1: \begin{abstract}

2:   We develop a Petrov-Galerkin stabilization method for multiscale

3:   convection-diffusion transport systems.  Existing stabilization

4:   techniques add a limited number of degrees of freedom in the form of

5:   bubble functions or a modified diffusion, which may not sufficient

6:   to stabilize multiscale systems. We seek a local reduced-order model

7:   for this kind of multiscale transport problems and thus, develop a

8:   systematic approach for finding reduced-order approximations of the

9:   solution. We start from a Petrov-Galerkin framework using optimal

10:   weighting functions.  We introduce an auxiliary variable to a mixed

11:   formulation of the problem. The auxiliary variable stands for the

12:   optimal weighting function. The problem reduces to finding a test

13:   space (a reduced dimensional space for this auxiliary variable),

14:   which guarantees that the error in the primal variable (representing

15:   the solution) is close to the projection error of the full solution

16:   on the reduced dimensional space that approximates the solution.  To

17:   find the test space, we reformulate some recent mixed Generalized

18:   Multiscale Finite Element Methods. We introduce snapshots and local

19:   spectral problems that appropriately define local weight and trial

20:   spaces.  In particular, we use energy minimizing snapshots and local

21:   spectral decompositions in the natural norm associated with the

22:   auxiliary variable. The resulting spectral decomposition adaptively

23:   identifies and builds the optimal multiscale space to stabilize the

24:   system.  We discuss the stability and its relation to the

25:   approximation property of the test space.  We design online basis

26:   functions, which accelerate convergence in the test space, and

27:   consequently, improve stability.  We present several numerical

28:   examples and show that one needs a few test functions to achieve an

29:   error similar to the projection error in the primal variable

30:   irrespective of the Peclet number.

31: \end{abstract}

32: