Driving Engineering Simulation and Design with AI/ML
DescriptionTraditionally, engineered products were designed with mechanical and electrical CAD tools, simulated and validated for correctness with CAE tools, prototypes were fabricated and tested, and products were then manufactured at scale in factories. This process required long product cycles often spanning years to build a new product. Today, virtually unlimited computing and storage available from the cloud is available for generative design to explore 10,000 design choices in near real-time, verify these products accurately through simulation (eliminating the need to build physical prototypes) and manufacture the products using additive manufacturing and factory automation. In the past, simulation tools were used to model specific, solitary physics such as mechanical structures, fluid dynamics, or electromagnetic interactions by solving second order partial differential equations using numerical methods. Today, simulation tools solve multi-physics problems (fluid-structure-electromagnetics interactions) at scale using the most complex solvers. We will explore the use of AI, Machine Learning and Deep Learning to accelerate these engineering simulations. We have identified four broad use cases of AI/ML applied to simulation: (1) Automatic parameter selection of simulation solvers to improve workflows and designer productivity (2) Augmenting simulation with AI/ML to accelerate simulation by factors of 100X (3) The use of AI/ML based generative design techniques to explore 10,000 designs automatically (4) Business intelligence to help improve engineering workflows. My talk will address three broad categories of AI/ML applied to simulation. (1) Top-down methods where we apply an AI/ML framework to a black box solver to train the ML models to improve run time (2) Bottom-up methods where we deeply embed the AI/ML methods inside the physics of our solvers. (3) Reduced order models where the order(?) of the solutions is reduced using AI/ML methods. We will illustrate each of these approaches on existing, commerical tools. As an example of a bottom-up approach, we will describe an ML-based Partial Differential Equation solver and apply it to accelerate Fluid Dynamics problems and will report our results on our Fluent CFD software. As an example of a top-down method, we will report on an ML framework to improve the productivity of any ML developer working in simulation. As an example of a reduced order model we will report on a hybrid digital twin tool called the Twin Builder. We will report on an end-to-end chip packaging solution using a combination of data-driven and physics-informed neural networks, as integrated within Ansys Redhawk/IcePak/Mechanical solutions for Conjugate Heat Transfer. We will describe approaches to support fast design exploration/optimization using ML frameworks. We will describe ML-enabled assistance in various steps of simulation workflows such as initial meshing, smart sub-modeling, user experience and automatic selection of parameters. We will report on automatically setting the best parameters in Fluent/Live AMG solver.
TimeTuesday, July 11th9:00am - 9:15am PDT
Location3020, 3rd Floor