This presentation was held at the 2020 NAFEMS UK Conference "Inspiring Innovation through Engineering Simulation". The conference covered topics ranging from traditional FEA and CFD, to new and emerging areas including artificial intelligence, machine learning and EDA.

Resource Abstract

Weld residual stresses play an important role in structural integrity assessments of engineering structures. In particular, they can reduce both the fatigue life of structures containing defects and increase the propensity of these structures to brittle failure. In a safety concious industry such as nuclear, the use of conservative assessment methodologies with residual stresses can lead to difficulties in justifying continued safe operations of aging infrastructure. Hence, the ability to quantify weld residual stresses both accurately and reliably can be important in ensuring the economic sustainability of engineering assets.

For some common materials, such as ferritic steels, the metallurgical behaviour has a direct relation on the residual stresses generated during welding. Finite element based computational methods have been shown to be able to accurately predict the quantity of residual stresses even for ferritic steels. It must be noted that these methods rely on accurate metallurgical models capturing sold state phase transformations. Further, these methods rely on many input quantities that are subject to sources of measurement uncertainty, such as alloy chemical composition. Thus, the reliability of predictions is not easily quantifiable, particularly because computational methods are expensive and time consuming to deploy.

In this study, a method for accurate model reduction of weld simulations, based on proper orthagonal decomposition and Gaussian process regression is demonstrated. The model reduction technique is used to enable a systematic uncertainty quantification study on weld residual stress models. Specifically, the sensitivity of residual stresses to the natural variation in material properties of ferritic steels is evaluated. Sensitivity is quantified from Sobol indices based on the Saltelli algorithm.

The study implies that state-of-the art residual stress prediction methods can be applied to ferritic welds with confidence. Furthermore, the use of statistical techniques using open source toolkits, such as openTURNS, demonstrates that powerful uncertainty quantification techniques can be deployed to support structural integrity assessments of welded structures at low cost to business. Such tools can be used to accelerate the acceptance of methods that have been hitherto limited to research and development.