Multi-Fidelity ML Based Approach to Predict Local Material Response

Ayush Rai, Purdue University

Abstract

High strain dynamic response of materials is of key importance for material development. An attempt was made to study the local material behavior of PMMA (polymethyl methacrylate) using a flat nail impact. The developed experiment aims to study the local material behavior by measuring the local strain and deformation response, and bulk load response. The material behavior is studied, and the prediction is made about the local material response with machine learning and predictive modeling techniques using training data set from the developed microscale nail impact experiment. PMMA samples were impacted with flat nails at high velocities using an in-house built gas gun. Sensors (force sensitivity resistors) are used to measure the material load response of examined samples. Digital image correlation (DIC) was used with a high-speed camera to record the local surface deformation and strains. The aim of this work holds to understand the wave propagation in the material and to predict the material behavior of the sample under the point of impact using the sensor data. Experiments and its data processing here were expensive in terms of time and complexity and hence concurrently, a FEM simulation was developed to mimic the experiment. A large amount of FEM data was generated to create the dataset for a low-fidelity model. The experiment (and simulation) parameters and the sensor data (force profile on the back surface of the sample in case of the simulation) are used to build the training data. The model predicts the strain profile (DIC data) of the sample at the point just below the impact. Composite neural-net architecture is implemented to model to multi-fidelity approach and comparisons were made between high-fidelity and multi-fidelity prediction.

Degree

M.Sc.

Advisors

Tomar, Purdue University.

Subject Area

Mechanics|Polymer chemistry

COinS