Limitations of ICME and ICMS due to Variability – Alternative Strategies for Alloy Design

Abstract

Integrated Computational Materials Engineering (ICME), and Integrated ComputationalMaterials Science (ICMS) are developing fields with an aim of alloy design, by combining physicalmodels describing materials behavior through lengthscales and processing steps. It has beensuspected, however, that uncertainties in input parameters may cumulate in a hereditary way andyield to a high variability in the final output, independently of the quality of models themselves.Such a variability is however rarely quantified. In this aim, an illustrative example is here given,using a set of “cascade models”, each model being voluntarily very simple (grain growth,precipitation, hardening…) whereas assumed to be exact, so that only the effect of parameteruncertainties on the variability of the output (yield stress of a Ni-base superalloy) can be studied. Itis demonstrated that, with usual uncertainty levels in input parameters, the final dispersion (error)can become very high. Additionally, considering that models are not exact themselves would renderthe situation even worse. Besides, global and implicit models, like neural networks or Gaussianprocesses, have been shown to be able to perform reliable predictions and to be used for alloydesign, with acceptable levels of error, the latter being estimated by statistical methods. In addition,unlike ICME or ICMS, predictions are very fast so that automatic alloy composition optimisation ispossible using, for instance, genetic algorithms. Other fast predictive tools, like computationalthermodynamics (Thermo-Calc), can then be used as constraints during alloy optimisation.