Predicting Simulation Curves from CAE Data with Machine Learning

Design engineers often look at curves of output data to assess performance.  The ability to predict curves enables more rapid product design.

Altair has had a forward-looking vision to expand the boundaries of computational science in engineering by providing easy to use tools integrated into everyday engineering software.  Training machine learning models from synthetic simulation data will continue to transform product design and improve efficiency by cutting design cycles and ultimately reducing the time it takes to bring a product to market.   Conceptually, our goal is summarized in the diagram below.

Applying data science to simulation data is not novel, but it has attracted unprecedented attention in recent years.  The design and analysis of computer experiments is a subtopic unto itself, with roots in design of experiments and response surface modeling.  Response surfaces themselves are, after all, just another predictive model to predict scalar outputs from design variables; frequently it is simply an ordinary least squares regression.   Several years back, Altair delivered field predictions in Design Explorer to visualize the the behavior across an entire body.  What about the last piece of the puzzle? Predicting curves is also important.  In many cases, engineers interpret curves to judge the merit of a proposed design, for example the forces in a car crash or the resistance of a reclining seat mechanism.  HyperStudy has recently released technology to predict vectors, which will be useful in many engineering applications.

To demonstrate, consider the automotive b-pillar subject to a side impact; a simulation result is shown in the video below.

Multiple simulation results are generated and collected by varying the parametric values. After training a machine learning model, an engineer can predict the amount of deformation or the force of the impact.  The video below captures the real-time feedback on design changes, which aids in making better and faster predictions.

Of course, this simple example is just the start of Altair’s journey into curve predictions.  Our goal is to expand into more advanced use cases supported by ever advancing numerical methods, but always with a focus on powerful and easy to use implementations.  Use the comments below to share some of your thoughts.