Presentation
Poster 125: Physics Informed Generative Adversarial Networks for Virtual Mechanical Testing
SessionResearch Posters Display
Event Type
Posters
Research Posters
TP
EX
EXH
TimeThursday, 21 November 20198:30am - 5pm
LocationE Concourse
DescriptionPhysics-informed generative adversarial networks (PI-GANs) are used to learn the underlying probability distributions of spatially-varying material properties (e.g., microstructure variability in a polycrystalline material). While standard GANs rely solely on data for training, PI-GANs encode physics in the form of stochastic differential equations using automatic differentiation. The goal here is to show that experimental data from a limited number of material tests can be used with PI-GANs to enable unlimited virtual testing for aerospace applications. Preliminary results using synthetically generated data are provided to demonstrate the proposed framework. Deep learning and automatic differentiation capabilities in Tensorflow were implemented on Nvidia Tesla V100 GPUs.
Archive