Inverse Problems and Machine Learning

Date: January 29, 2020

Time: 12:00PM - 1:00PM

Location: BLOC 628

Speaker: Mauricio Tano, Department of Nuclear Engineering, Texas A&M University

Title: Artificial Neural Networks to Accelerate Radiation Transport Sweeps

Abstract: Transport sweeping is a fast and efficient solution technique for solving the Sn (discrete ordinates) first-order transport equation discretized using Discontinuous Finite Elements. In transport sweeps, the angular flux solution is obtained for one angle and one spatial cell at a time by solving a small local system of equations. Gaussian Elimination (GE) is typically invoked to perform this task. However, GE may take up to 50% of the computing time spent in the sweeping routine. Here, we investigate the use of Machine Learning techniques to bypass assembling and solving the local system. In our approach, a shallow feed-forward artificial neural network (ANN) is chosen. The architecture of the ANN is to be optimized in order to reduce the number of operations compared to the Gaussian Elimination solve, while keeping a bounded error in the solution. The resulting ANN consists of an ANN with only one hidden layer. Numerical results show that speedups of a factor 3-4 are attainable by training a shallow artificial neural network.