How to Enable Complex Simulations: The Power of Multiphysics & Digital Thread Seminar

Machine Learning for Time Consuming or Complex Simulations

Author: Gavin Jones - Smart UQ

Abstract

Simulations can often have long run times from hours to many days for a single case. This can be the result of factors such as the complexity of the system to be modeled, the type of simulation, and the level of fidelity. Simulations with long run times present several challenges to their use. When computational cost limits the number of simulation runs that can be performed, direct approaches to optimization and uncertainty quantification (UQ) become infeasible and the number of inputs, scenarios, and design possibilities that can be explored are severely limited. Surrogate modeling, whereby a machine learning (ML) model is trained to predict the results of a simulation, has become a popular approach to address these challenges. However, often long run times, possibly combined with complex response behavior (e.g. highly non-linear, spatially and/or temporally distributed) or large numbers of model inputs makes collecting enough data to train an accurate surrogate model its own challenge. A further challenge concerns the time involved in training models for such cases.

This will include:

• Adaptive Design of Experiment (DOE) techniques for efficient data collection, i.e. to minimize the number of simulation runs needed.
• Efficient ML models for addressing the challenges of many inputs or many correlated outputs, e.g. a ML model trained to predict CFD results representing a time varying field of temperature values.
• ML models for multi-fidelity modeling approaches, allowing training sets for the most complex expensive to run simulations to be augmented by data from lower fidelity, cheaper to run models.

Sponsor

Sponsorships

Get involved and engage with a diverse, enthusiastic audience by becoming a valued sponsor of our event.

W​e have a number of excellent sponsorship and exhibition opportunities available at this event.

Click here to explore your options

Contact lisa.rimback@nafems.org for more information

Back to Event Page