Solving inverse problem of material characterization of hyperelastic materials using experimental data and leveraging Bayesian sparse regression and Input Convex Neural Networks

Abstract: Material characterization or constitutive modelling of materials primarily involves being able to estimate the stress-tensor in the material given the strain (or deformation gradient) and a few other history variables (for inelastic material behavior). Classical constitutive modelling involves assuming a form for the stress-strain relationship and using experimental data in conjunction with finite element simulations to calibrate the assumed model. This tends to be limited in the class of materials it is applicable to and is also computationally intensive. On the other hand, data-driven constitutive modelling, specifically the EUCLID framework which will be discussed, are able to leverage recent advances in Machine Learning techniques and hardware to perform efficient and versatile constitutive modelling of materials. The talk will describe how sparse Bayesian regression and Input Convex Neural Networks are integrated into the EUCLID framework to constitutively model hyperelastic materials in a stress-unsupervised manner.

Speaker Bio: Akshay Joshi is a faculty member in the Department of Mechanical Engineering, Indian Institute of Science (IISc), Bangalore. He leads the Dynamics, Data, and Design lab, where he integrates machine learning with dynamic experiments to characterize materials and optimize structural design. His research spans machine learning and data-driven mechanics, high strain-rate and pressure behavior of materials, wave propagation, shock physics, and fracture and fatigue studies.