Autoencoders for Emulation and Calibration of Dynamic Compression Experiments

• Activity Number: 462 - Making an Impact When Things Make Impacts
• Type: Topic Contributed
• Date/Time: Wednesday, July 31, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Uncertainty Quantification in Complex Systems Interest Group
• Author(s): Natalie Klein* and Earl Christopher Lawrence
• Companies: Carnegie Mellon University and Los Alamos National Laboratory
• Keywords: model calibration; uncertainty quantification; Gaussian process; functional data analysis; computer experiments; neural network

Abstract

In this work, we address statistical model calibration and uncertainty quantification based on dynamic materials simulations and experiments in which the outputs are functions of velocity over time following an impact. While previous work on Bayesian calibration for multivariate outputs uses linear dimensionality reduction, in the current application, such representations are not parsimonious due to the presence of both phase and amplitude variability in the functions. Instead, we propose using an autoencoder neural network, a powerful machine learning-based nonlinear dimensionality reduction method, to represent functional variation in a low-dimensional space. Our results suggest that the autoencoder representation leads to accurate and computationally efficient Gaussian process emulation and Bayesian calibration in real data. Based on the empirical success of neural networks in representing complex multivariate data, such as images, we expect that the autoencoder approach could be useful for emulation and calibration of a variety of complex simulation and experimental outputs.