Scott Moe

Donsub Rim, Scott Moe, Randall J. LeVeque

Snapshot matrices built from solutions to hyperbolic partial differential equations exhibit slow decay in singular values, whereas fast decay is crucial for the success of projection- based model reduction methods. To overcome this problem, we build on previous work in symmetry reduction [Rowley and Marsden, Physica D (2000), pp. 1-19] and propose an iterative algorithm that decomposes the snapshot matrix into multiple shifting profiles, each with a corresponding speed. Its applicability to typical hyperbolic problems is demonstrated through numerical examples, and other natural extensions that modify the shift operator are considered. Finally, we give a geometric interpretation of the algorithm.

Scott A. Moe, James A. Rossmanith, David C. Seal

This work introduces a single-stage, single-step method for the compressible Euler equations that is provably positivity-preserving and can be applied on both Cartesian and unstructured meshes. This method is the first case of a single-stage, single-step method that is simultaneously high-order, positivity-preserving, and operates on unstructured meshes. Time-stepping is accomplished via the Lax-Wendroff approach, which is also sometimes called the Cauchy-Kovalevskaya procedure, where temporal derivatives in a Taylor series in time are exchanged for spatial derivatives. The Lax-Wendroff discontinuous Galerkin (LxW-DG) method developed in this work is formulated so that it looks like a forward Euler update but with a high-order time-extrapolated flux. In particular, the numerical flux used in this work is a linear combination of a low-order positivity-preserving contribution and a high-order component that can be damped to enforce positivity of the cell averages for the density and pressure for each time step. In addition to this flux limiter, a moment limiter is applied that forces positivity of the solution at finitely many quadrature points within each cell. The combination of the flux limiter and the moment limiter guarantees positivity of the cell averages from one time-step to the next. Finally, a simple shock capturing limiter that uses the same basic technology as the moment limiter is introduced in order to obtain non-oscillatory results. The resulting scheme can be extended to arbitrary order without increasing the size of the effective stencil. We present numerical results in one and two space dimensions that demonstrate the robustness of the proposed scheme.

Naman Maheshwari, Nicholas Malaya, Scott Moe et al.

For Deep Neural Networks (DNNs) to become useful in safety-critical applications, such as self-driving cars and disease diagnosis, they must be stable to perturbations in input and model parameters. Characterizing the sensitivity of a DNN to perturbations is necessary to determine minimal bit-width precision that may be used to safely represent the network. However, no general result exists that is capable of predicting the sensitivity of a given DNN to round-off error, noise, or other perturbations in input. This paper derives an estimator that can predict such quantities. The estimator is derived via inequalities and matrix norms, and the resulting quantity is roughly analogous to a condition number for the entire neural network. An approximation of the estimator is tested on two Convolutional Neural Networks, AlexNet and VGG-19, using the ImageNet dataset. For each of these networks, the tightness of the estimator is explored via random perturbations and adversarial attacks.

Scott A. Moe, James A. Rossmanith, David C. Seal

The discontinuous Galerkin (DG) finite element method when applied to hyperbolic conservation laws requires the use of shock-capturing limiters in order to suppress unphysical oscillations near large solution gradients. In this work we develop a novel shock-capturing limiter that combines key ideas from the limiter of Barth and Jespersen [AIAA-89-0366 (1989)] and the maximum principle preserving (MPP) framework of Zhang and Shu [Proc. R. Soc. A, 467 (2011), pp. 2752--2776]. The limiting strategy is based on traversing the mesh element-by-element in order to (1) find local upper and lower bounds on user-defined variables by sampling these variables on neighboring elements, and (2) to then enforce these local bounds by minimally damping the high-order corrections. The main advantages of this limiting strategy is that it is simple to implement, effective at shock capturing, and retains high-order accuracy of the solution in smooth regimes. The resulting numerical scheme is applied to several standard numerical tests in both one and two-dimensions and on both Cartesian and unstructured grids. These tests are used as benchmarks to verify and assess the accuracy and robustness of the method.