Integrating AI for Data Generation and Surrogate Model Training
AI-assisted multiphysics models are only as good as the data used to train them, but high-quality datasets are often scarce, expensive, or difficult to obtain through experimentation. The most efficient way to acquire training data is actually to create it synthetically through simulation. In this archived webinar, COMSOL's Bjorn Sjodin explains how the COMSOL Multiphysics® software can facilitate the creation of high-fidelity surrogate models through the generation of synthetic training data and the design, training, and evaluation of the models themselves.
The webinar begins with an overview of the COMSOL Multiphysics® software's capabilities for AI integration, followed by an explanation of the process to train deep neural network (DNN) surrogate models. After this, Sjodin demonstrates how datasets can be imported from third-party tools or from tabulated experimental data. Then, he presents computationally efficient methods for training models, such as Latin hypercube sampling (LHS), cluster computing, and selective error handling.
From here, Sjodin discusses how the software can be used to evaluate completed surrogate models across multiple areas of physics. The webinar concludes with a demonstration of a Gaussian process surrogate model.
Tip: View this Learning Center course on how to train and utilize surrogate models in the COMSOL® software.
