Predicting the Structural Performance of Heat-Treatable Al-Alloys

Abstract

Advances in physically-based and adaptive numeric modelling (ANM) have lead to a significant elevation of the role of modelling in commercial alloy development and process optimisation. Within these two categories of modelling a wide variety of techniques exists, whilst hybrid approaches, taking advantage of the benefits of the both methods, are also being developed. In this paper various ANM techniques and physically-based models relevant for modelling and predicting the properties of heat treatable Al-based alloys are reviewed, and case studies involving the modelling of proof stress, toughness and conductivity of a range of Al-based alloys are presented. These case studies illustrate that successful modelling requires an in-depth understanding of the various modelling options. Selection of the optimum modelling approach is driven by a range of factors, such as size of database available (large databases are best analysed using adaptive numeric modelling approaches), and availability of sound physical understanding (the latter benefiting physically-based approaches). 1.Introduction Throughout the history of alloy development, the modelling of properties and microstructure-property relationships has lagged behind the commercialisation of new products and processes. In terms of agehardened Al-alloys, commercial aerospace applications were established by 1919, however the first basic models of age hardening, based on dislocation movement and dislocation pinning, were not developed until the 1940's [1,2], whilst the first detailed models of strengthening in multi-component Al-based alloys were only published in the 1980/90s [3,4,5]. In recent years, this time lag has been reduced, with modelling preceding or directly leading product development in some cases. A specific example is the formulation of new compositions of Ni-based superalloys by Rolls-Royce Plc. which are based on thermodynamic modelling, and the recent development of 7449 and 7040 Al-alloy plate which has been influenced by microstructure-property modelling [6]. The use of well-founded modelling approaches has clear advantages in directing product development in a cost effective manner.Whilst the above examples concern modelling through physical understanding of microstructures and microstructure development, modelling using adaptive numeric (AN) approaches such as artificial neural networks has recently made important advances. Current state-of-the-art adaptive numeric modelling combines flexible, adaptive algorithms for model construction with a rigorous mathematical/statistical basis [7,8].The purpose of the present paper is to highlight recent advances in physically-based and AN modelling, and to consider the relative benefits of the two approaches as well as hybrid approaches which use elements of both. In exploring this wide ranging and continuous field of modelling approaches we will start by introducing some of the better known techniques. On one side of the spectrum of modelling

A model for the yield strength of overaged Al–Zn–Mg–Cu alloys

“…In relation to the strength of commercial Al based alloys, this assessment thus indicates that after accounting for relatively easily accessible parameters such as average plate compositions, heat treatment temperatures and average M factors of nominally identical alloys, variations in the strength due to other, less accessible sources such as (local) M factor and microsegregation will occur. This view is supported by results of analysis of large databases on yield strengths and processing parameters of nominally identical Al based alloys using adaptive numeric modelling (ANM) which show residual RMSE(test) values of about 1.5% for σ y determined for commercially produced 2024-T351 (Al-Cu-Mg-Mn) [57], similar to the residual deviations of the present model. A further observation supporting this interpretation of the source of deviations is that for the majority of the alloys investigated predicted values were either all higher or all lower than the measured ones.…”

“…(But, adaptive numeric modelling approaches have its own advantages, i.e. they are more 'flexible' in dealing with large and complex data sets where prior physical understanding is very limited [57]. )…”

“…In order evaluate the model and test its accuracy in an objective way we used train and test methods developed originally in the context of adaptive numeric modelling [33,57]. Thus, training and testing was performed by repeatedly splitting the available data in sets for training and testing (about ~50% training data and ~50% test data), and the root of the mean squared error (RMSE) in prediction on the test data ('unseen' data) represents the accuracy of the model.…”

A model for the electrical conductivity of peak-aged and overaged Al-Zn-Mg-Cu alloys

“…According to the regular solution model, the solvus of η′ for η′ particles that are large (i.e. in the absence of any capillary effects [11]) can then be approximated as: (23) This type of simplified approach was explored in an earlier, simpler version of a conductivity model, which led to an accurate description of the conductivity data [46]. In the present, more detailed model, we will take account of coarsening of precipitates and the effect on metastable solubility through the capillary effect (see e.g.…”

Invited review: Adaptive numerical modelling and hybrid physically based ANM approaches in materials engineering – a survey

“…A range of modelling techniques were compared: MLR, Bayesian MLPs, neurofuzzy models and a support vector machine derived approach with additional simple model representation based on the parsimonious ANOVA representation (SUPANOVA). 9 The key approach here was not only to consider a range of modelling approaches, but also to repeatedly sample and resample the train and test data splits (90% : 10%) and examine the multiple model runs for each technique, both in terms of averaged test MSE (effectively adopting a committee approach) but also in terms of underlying relationships revealed. The more complex ANM models outperformed the MLR approach, with the evaluation of predicted trends being considered a key element in assessing the model validity.…”