-
Mesoscale simulation of woven composite design decisions
Authors:
Lincoln N. Collins,
Scott A. Roberts
Abstract:
Characterizing the connection between material design decisions/parameters and their effective properties allows for accelerated materials development and optimization. We present a global sensitivity analysis of woven composite thermophysical properties, including density, volume fraction, thermal conductivity, specific heat, moduli, permeability, and tortuosity, predicted using mesoscale finite…
▽ More
Characterizing the connection between material design decisions/parameters and their effective properties allows for accelerated materials development and optimization. We present a global sensitivity analysis of woven composite thermophysical properties, including density, volume fraction, thermal conductivity, specific heat, moduli, permeability, and tortuosity, predicted using mesoscale finite element simulations. The mesoscale simulations use microscale approximations for the tow and matrix phases. We performed Latin hypercube sampling of viable input parameter ranges, and the resulting effective property distributions are analyzed using a surrogate model to determine the correlations between material parameters and responses, interactions between properties, and finally Sobol' indices and sensitivities. We demonstrate that both constituent physical properties and the mesoscale geometry strongly influence the composite material properties.
△ Less
Submitted 27 April, 2021;
originally announced April 2021.
-
Quantifying the unknown impact of segmentation uncertainty on image-based simulations
Authors:
Michael C. Krygier,
Tyler LaBonte,
Carianne Martinez,
Chance Norris,
Krish Sharma,
Lincoln N. Collins,
Partha P. Mukherjee,
Scott A. Roberts
Abstract:
Image-based simulation, the use of 3D images to calculate physical quantities, fundamentally relies on image segmentation to create the computational geometry. However, this process introduces image segmentation uncertainty because there is a variety of different segmentation tools (both manual and machine-learning-based) that will each produce a unique and valid segmentation. First, we demonstrat…
▽ More
Image-based simulation, the use of 3D images to calculate physical quantities, fundamentally relies on image segmentation to create the computational geometry. However, this process introduces image segmentation uncertainty because there is a variety of different segmentation tools (both manual and machine-learning-based) that will each produce a unique and valid segmentation. First, we demonstrate that these variations propagate into the physics simulations, compromising the resulting physics quantities. Second, we propose a general framework for rapidly quantifying segmentation uncertainty. Through the creation and sampling of segmentation uncertainty probability maps, we systematically and objectively create uncertainty distributions of the physics quantities. We show that physics quantity uncertainty distributions can follow a Normal distribution, but, in more complicated physics simulations, the resulting uncertainty distribution can be both nonintuitive and surprisingly nontrivial. We also establish that simply bounding the uncertainty can fail in situations that are sensitive to image segmentation. While our work does not eliminate segmentation uncertainty, it makes visible the previously unrecognized range of uncertainty currently plaguing image-based simulation, enabling more credible simulations.
△ Less
Submitted 9 September, 2021; v1 submitted 17 December, 2020;
originally announced December 2020.
-
We Know Where We Don't Know: 3D Bayesian CNNs for Credible Geometric Uncertainty
Authors:
Tyler LaBonte,
Carianne Martinez,
Scott A. Roberts
Abstract:
Deep learning has been successfully applied to the segmentation of 3D Computed Tomography (CT) scans. Establishing the credibility of these segmentations requires uncertainty quantification (UQ) to identify untrustworthy predictions. Recent UQ architectures include Monte Carlo dropout networks (MCDNs), which approximate deep Gaussian processes, and Bayesian neural networks (BNNs), which learn the…
▽ More
Deep learning has been successfully applied to the segmentation of 3D Computed Tomography (CT) scans. Establishing the credibility of these segmentations requires uncertainty quantification (UQ) to identify untrustworthy predictions. Recent UQ architectures include Monte Carlo dropout networks (MCDNs), which approximate deep Gaussian processes, and Bayesian neural networks (BNNs), which learn the distribution of the weight space. BNNs are advantageous over MCDNs for UQ but are thought to be computationally infeasible in high dimension, and neither architecture has produced interpretable geometric uncertainty maps. We propose a novel 3D Bayesian convolutional neural network (BCNN), the first deep learning method which generates statistically credible geometric uncertainty maps and scales for application to 3D data. We present experimental results on CT scans of graphite electrodes and laser-welded metals and show that our BCNN outperforms an MCDN in recent uncertainty metrics. The geometric uncertainty maps generated by our BCNN capture distributions of sigmoid values that are interpretable as confidence intervals, critical for applications that rely on deep learning for high-consequence decisions. Code available at https://github.com/sandialabs/bcnn.
△ Less
Submitted 1 April, 2020; v1 submitted 23 October, 2019;
originally announced October 2019.
