Scribd
Upload
12 views3 pages

TSFP-7 Original

This paper introduces a new approach for generating transitional initial and inflow boundary conditions for direct numerical simulations (DNS) using variational data assimilation (VDA). The method optimizes the initial conditions based on observed data to accurately reproduce the dynamics of spatially evolving flows, utilizing automatic differentiation for gradient computation. Future work aims to extend this technique to include three-dimensional effects and assess its performance with experimental data.

Uploaded by

Alejandro Gronskis
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
12 views3 pages

TSFP-7 Original

This paper introduces a new approach for generating transitional initial and inflow boundary conditions for direct numerical simulations (DNS) using variational data assimilation (VDA). The method optimizes the initial conditions based on observed data to accurately reproduce the dynamics of spatially evolving flows, utilizing automatic differentiation for gradient computation. Future work aims to extend this technique to include three-dimensional effects and assess its performance with experimental data.

Uploaded by

Alejandro Gronskis
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Inflow and initial conditions for direct numerical simulation based on

adjoint data assimilation


A. Gronskis1 , D. Heitz2,3 and E. Mémin1
1
INRIA Rennes - Bretagne Atlantique, Campus universitaire de Beaulieu, F-35042 Rennes, France.
2 Cemagref, 17 avenue de Cucillé, CS64427, F-35044 Rennes, France.
3 Université européenne de Bretagne, Rennes, France.

Abstract
In order to limit the computational cost associated to direct numerical (DNS) or large-eddy
simulations (LES) of spatially evolving flows, new strategies that aim at simulating only the
region of interest of the flow have recently been developed [1]. A potential implication is that
any results computed may be strongly influenced by the prescribed instantaneous inflow ve-
locity profiles. Such profiles are practically never available, and a usual practice is to generate
synthetic inflow data satisfying certain properties, which may be known from experimental
data or empirical correlations. This paper presents a new approach for generating simultane-
ous transitional initial and inflow boundary conditions based on variational data assimilation
(VDA) that is capable of reproducing the spatiotemporal dynamics of the targeted flow [2].
VDA is devoted to the solution of a minimization problem combining quantitative information
on a phenomenon with modelling results by means of the optimal control theory. Such an
inverse problem makes use of an objective function that measures some discrepancies between
acquired data and numerical solutions. A minimization problem involving a compound func-
tion (objective function plus model) is set up for the identification of some inputs of the model
(our control parameters) from which the solution fits at best the observed data. Assuming that
both the model and the objective function are differentiable, VDA proposes to solve this inverse
problem looking for a control that cancels out the gradient of this compound function. In the
present study, the optimisation is carried out using the limited memory BFGS deterministic
gradient based optimization algorithm (L-BFGS) [3]. In order to generate the gradient of the
objective function, we use the algorithms and tools of Automatic Differentiation (AD). This
can be viewed as a program transformation carried out by regarding the original computer
code as a sequence of elementary functions having at least a first derivative and then using the
chain rule to automatically evaluate the function represented by the given code as well as its
derivative. This is indeed very close to the mathematical resolution with the adjoint state: the
reverse mode of AD actually computes a discretized adjoint on the program. The AD tool used
for our purpose is TAPENADE [4]. The dynamical model that relates the state function to
the unknown of our inverse problem is based on pressure-velocity formulation of Navier-Stokes
equations. A DNS code fully based on sixth-order compact finite difference schemes and a
Cartesian grid is used to solve the incompressible Navier-Stokes equations. In this context, the
incompressibility condition is ensured via a fractional step method introducing a Poisson equa-
tion for the pressure [5]. We generated tangent (forward mode AD) and adjoint (reverse mode
AD) codes for this DNS code. Subsequently, we performed validation experiments to check for
correctness of both AD generated codes. The results are shown in table 1. The tangent norm
and the adjoint norm match very well, up to the last few digits, which shows that the tangent
and adjoint codes really compute the same derivatives, only in a different computation order.
The norm obtained with Divided Differences (DD) matches only to half the machine precision,

Preprint submitted to TSFP-7 September 30, 2010


because of the weakness of DD approximation. In order to experimentally assess the benefits of
our technique, we carried out a VDA experiment applied to the identification of the initial con-
ditions from a spatially evolving two-dimensional mixing layer flow. The velocity ratio between
low and high speed is 0.66 and the Reynolds number based on the velocity difference of two
streams and the inflow vorticity thickness is Re = 400. As a first step, a precursor simulation
was run to provide artificial inflow conditions for the assimilation experiment. The spanwise
section retained which will correspond to the inlet section of the assimilation domain is lo-
cated at the beginning of the convection region of vortical structures. Subsequently, synthetic
observations generated by the direct model trajectory are perturbed using zero mean Gaus-
sian noise and an additional spatial filtering is applied to these vector velocity fields in order
to provide smoother motion flow. Figure 1(a) illustrates the performance of the optimization
loop for an integration period of 2 s (i.e. 200 time steps) and an observation frequency equal
to 0.1 s. Figure 1(b) shows the temporal evolution of the assimilation error with respect to
the unperturbed observations (our reference or true state). The initial trajectory is generated
from the model output starting from the perturbed observation at t = t0 (initial estimate for
L-BFGS algorithm), and the assimilated one is obtained from the best estimate for the initial
state found by L-BFGS. As can be seen, the minimization algorithm corrects the error of the
initial guess and converges to a global minimum point that is close to the true state point,
allowing to retrieve the initial condition of the model that yields a good agreement with the
perturbed velocity fields. Figure 2 indicates that the true state at the end of the assimilation
window (a) is recovered with a good approximation (d) from the perturbed one (b), showing
the quality of the derivatives obtained. The proposed method allows to recover the state of an
unknown function on the basis of a DNS model and noisy measurement of this function, by
modifying the initial condition of the system. To go further, it could be interesting to introduce
the inflow boundary condition as a control parameter in our optimization problem. Future work
will also include extending the adjoint code capabilities to include three-dimensional effects in
data assimilation system. The quality of the assimilated state will be assessed in situations
like those in which the sequence of observations come from experimental database acquired by
particle image velocimetry. In addition, another flow configurations will be considered to fully
validate the proposed method. For example, in wake flows, one can consider inlet boundary
conditions taken far downstream from the body in order to avoid the expensive computation
of the near wall region.

References
[1] L. Perret, J. Delville, R. Manceau and J. P. Bonnet. Turbulent inflow conditions for large-
eddy simulation based on low-order empirical model. Phys. Fluids 20, 075107 (2008).

[2] N. Papadakis and E. Memin. Variational assimilation of fluid motion from image sequences.
SIAM Journal on Imaging Science 1(4), 343-363, 2008.

[3] D. Liu and J. Nocedal. On the limited memory bfgs method for large scale optimization.
Mathematical Programming, Series B 45(3), 503-528 (1989).

[4] L. Hascoet, R. Greborio and V. Pascual. Computing adjoints by automatic differentiation


with tapenade. Research report, INRIA (2003).

[5] P. Parnaudeau, J. Carlier, D. Heitz and E. Lamballais. Experimental and numerical studies
of the flow over a circular cylinder at Reynolds number 3900. Phys. Fluids 20(8), 2008.

2
hDD, DDi ( = 10−7 ) 6.30295013229048399E-004
hẎ , Ẏ i 6.30295332305800073E-004
hX̄, Ẋi 6.30295332305800941E-004
Table 1: Dot product test for AD generated codes. F being the cost function and choosing an arbitrary initial
state X and an arbitrary direction Ẋ, we compute the divided difference DD = [F (X + Ẋ) − F (X)]/. Using
0
the tangent differentiated program, we compute Ẏ = F (X) × Ẋ. Using the adjoint differentiated program, we
0
compute X̄ = F ∗ (X) × Ẏ .

(a) (b)

Figure 1: (a) Reduction of the cost function and the gradient norm by varying the number of minimization
Rt R
iterations with the L-BFGS algorithm. The cost function to be minimized is F (~u(~x, t0 )) = t0N Ω k~u(~x, t) −
~uobs (~x, t)k2 d~x dt where the superscript obs stands for perturbed observation and ~u(~x, t) is the velocity field
generated from the model output. We have chosen a stop criterion based on gradient reduction from its initial
value, i.e. k∇F (Xk )k/k∇F (X0 )k < 10−4 . (b) Temporal evolution of the squared norm of the discrepancy
between each vorticity field generated from the model output wz (~x, t) and the reference value wzref (~x, t).

Figure 2: Vorticity fields at the end of the assimilation window. Computational domain Ω is discretized with
401 × 257 nodes on a stretched grid. (a) True field; (b) Perturbed field; (c) Field obtained by starting from
the initial estimate for L-BFGS algorithm; (d) Field obtained by starting from the retrieved optimal initial
conditions.

You might also like