Material Interface Reconstruction Using Machine Learning

Authors: Daniel S. Fenn, Katie E. Lewis, and Charles Doutriaux (Lawrence Livermore National Laboratory)

Abstract: In multi-material simulations, it is important to track material interfaces. These are frequently not tracked explicitly and must be reconstructed from zone data. Current methods provide either material conservation or interface continuity, but not both, meaning that many interfaces may be constructed erroneously, affecting simulation accuracy. Meanwhile, progress in image-related machine learning (ML) fields is noteworthy, and several of such fields exhibit conceptual similarity to the material interface reconstruction (MIR) problem. Here we investigate the application of image-based, supervised learning methods to MIR. We generate images by taking “snapshots” of a mesh, with the material information encoded as a pixel value. We feed these images to a network that infers the interface morphology. We use an autoencoder design and generate synthetic data. Our network is able to accurately reproduce correct interfaces for most cases. Our promising results indicate that the application of ML to MIR warrants further study.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF


Back to Poster Archive Listing