) involves the evaluations of ‖

′ [𝝁𝛿𝑡,𝑛] = (𝜕𝑅(𝜇𝜕𝑊𝑡,𝑛𝛿 ,𝑊 ))−1 𝜕𝜇𝜕𝑅𝑡 , (25) log⎛⎜⎜⎜⎝‖𝑊 (𝐹𝜇𝑡,𝛿(0𝜇)𝑡,𝑛𝛿 −)‖𝑊 †‖⎞⎟⎟⎠⎟ = log⎜⎛⎜‖𝑊
((𝜇𝑡,𝑛𝛿𝛿 )) −− 𝑊𝑊 𝛿†‖‖⎟⎞⎟⎟⎠.
7
 satisfied simultaneously, that is (28) simplifies to

S. Langer Computers and Fluids 284 (2024) 106438

⎛ 𝑃1 (𝑅1𝜇𝑡 )−1 (0))⎞⎟ ⎛𝑔1⎞
⎜⎝‖𝑊 𝜇𝑡,0
⎜

The convergence follows an expected behavior of the

(29)
IRGNM [33]. This example shows that it is in principle ⎜ ( (
possible to directly reconstruct the parameter 𝜇𝑡 given a 𝐹(𝜇𝑡) = ⎜ ⎜ 𝑇 𝑇𝜇⋮𝑡 )−1 (0))⎟⎟⎟ − ⎜⎜⎝𝑔⋮𝑇 ⎟⎠⎟,
sufficient number of prescribed data. Additionally, Fig. 3
demonstrates that for these exact data the surface ⎜⎝ 𝑃 𝑅 ⎟⎠
pressure (left) and skin-friction (right) distributions are
perfectly recovered. or
In a second computation we have added about 10% (
noise to the data. The final iterate was selected by (19)
⎛ 𝑃1 (𝑅1𝛽)−1 (0))⎞⎟ ⎛𝑔1⎞
with 𝜏 = 1.5. Fig. 5 shows the reconstructed pressure (left)
and skin-friction (right) as well as these distributions ⎜ 𝐹(𝛽) = ⎜ ⋮ ⎟⎟ −
perturbed by noise. As expected, the regularization has an ⎜ ⋮ ⎟.
effective smoothing effect while the shock position is still
perfectly resolved. Moreover, Figs. 5 and 6 also show the ⎜ (30)
pressure and skin-friction distributions when violating the
⎜⎜⎝( ( 𝑇 )−1 (0))⎟⎟⎠ ⎝⎜𝑔𝑇 ⎠⎟
stopping criterion (19), that is, we have plotted the 15th
and 20th iterate. One clearly notices the influence of the 𝑃𝑇 𝑅𝛽
noisy data, the pressure distribution is perturbed and the
skin-friction distribution starts to oscillate. In particular, as For both (29) and (30) we look for a scalar function 𝜇𝑡 or 𝛽,
expected, the skin-friction in a neighborhood of the shock such that 𝑇 equations are satisfied simultaneously, at least
is significantly amplified. As expected from theory, for the in the sense of minimizing the least-squares problem
15th iterate the regularization parameter 𝛾𝑛 is too small to ‖𝐹(𝜇𝑡)‖ → min! or ‖𝐹(𝛽)‖ → min!, that is
balance the data noise error. After 20 iterations the ∑‖‖( ( )−1 )
obtained solution is completely worthless. The
corresponding convergence history for noisy data is shown
min𝜇𝑡 𝑘𝑇 ‖ 𝑃𝑘 𝑅𝑘𝜇𝑡 (0) − 𝑔𝑘‖‖‖‖‖ ormin𝛽 ∑𝑘𝑇=1 ‖‖‖‖‖(𝑃𝑘
in Fig. 4. (𝑅𝑘𝛽)−1 (0)) − 𝑔𝑘‖‖‖‖‖.
‖
6. Machine learning of a turbulent viscosity
=
1

6.1. Turbulence modeling as an inverse problem ‖

This can also be interpreted the other way around. For
From a general point of view, one is interested in example, considering the original Spalart–Allmaras model
learning a turbulent viscosity or improving an existing (9), then we obtain
turbulence model for a large number of different test
∑‖‖( ( )−1 ) ‖
cases. For example, this could be either for the same
geometry and a number of different flow conditions or
even a change in the geometry can be considered. Then min𝛽 𝑘𝑇 ‖‖ 𝑃𝑘 𝑅𝑘𝛽 (0) − 𝑔𝑘‖‖‖‖≤ min𝛽≡1 ∑𝑘𝑇=1 ‖‖‖‖‖(𝑃𝑘 (𝑅
−1 (0)) − 𝑔𝑘‖‖‖‖‖
we can generalize the considered idea in this article in the
following way. Assume that we want to reconstruct (learn) =
the field parameters 𝜇1,…,𝜇𝑠, 𝑠 ∈ N and that 𝑇 sets of data 1

𝑔1,…,𝑔𝑇 are given. Similar as (12) we can write these as ‖

∑𝑇 ‖‖( (
‖

⎜⎜ ⎟ = 𝑘 ‖‖‖ 𝑃𝑘 𝑅𝑘𝜇𝑡=𝜇𝑡(SA model)

⎛⎜⎝⎜⎜(𝑃1 (𝑅1𝜇11,…⋮,𝜇𝑠𝑠)−1 (0))⎟⎞⎟⎟ − ⎛⎝𝑔𝑔⋮𝑇1⎞ ⎠
(0) − 𝑔𝑘‖‖‖‖.
(28)
=1

(31)

⎜( ( )⎟ In this sense, the original Spalart–Allmaras model is only

𝐹(𝜇1,…,𝜇𝑠) = )−1
one instantiation of a more general representation of the
model including one or even more parameters. On the
⎜ 𝑇 (0) ⎟⎠ other hand, it is also clear that in practice one will only
ever find one optimizer for the given data and conditions,
𝑃𝑇 𝑅𝜇 ,…,𝜇
and it is unclear whether this minimizer is unique and also
which represents in general a large-scale, possibly ill-posed applies to other data sets. The argument can be extended
nonlinear to other turbulence models. For example, we can simply
system of equations. For example, following either the estimate
FIML approach or the reconstruction of a turbulent
viscosity field, then one parameter either 𝛽 or 𝜇𝑡 needs to ∑𝑇 ‖‖‖( ( )−1 ) ‖
be determined such that 𝑇 equations must be

8
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.
min𝜇𝑡 ‖‖ 𝑃𝑘 𝑅𝑘𝜇𝑡 (0) − 𝑔𝑘‖‖‖‖
𝑘=1

‖𝑃 𝑅𝑘 −1 (0) −
≤ min⎧⎨⎪⎪∑𝑇𝑘=1 ‖‖ ( 𝑘( 𝜇𝑡=𝜇𝑡(SA model)) )
𝑔 ‖‖ ⎪
𝑘 ‖‖‖⎪ ⎫.

‖
‖

( ( )−1 ) ‖⎬

⎪⎪⎩∑𝑇𝑘=1 ‖‖‖‖‖ 𝑘𝑡 𝑡 (0)−

𝑔𝑘‖‖‖‖⎪⎪⎭
Fig. 4. Convergence history for the IRGNM using noisy data.
𝑃𝑘 𝑅𝜇 =𝜇 (𝑘-𝜔 model)

That is, among all possible turbulent viscosities 𝜇𝑡 that

minimize the initial minimization problem ((29) (left)), the
SA model and the 𝑘-𝜔 model represent only two possible
variants. Then, a given turbulence model is only one way that minimizing a parameter in a given turbulence model
of determining an approximation to this minimizer, which, is probably generally an easier task than directly
depending on the given test case, sometimes reflects the minimizing for the turbulent viscosity, since the basic
available data better and sometimes worse. Interpreting turbulence model serves already as some kind of a priori
the topic of turbulence modeling in this way, it knowledge.
corresponds mathematically to the inverse problem
In order to find a better minimizer, one obvious way is
formulated in Sections 3.2 and 3.3. Though not given a
proof, due to the nature of the problem it can be assumed to allow for a vector-valued function instead of the scalar
to be ill-posed, in general. Because of (32) and (31) we can function 𝜇𝑡. For example, a natural generalization, instead
also assume

9
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.

of using a linear turbulence model, is to determine the six 6.3. Generalization to a variety of fluid flow test cases

components (𝑟𝑖𝑗)1≤𝑖<𝑗≤3 of the Reynolds stress tensor.

So far the proposed methodology depends on the
Then we have not only one but six free parameters to
mesh in the sense that the specific parameter is
minimize (28), that is we can minimize
determined at the grid nodes. If we want to carry over the
‖ ‖
‖𝐹(𝑟11,𝑟12,𝑟13,𝑟21,𝑟23,𝑟33) ‖. determined parameters 𝜇1,…,𝜇𝑠 to other geometries, then
we have to get away from considering this as a field
parameter for
6.2. Computational complexity
a given mesh. Instead we need to find a suitable ansatz
function. That is we consider
Note that (28), (29) and (30) represent in general large-
scale, nonlinear ill-posed problems, which, in addition to 𝜇𝑗 = 𝜇𝑗(𝜔𝑠1,…𝜔𝑠𝑁 ), 𝑗 = 1,…,𝑠,
the algorithmic challenges of finding approximate
solutions, also contain many technical challenges. Even on and 𝜔𝑠1,…,𝜔𝑠𝑁 represent the coefficients in the ansatz
today’s High Performance Cluster infrastructure, it could function. For example, the approach presented here can
be difficult to provide the necessary storage space with a be adapted for this directly. Then the derivative needs to
naive approach, if 𝑇 is large. On the other hand, only if 𝑇 is be adjusted to the ansatz coefficients, that
large enough, that is sufficiently data is available, we may 𝜕𝜇
is replaced by
𝑡
expect a representative determined parameter which can
is 𝜕 𝜕 𝜕𝜇𝑡 𝜕𝑅
be applied to a large number of test cases.
If implementation and solving the problem does not = , 𝜇𝑡 = 𝜇𝑡(𝜔1,…,𝜔𝑑),
appear possible due to technical requirements, another 𝑑 ∈ N.
option is to examine, weight and summarize the data in 𝜕(𝜔1,…,𝜔𝑑) 𝜕𝜇𝑡 𝜕(𝜔1,…,𝜔𝑑)
advance. If we express this process mathematically by Hence, additionally the derivative of the ansatz function
another projection operator 𝑃, we can solve a projected with respect to its coefficients is required. Possible ansatz
equation instead of solving (28), (29) or (30), that is for functions which can be considered in future work are for
example ‖𝑃𝐹‖ → min𝜇1,…,𝜇𝑠 !. Such an approach can example convolutional neural networks.
possibly reduce the number of unknowns, such that
technical feasibility becomes possible.

10
S. Langer Computers and Fluids 284 (2024) 106438
7. Conclusion and future work may focus on the possibility of data reduction (i.e., instead
of using distributed measurements one assumes that only
In this article, it was shown that the relatively new surface data are available). Theoretically this is also
approach of modifying existing turbulence models by the interesting in the sense to clarify how much data is
introduction of field parameters can be interpreted as an required to determine the considered parameter, for
application of parameter identification problems studied example 𝜇𝑡 or 𝛽, uniquely. Furthermore, to improve results
for a long time in the field of Inverse Problems. For the a stochastic modeling of the noise level can be included,
examples shown, it was demonstrated at least heuristically, requiring often a deep understanding of the underlying
that these problems can be severely ill-posed, on the problem as well as the generation of measurements. Then,
discrete level at least ill-conditioned, in the sense that the one may also replace (19) by some other, possibly better
singular values of the linearized operator tend at an suited stopping criterion.
exponential decay to zero. Furthermore, as discussed in Section 6, when using
Moreover, we have demonstrated that it is possible, only one scalar function 𝜇𝑡 or the modification of a linear
provided sufficient data are available, to interpret the turbulence model, which ultimately only affects this
turbulent viscosity 𝜇𝑡 directly as a parameter in the RANS turbulent viscosity, we gave arguments that the
equations and to determine it on the basis of the given possibilities to improve in general solutions of the RANS
data. The feasibility of this approach was demonstrated for equations appear limited. As shown in (32), one such
a well-known example in CFD, namely a turbulent flow solution, just like any turbulence model, only represents a
over the RAE2822 airfoil at transonic conditions where the certain approximation to a possibly non-unique minimizer.
flow on the upper side of the airfoil has a shock. To solve Major improvements may only be expected when one
the problem the iteratively regularized Gauss–Newton allows for more degrees of freedom such as in Reynolds
method (17) was applied together with the discrepancy stress models. Therefore, it is certainly an interesting
principle (19) as a stopping criterion. It was shown that for question for future work, whether the components of the
noisy data this stopping criterion leads to reasonable Reynolds stress tensor can directly be reconstructed,
reconstructions of the turbulent viscosity such that instead of the turbulent viscosity 𝜇𝑡 or a parameter within
aerodynamic data such as 𝑐𝑝 and 𝑐𝑓 distributions are well a given turbulence model.
represented. Moreover, it was also demonstrated that in A further important issue for future work is the
particular for noisy data a violation of the stopping replacement of the direct linear solvers. Instead one may
criterion blurrs the turbulent viscosity 𝜇𝑡 such that for the consider to solve (23) by the conjugate gradient method
given test case strong oscillations in particular in a for the normal equations (CGNE). Then a direct evaluation
neighborhood of the shock in the 𝑐𝑓 distribution occur. of (25) and (26) can be avoided since only the matrix–
To solve such problem based on a finite-volume vector products are required. These can be realized, for
discretization, a methodology was suggested to implement example, by the application of a possibly preconditioned
on a discrete level suited function spaces approximating a GMRES method.
Sobolev 𝐻1-norm, which in our implementation was
CRediT authorship contribution statement
crucial for the success of the iteratively regularized Gauss–
Newton method. Neglecting such gradient information in
Stefan Langer: Data curation, Formal analysis, Funding
the formulation and regularization of the problem typically
acquisition, Investigation, Methodology, Project
gave instabilities in our implementation of the algorithm
administration, Resources, Software, Supervision,
and the reconstructions showed up strong oscillations.
Validation, Visualization, Writing – original draft, Writing –
Beyond that, based on the interpretation of the FIML
review & editing.
approach as a parameter identification problem for the
RANS equations, in Section 6 it was discussed in which way Declaration of competing interest
this technique can possibly be exploited if one aims for a
general improvement of a given turbulence model or if The authors declare that they have no known
one wants to represent a turbulent viscosity 𝜇𝑡 directly competing financial interests or personal relationships that
with the help of ansatz functions for a large number of test could have appeared to influence the work reported in this
cases. Due to the possibly ill-posed character of the paper.
problem and the large number of required data, the
determination of the unknown parameter seems very Data availability
challenging, since many large-scale problems need to be
solved at the same time to approximate a minimizer of a Data will be made available on request.
corresponding leastsquares problem. This might even be Acknowledgments
infeasible due to the numerical complexity and memory
requirements. A first way out would therefore possibly be The author gratefully acknowledges the scientific
to use the approach only for an improvement of a certain support and HPC resources provided by the German
class of problems in order to try to improve the underlying Aerospace Center (DLR). The HPC system CARA is partially
model for these applications. This seems under the current funded by ‘‘Saxon State Ministry for Economic Affairs,
technical possibilities, from the author’s point of view, Labour and Transport’’ and ‘‘Federal Ministry for Economic
more promising than trying to achieve general Affairs and Climate Action’’. This work has been done
improvements. within the framework of the DLR internal project ‘Machine
As a further way to reduce the numerical complexity of Learning and Quantum Computing - Digitalization of
the method, it was discussed to possibly summarize data Aircraft Development 2.0’.
and to solve a projected problem. Therefore, future work
11
S. Langer Computers and Fluids 284 (2024) 106438
References [24] Wilcox David C. Turbulence modeling for CFD. DCW Industries,
Incorporated; 1994.
[1] Spalart PR, Allmaras SR. A one-equation turbulence model for [25] Menter Florian R. Improved two-equation 𝑘-𝜔 turbulence
aerodynamic flows. In: AIAA computational fluid dynamics models for aerodynamic flows. Technical report, NASA; 1992.
conference. Conference proceeding series, AIAA; 1992, number [26] Baldwin BS, Lomax H. Thin layer approximation and algebraic
1992-439. model for separated turbulent flows. AIAA J 1978;78–257.
[2] Allmaras Steven R, Johnson Forrester T, Spalart Philippe R. [27] Engl Heinz W, Hanke Martin, Neubauer Andreas. Regularization
Modifications and clarifications for the implementation of the of inverse problems. Dordrecht, Boston, London: Kluwer
Spalart-Allmaras turbulence model. In: International conference Academic Publisher; 1996.
on computational fluid dynamics 7, hawaii. Conference [28] Langer Stefan. Complexity analysis of the iteratively regularized
proceeding series, 2012, number ICCFD7-1902. Gauss-Newton method with inner CG-iteration. J Inverse Ill-posed
[3] Wilcox David C. Reassessment of the scale-determining equation Probl 2009;17:871–90.
for advanced turbulence models. AIAA J 1988;26(11):1299–310. [29] Langer Stefan. Application of a line implicit method to fully
[4] Wilcox David C. Formulation of the 𝑘-𝜔 turbulence model coupled system of equations for turbulent flow problems. Int J
revisited. AIAA J 2008;46(11):2823–38. Comput Fluid Dyn
[5] Menter Florian R. Two-equation eddy-viscosity turbulence 2013;27(3):131–50.
models for engineering applications. AIAA J 1994;32:1598–605. [30] Langer Stefan. Agglomeration multigrid methods with implicit
[6] Menter Florian R, Kuntz M, Langtry R. Ten years of industrial Runge-Kutta smoothers applied to aerodynamic simulations on
experience with the SST turbulence model. Turbul Heat Mass unstructured grids. J Comput Phys 2014;277:72–100.
Transf 2003;4. [31] Langer Stefan. Preconditioned Newton methods to approximate
[7] Menter Florian R. Review of the shear-stress transport solutions of the Reynolds averaged Navier-Stokes equations.
turbulence model experience from an industrial perspective. Int J Technical report, Institut für Aerodynamik und Strömungstechnik;
Comput Fluid Dyn 2018.
2009;23(4):305–16. [32] Morozov VA. On the solution of functional equations by the
[8] Rodi W, Spalding DB. A two-parameter model of turbulence, and method of regularization. Sov Math Dokl 1966;7:414–7.
its application to free jets. Wärme Stoffübertragung [33] Langer Stefan, Hohage Thorsten. Convergence analysis of an
1970;3(2):85–95. inexact iteratively regularized Gauss-newton method under
[9] Jones WP, Launder BE. The prediction of laminarization with a general source conditions. J Inverse and Ill-posed Prob
two-equation model of turbulence. Int J Heat Mass Transfer 2007;15:311–27.
1972;15(2):301–14.
[10] Bredberg Jonas. On two-equation eddy-viscosity models. Internal
report, Göteborg, Sweden: Department of Thermo and Fluid
Dynamics, Chalmers University of Technology; 2001.
[11] Kamenetskii Dmitry S, Bussoletti John E, Hilmes Craig L,
Venkatakrishnan Venkat, Wigton Laurence B, Johnson Forrester T.
Numerical evidence of multiple solutions for the Reynolds-
averaged Navier-Stokes equations. AIAA J 2014;52(8).
[12] Hadamard Jacques. Jacques Hadamard lectures on Cauchy’s
problem in linear partial differential equations. New York: Dover
publications; 1952.
[13] Parish Eric J, Duraisamy Karthik. A paradigm for data-driven
predictive modeling using field inversion and machine learning. J
Comput Phys 2016;305:758–74.
[14] Jäckel Florian. Sensitivity analysis of discrepancy terms
introduced in turbulence models using field inversion. In:
Dillmann Andreas, Heller Gerd, Krämer Ewald, Wagner Claus,
editors. 22. STAB/DGLR symposium 2020. Notes on numerical
fluid mechanics and multidisciplinary design, vol. 151, Springer
Nature; 2021, p. 625–34, Hardcover ISBN 978-3-030-79560-3
Softcover ISBN 978-3-030-79563-4 eBook ISBN 978-3-030-79561-
0 Series ISSN 1612-2909 Series E-ISSN 1860-0824.
[15] Duraisamy Karthik. Perspectives on machine learning-augmented
reynoldsaveraged and large eddy simulation models of
turbulence. Phys Rev Fluids 2021;6:050504.
[16] Grabe Cornelia, Jäckel Florian, Khurana Parv, Dwight Richard P.
Data-driven augmentation of a RANS turbulence model for
transonic flow prediction. Internat J Numer Methods Heat Fluid
Flow 2023;33:1544–61.
[17] Jäckel Florian. A closed-form correction for the Spalart-Allmaras
turbulence model for separated flows. AIAA J 2023;61(6):2319–
30.
[18] Colton David, Kreß Rainer. Inverse acoustic and electromagnetic
scattering theory. 2nd ed.. Berlin, Heidelberg, New York: Springer
Verlag; 1997.
[19] Hohage Thorsten. On the numerical solution of a three-
dimensional inverse medium scattering problem. Inverse
Problems 2001;17:1743–63.
[20] Hohage Thorsten. Fast numerical solution of the electromagnetic
medium scattering problem and applications to the inverse
problem. J Comput Phys 2006;214:224–38.
[21] Langer Stefan. Preconditioned Newton methods for ill-posed
problems (Ph.D. thesis), University of Göttingen; 2007.
[22] Langer S. Investigation of preconditioning techniques for the
iteratively regularized Gauss—Newton method for exponentially
ill-posed problems. SIAM J Sci Comput 2010;32(5):2543–59.
[23] Hohage Thorsten, Langer Stefan. Acceleration techniques for
regularized Newton methods applied to electromagnetic inverse
medium scattering problems. Inverse Problems
2010;26(7):074011.
12