Machine Learning to Massively Accelerate Nonlinear Design Exploration in Full 3D

Physics-based simulations are often used to drive product design. To extract meaningful information to support design decisions, model surrogates are often used to reduce execution times to allow a high number of parameter evaluations in the design space. Historically, these model surrogates only provide limited information through a few scaler KPIs. To retain comprehensive 3D simulation results while massively reducing execution time, this work presents a neural network approach towards 3D interactive design exploration. 3D multiphysics-multiscale simulations (FEA analyses of structural statics, dynamics, manufacturing, packaging and safety; CFD analyses of compressible fluids) are used in a Design of Experiments (DOEs) to generate the parametric design data. This data is then processed and used to train neural networks as full 3D surrogates for fast and accurate predictions both transient physical responses and full 3D fields.