S. Guerin, Z. Rogers
Additive Rocket Corporation,
United States
Keywords: system design, automation, simulation, deep learning
Summary:
Designing multistage physical systems involves modeling individual subsystems and then aggregating the subsystems to model the complete system. For systems with complex phenomena where analytical solutions are not readily available, multi-physics simulation software such as ANSYS is leveraged to provide engineers and designers accurate models of physical phenomena. The model accuracy comes at the cost of time or computational expense, as it often takes hours to conduct a simulation for each design element. For systems with consistent architectures and a high demand for customization, we propose an alternative method to the current simulation and design framework through the use of deep learning. By running many simulations that span the design space for each subsystem, a corpus of simulation data can then be used to train a neural network to model the subsystem. This Neural Simulation Model (NSM) provides results within 5 – 10% of multi-physics simulation results and several orders of magnitude improvement with regard to speed. Because of this rapid simulation capability, design optimization can be conducted in minutes instead of weeks, creating a paradigm shift in cost and lead time associated with custom design.