This presentation was made at the NAFEMS Americas Seminar - Confidence in Engineering Simulation: The Next 10 Years of CAE in Mexico.

What is the future for engineering analysis and simulation in Mexico? Discover innovative engineering simulation processes and tools which are helping companies in Mexico improve production capabilities. Engage with domain experts, industry leaders, and peers in a focused, comprehensive one-day event that covers topics on engineering analysis, simulation, and systems modeling and simulation that every engineer in Mexico should know.

Resource Abstract

Due to the complexity of simulating multiple FEA configurations and limited resources, we began an exercise using a statistical methodology to identify physical test approach based on virtual crash testing for Energy Absorber design.



The complexity of the design of this component lies in the interaction of multiple parts that are responsible for the dissipation of energy. In addition to this, the system must meet the opposite tests in terms of rigidity.



Initially, the baseline model was correlated against physical testing then, a DOE with several inputs for the base model was performed providing variability in results, and those were used as raw data. Monitoring all outputs allow us to use dimensionality reduction for the inputs to improve the performance of the data mining.



Being this the starting point, we utilized various techniques applied for data miner commonly applied for linear regressions, neuronal networks, and decision trees.



Approaches combine several machine learning techniques into the predictive model for decrease the variance (bagging), bias (boosting) to improve the prediction, however, the results were not as expected due the nature of the data (continuous) and inaccurate mining algorithm.



Stepping back, using linear regression, neural networks, and decision trees mining algorithms with raw data, the prediction model was built and lead us to obtain a mean predicted which has a statistical representation around 90% of fidelity by applying using three inputs and one output at each time. This prediction was verified with the prediction model.



Being these results as a starting point we need to add new specialized learning systems with different environments to consider more design variables, for example, materials, geometric changes, among others



This whole study is focused on simulations of impact at low speed but considering the potential of statistical algorithms and building a broad-spectrum predictive model can be used to support other areas focused on simulation.