To first-order in the parameter variations, the 1st-order variation
around the nominal temperature distribution,
, in the electrically heated rod is the solution of the 1st-LFSS obtained by determining the G-differentials of Equations (3)–(5). By definition, the G-differentials of Equations (3)–(5) are obtained as follows
The interface condition expressed by Equation (70) couples the variation in the rod temperature , to the variation in the fluid temperature, , thereby coupling Equations (67), (69) and (70) to Equations (10) and (11), the totality of which constitute the 1st-LFSS. As has been discussed throughout this work, it is computationally expensive to repeatedly solve the 1st-LFSS, namely Equations (10), (11), (67), (69) and (70), for all possible parameter variations.
4.1. Spectral Expansion of the Response Sensitivities
In this Section, the 1st-LASS corresponding to the 1st-LFSS comprising Equations (10), (11), (67), (69) and (70) will be developed by applying the general principles outlined in [
1], to represent the temperature distribution variation,
, in the heated rod by means of a spectral expansion. As in
Section 3.1, it is convenient to use the Legendre polynomials,
as the spectral basis functions for representing the function
in the
-direction. It is therefore convenient to rescale Equations (67), (69) and (70) in the
-direction by using the scaling transformation indicated in Equation (13) the 1st-LFSS. Consequently, the re-scaled 1st-LFSS for the coupled variations
and
are as follows
For convenient reference and completeness of the 1st-LFSS, Equations (18) and (19) were reproduced above as Equations (74) and (75).
The recommended [
5,
6] spectral basis functions for representing a continuous function like
on the finite interval
would be Chebyshev Polynomials. However, their use would be very similar to the use of the Legendre Polynomials as the spectral basis functions for the expansion in the axial direction
. Consequently, to illustrate the use of a spectral basis for which there are no recursion relations among the underlying spectral elements, Fourier series expansions will be used in the radial direction for representing
. Since
is defined on the half-range
, either Fourier Sine or Cosine Series can be employed to represent
in the radial direction. Since the result in Equation (7) indicates that
is an even function in
and since the Cosine Series converges faster than the Sine Series, the former will be used. Thus,
can be represented over the domain
using the following spectral representation
where the generalized Fourier spectral coefficients
are defined as follows
The appearance of the unknown temperature variation
in the expressions of the generalized Fourier spectral coefficients
will be eliminated by constructing equivalent expressions for these coefficients in terms of adjoint functions, which will be the solutions of the 1st-LASS corresponding to the 1st-LFSS comprising Equations (71)–(75). The Hilbert space appropriate for constructing this 1st-LASS comprises square-integrable two-component vector functions of the form
, endowed with an inner product
of the form
Using the definition provided in Equation (79), construct the inner product of a square integrable vector function
with Equations (71) and (74), respectively, to obtain the following relation
The left-side of Equation (80) is now integrated by parts (twice over the variable
and once over the variable
) to obtain the following relation
Using the boundary condition given in Equation (72) and imposing the boundary condition
eliminates the last term on the right-side of Equation (81), including the unknown function
. Imposing the boundary condition
eliminates the unknown function
in the third term on the right-side of Equation (81). Using the boundary condition given in Equation (75) to replace the quantity
which appears in the third term on the right side of Equation (81) and replacing the left-side of Equation (81) by the right-side of Equation (80) yields the following expression, equivalent to Equation (81)
The unknown quantity
, which appears in the last term on the right-side of Equation (84) is eliminated by using the boundary condition given in Equation (73); this operation transforms Equation (84) into the following form
The unknown quantity
, which appears in third and fourth terms on the right-side of Equation (85) is eliminated by imposing the following interface condition on the (adjoint) function
Inserting Equation (86) into the right-side of Equation (85) reduces the latter equation to the following form
The two terms that contain the unknown function
in Equation (87) are grouped together, transforming Equation (87) into the following form
The second-term on the right-side of Equation (88) will represent the generalized Fourier spectral coefficients
defined in Equation (77) by requiring that the following equations be satisfied
Altogether, the relations required in Equations (89) and (90), and the boundary and interface conditions already imposed in Equations (82), (83), and (86) for
constitute the 1st-LASS for the adjoint functions
. Solving this 1st-LASS yields the following expressions for
Inserting Equations (89) and (90) into Equation (88) and re-arranging the resulting relation yields the following expression for the generalized Fourier spectral coefficients
Inserting the definitions provided for
and
in Equations (71) and (74), respectively, into Equation (92) yields the following expression for the generalized Fourier spectral coefficients
The second-term on the right-side of Equation (88) also represents the generalized Fourier spectral coefficients
, defined in Equation (78) by requiring that the following equations be satisfied
Inserting Equations (94) and (95) into Equation (88) and re-arranging the resulting relation yields the following expression for the generalized Fourier spectral coefficients
Inserting the definitions provided for
and
in Equations (71) and (74), respectively, into Equation (96) yields the following expression for the generalized Fourier spectral coefficients
Altogether, the relations required in Equations (94) and (95), and the boundary and interface conditions already imposed in Equations (82), (83), and (86) for
constitute the 1st-LASS for the adjoint functions
. Solving this 1st-LASS yields the following expressions for
Noting from Equation (98) that
inserting Equations (93) and (97) into Equation (76) and equating the expressions that multiply each of the respective arbitrary parameter variation yields the following expressions for the sensitivities of the rod temperature
with respect to the various parameters
Inserting the results provided in Equations (91) and (98) in Equations (100)–(107), noting that
and using the relations provided in Equation (13) to revert from the variable
to the variable
yields, after considerable algebra, the following final exact expressions for the sensitivities of the temperature distribution
within the heated rod to the model and boundary parameters
Notably, the 1st-LASS is solved in a manner which is “reverse/backwards” by comparison to the way in which solution proceeds for solving the First-Level Forward Sensitivity System (1st-LFSS) as well as the original heat transport model. Thus, while the 1st-LFSS and the original heat transport model are solved by starting with the fluid flow equation (which is solved from the inlet to the outlet of the fluid flow) and subsequently solving the heat conduction equation in the rod, the solution of the 1st-LASS proceeds in the reverse manner, by first solving the heat conduction in the rod, followed by solving the fluid flow equation from the outlet to the inlet.
In practice, it is evidently not possible to infinitely compute many basis functions and corresponding expansion coefficients. Consequently, the exact expressions for the response sensitivities presented in Equations (109)–(116) cannot be attained. In practice, known convergence criteria for the various expansions see, e.g., [
5,
6] would be used in order to decide the number of expansion coefficients to be computed by solving the 1st-LASS. It is important to note that the issue of computational accuracy regarding the spectral expansion refers not to the accuracy of the functional derivative of the response in the phase-space of model parameters (in which space the 1st-CASAM provides exact expressions), but refers to the representation of the respective sensitivity as a function in the phase-space of independent variables.
4.2. Collocation Pseudo-Spectral Expansion of the Response Sensitivities
The collocation pseudo-spectral expansion of the fluid temperature variation
can be written in the following form
where
and
denotes the chosen cardinal functions and where
In Equation (118), the quantities
denote the collocation points in the radial direction while the quantities
denote the collocation points in the axial direction. The functional
in Equation (118) can be expressed in terms of the solution of a 1st-LASS that is constructed by following the same conceptual steps as those leading to Equation (82), (83), (86), (94) and (95). Thus, denoting as
and
the adjoint sensitivity functions that correspond to the forward functions
and
, respectively, and following the procedure outlined in [
1] leads to the following expressions for the sensitivities of
in terms of the adjoint functions
and
In Equations (119)–(126), the adjoint functions
and
are the solutions of the following 1st-LASS
Solving Equations (127)–(131) yields the following closed-form expressions for the adjoint functions
and
where
denotes the customary Heaviside functional. i.e.,
and
.
Inserting the expressions obtained in Equations (132) and (133) into Equations (119)–(126) yields the following expressions for the response sensitivities
The 1st-LASS is solved in a manner that is “reverse/backwards” by comparison to the way in which solution proceeds for solving the 1st-LFSS as well as the original heat transport model, by first solving the heat conduction in the rod, followed by solving the fluid flow equation from the outlet to the inlet.
4.3. Mixed Spectral/Collocation Expansion of the Response Sensitivities
In contrast to
, the total sensitivity
of the rod temperature (with respect to the model parameters) depends on more than one independent variable. Therefore, a mixed spectral/collocation representation of
over the domain
may be contemplated, as provided below
where
denotes the
nth-order Legendre polynomial,
denotes the
mth-cardinal function chosen for the radial direction and the functionals
denote the Legendre spectral coefficients defined at collocation points
, in the radial direction, as follows
The 1st-LFSS appropriate for expressing the functionals
in terms of adjoint functions is constructed by following the same conceptual steps as followed in
Section 3.1 and
Section 3.2. Denoting the adjoint functions that would correspond to
as
and
, respectively, and following the same steps as those leading to Equation (88) yields the following counterpart of Equation (88)
The second-term on the right-side of Equation (144) will represent the coefficients
defined in Equation (143) by requiring that the following equations, which represent the 1st-LASS for the adjoint functions
and
, be satisfied
inserting Equations (93) and (97) into Equation (76) and equating the expressions that multiply each of the respective arbitrary parameter variation yields the following expressions for the sensitivities of the rod temperature
with respect to the various parameters
Solving the 1st-LASS represented by Equations (145)–(149) yields the following expressions for the adjoint functions
and
Inserting the results from Equations (158) and (159) into Equations (150)–(157), carrying out the respective algebraic operations and using Equation (13) to revert from the independent variable
to the independent variable
yields the following expressions for the sensitivities of the time-dependent temperature
at a collocation point
within the heated rod to the model and boundary parameters
As Equations (150)–(157) indicate, if only a finite amount of adjoint functions and are computed, then the expressions in these equations will be truncated and the time-dependent accuracy of the respective sensitivities at the collocation point will be limited by the order of the respective expansion in Legendre polynomials. Spatially, the accuracy of the respective sensitivities will be limited by the amount of collocation points taken into consideration.