Webinars

Optimization has played an important role in simulation for several decades now. Since that time advances in machine learning have led to the ability to build highly accurate emulators (aka predictive models). These emulators play a key role in Uncertainty Quantification (UQ) as many of the techniques that make up UQ can be too computationally costly to implement directly on the simulation and so the training of a much cheaper to evaluate emulator of the simulation is required in practice.

Using emulators and ideas from UQ, scientists, engineers, and data scientists familiar with optimization can get more value out of their simulations and achieve faster and more reliable optimization results. Examples of the benefits include:

• Accelerate simulation design optimizations by using emulator predictions in place of full fidelity simulation runs at each iteration of the optimization algorithm.
• Some optimization methods such as genetic algorithms require hyperparameters, which themselves are extremely time consuming to optimize. An emulator can be used to speed up the process of discovering the correct hyperparameters to utilize.
• Optimization is often performed deterministically, ignoring the reality of uncertainty in some model inputs, e.g. simulation boundary conditions, material property parameters, and loading conditions. Emulation and UQ allow for optimization that takes input uncertainties into account.
• Move away from purely optimization based model calibration, which only accounts for parameter uncertainty, towards statistical calibration which can also account for model form uncertainty.

This webinar will provide an introduction to machine learning based UQ and emulation and their benefits to optimization. Examples will be used to highlight the additional benefits of such an approach over more basic optimization techniques.