A machine learning approach to automate ductile damage parameter selection using finite element simulations

Abstract

Ductile damage models require constitutive model parameter values that are difficult to derive experimentally or analytically. The calibration procedure for ductile damage model parameters, typically performed manually, is labour-intensive. In this work we detail a fully autonomous framework that integrates Bayesian optimisation and finite element analysis to identify ductile damage model parameters. The framework detailed here selects ductile damage model parameters from inputs that can be derived from a simple tensile test. This framework has been successfully deployed to three datasets of P91 material tested at ambient (20°C) and higher (500°C) temperatures. The Bayesian optimisation derived material model parameters result in simulated output with less than 2% error compared to experimental data. This research demonstrates that algorithm hyperparameters can significantly affect the Bayesian optimised ductile damage parameter values resulting in non-unique ductile damage parameters. We show that the non-unique solutions can be further assessed using a second test geometry.