Training

Introduction to Surrogate Modeling for Engineering Simulation

Simulation plays an indispensable role in engineering activities to achieve objectives such as rapid prototyping, optimization of designs and processes, and the analysis of complex systems without the need for costly physical testing. However, the effective use of simulation still presents several challenges including:

• Long run times which can render statistical or iterative tasks like optimization, sensitivity analysis, and uncertainty quantification infeasible.
• Inaccuracy due to the presence of uncertainties arising from sources including initial conditions, model parameters, model form, manufacturing, and operating environment.
• Difficulty in calibrating to real-world data, including test results and sensor readings.

Surrogate modeling is an approach which can help address all the above. With surrogate modeling, first a machine learning (ML) model is trained to predict the simulation’s results. The ML model is then used in place of the simulation to run the desired analyses. The ML model’s rapid prediction of simulation results allows more analyses to be performed and in less time.

The course provides an introduction to surrogate modeling with topics including:

• Use of design of experiments (DOEs) for optimal training data collection, i.e. to decide which simulations to run.
• How to train and validate the accuracy of surrogate models
• Use of surrogate models for sensitivity analysis, uncertainty propagation, and optimization under uncertainty to maximize the knowledge gained from simulation and understand uncertainties
• Statistical calibration for addressing parameter and model form uncertainty in calibrating simulations and surrogate models to physical data.

Points will be illustrated with SmartUQ customer use cases and example problems demonstrated in SmartUQ. Use cases and examples can be tailored to be relevant to the needs and industry of the attendees.

Request More Information