Generative Adversarial Networks for Real-Time Realistic Physics Simulations

NASA is seeking to develop real-time realistic nondestructive evaluation (NDE) and structural health monitoring (SHM) physics-based simulations and automated data reduction/analysis methods for large datasets. We propose the combination of a neural network approach with a traditional finite element simulation to generate realistic thermal-based NDE methods for precise determination of structural defects such as cracks, delaminations, and ageing. The proposed approach will allow simulating the structural behavior of complex structures and different types of materials, including any metal alloy and composites. Although the method will be first developed to simulate thermal-based measurements such as thermography, flash thermography, and vibrothermography, the framework could be expanded to other domains including, ultrasonic, microwave, Terahertz, and X-ray. The proposed method has the potential to reduce simulation time by 2 orders of magnitude and an increase the compression rate by 2 orders of magnitude also. Due to the machine learning approach of the method, the accuracy and reliability will increase overtime as the number of validated experimental data increases.