Wednesday  |  Monongahela  |  02:50 PM–03:10 PM

#13542, Multi-Fidelity Machine Learning Based Approach to Predict Local Dynamic Strain Response

Modeling and prediction of the local wave propagation within a material in response to high macroscale impact forces can provide key information for material performance. This work aims to fill current knowledge gap in modeling local high strain dynamic response of a material when impacted by a large impactor using machine learning techniques. A PMMA (polymethyl methacrylate) sample is subjected to flat nail impact at sub-mm scale on one end and macroscale wave propagation response is recorded using force sensitivity resistors at the other end. Nail impact is used to represent local dynamic material behavior. FSR measurement is used to represent macroscale average dynamic material behavior. Using a combination of multi fidelity machine learning with finite element method (FEM) and digital image correlation (DIC), a multifidelity numerical framework is laid out to related local material response with macroscale material response. Setup parameters, sensor data, and FEM simulations were used to build the training data. Composite neural-net architectures were implemented to model and predict the strain profile (DIC data) of PMMA during dynamic impact at varying levels of fidelity. Comparisons are then drawn between the low, high, and multi fidelity predictions and measured strain field data. Ultimately the training is used to predict local material response from macroscale FSR measurements.

Tyler Dillard   Purdue University
Sushrut Karmarkar   Purdue University
Ayush Rai   Purdue
Vikas Tomar   Purdue University