1: \begin{abstract} 

2:   Surrogate models for computational simulations are input-output

3:   approximations that allow computationally intensive analyses, such

4:   as uncertainty propagation and inference, to be performed

5:   efficiently. When a simulation output does not depend smoothly on

6:   its inputs, the error and convergence rate of many approximation

7:   methods deteriorate substantially. This paper details a method for

8:   efficiently localizing discontinuities in the input parameter

9:   domain, so that the model output can be approximated as a piecewise

10:   smooth function.

11:   % 

12:   The approach comprises an initialization phase, which uses

13:   polynomial annihilation to assign function values to different

14:   regions and thus seed an automated labeling procedure, followed by a

15:   refinement phase that adaptively updates a kernel support vector

16:   machine representation of the separating surface via active

17:   learning. The overall approach avoids structured grids and exploits

18:   any available simplicity in the geometry of the separating surface,

19:   thus reducing the number of model evaluations required to localize

20:   the discontinuity.

21:   % 

22:   The method is illustrated on examples of up to eleven dimensions,

23:   including algebraic models and ODE/PDE systems, and demonstrates

24:   improved scaling and efficiency over other discontinuity

25:   localization approaches.

26: \end{abstract}

27: