Brian Wetton

Wan Chen, Brian Wetton

This paper describes a technique to determine the linear well-posedness of a general class of vector elliptic problems that include a steady interface, to be determined as part of the problem, that separates two subdomains. The interface satisfies mixed Dirichlet and Neumann conditions. We consider ``2+2'' models, meaning two independent variables respectively on each subdomain. The governing equations are taken to be vector Laplacian, to be able to make analytic progress. The interface conditions can be classified into four large categories, and we concentrate on the one with most physical interest. The well-posedness criteria in this case are particularly clear. In many physical cases, the movement of the interface in time-dependent situations can be reduced to a normal motion proportional to the residual in one of the steady state interface conditions (the elliptic interior problems and the other interface conditions are satisfied at each time). If only the steady state is of interest, one can consider using other residuals for the normal velocity. Our analysis can be extended to give insight into choosing residual velocities that have superior numerical properties. Hence, in the second part, we discuss an iterative method to solve free boundary problems. The advantages of the correctly chosen, non-physical residual velocities are demonstrated in a numerical example, based on a simplified model of two-phase flow with phase change in porous media.

Maricela Best McKay, Nathan P. Lawrence, Brian Wetton et al.

The Gauss-Newton matrix is widely viewed as a positive semidefinite approximation of the Hessian, yet mounting empirical evidence shows that Gauss-Newton descent outperforms Newton's method. We adopt a function space perspective to analyze this phenomenon. We show that the generalized Gauss-Newton (GGN) matrix projects the Newton direction in function space onto the model's tangent space, while a Jacobian-only variant obtained by applying the least squares Gauss-Newton matrix to non-least squares losses projects the function space loss gradient onto this same tangent space. Both projections eliminate distortions from the model's parameterization. Specifically, the evolution of the prediction-target mismatch depends on the model's parameterization through the matrix $JJ^\top$ where $J$ is the Jacobian of the model with respect to its parameters. The projections effectively replace $JJ^\top$ with the identity. We call this effect error whitening. Once the parameterization is removed, the prediction-target mismatch evolves according to dynamics dictated by the structure of the loss and the projection produced by the optimizer. Error whitening is a special property of Gauss-Newton descent that rigorously distinguishes it from Newton's method. We empirically demonstrate that Gauss-Newton optimizers follow the theoretically predicted function space dynamics and outperforms Newton's method, Adam, and Muon across case studies spanning supervised learning, physics-informed deep learning, and approximate dynamic programming.

Maricela Best McKay, Bhushan Gopaluni, Brian Wetton

Lithium ion batteries are widely used in many applications. Battery management systems control their optimal use and charging and predict when the battery will cease to deliver the required output on a planned duty or driving cycle. Such systems use a simulation of a mathematical model of battery performance. These models can be electrochemical or data-driven. Electrochemical models for batteries running at high currents are mathematically and computationally complex. In this work, we show that a well-regarded electrochemical model, the Pseudo Two Dimensional (P2D) model, can be replaced by a computationally efficient Convolutional Neural Network (CNN) surrogate model fit to accurately simulated data from a class of random driving cycles. We demonstrate that a CNN is an ideal choice for accurately capturing Lithium ion concentration profiles. Additionally, we show how the neural network model can be adjusted to correspond to battery changes in State of Health (SOH).

Zhenguo Pan, Brian Wetton

We study the coupled surface and grain boundary motion in a bicrystal in the context of the "quarter loop" geometry. Two types of physics motions are involved in this model: motion by mean curvature and motion by surface diffusion. The goal is finding a formulation that can describe the coupled motion and has good numerical behavior when discretized. Two formulations are proposed in this paper. One of them is given by a mixed order parabolic system and the other is given by Partial Differential Algebraic Equations. The parabolic formulation constitutes several parabolic equations which model the two normal direction motions separately. The performance of this formulation is good for a short time simulation. It performs even better by adding an extra term to adjust the tangential velocity of grid points. The PDAE formulation preserves the scaled arc length property and performs much better with no need to add an adjusting term. Both formulations are proven to be well-posed in a simpler setting and are solved by finite difference methods.