Bridging Length Scales Efficiently Through Surrogate Modelling

Safety sensitive industries are increasingly facing challenges such as reducing their environmental impact and bringing down their cost. Part of the solution to such challenges is economical use of assets while maintaining the safety level if not increasing it. Thus, more informed and reliable decision making on repairing or replacing key components is becoming even more important where a simple binary safe/unsafe choice is no longer desirable. Instead, a realistic assessment which inevitably would be probabilistic is needed. However, obtaining the required level of data to suitably underpin a probabilistic assessment can be prohibitively expensive as carrying out hundreds if not thousands of full-scale tests is no longer economically possible. In this work, we explore an alternative approach in which micromechanical characterisations, which due to their small scale, are more affordable, are carried out and informed a meso-scale model of the material behaviour. The meso-scale simulation, that is a crystal plasticity finite element model, is informed by the variations within the material microstructure thus returning a representative material response. The model variation can be estimated by machine learning algorithm such as polynomial chaos expansion thus returning material response variability in a sensible time-scale. The material variability, in turn, is input into a surrogate model of a process modelling, in our case welding simulation, to produce variability in a parameter important for assessment such as weld residual stress.