Application of neural networks in combination with complex thermodynamics for the prediction of creep rupture strength of 9-12 % Cr steels

Abstract

The multilayer perceptron type of neural network (MLP) was used to investigate the influence of chemical composition, heat treatment and thermodynamically stable phases on the creep rupture strength of 9-12 % Cr-steels. The model is based on extensive sets of data from literature. Additionally, methods for the minimization of Gibb's energy in complex multi-component systems are employed for thermodynamic generation of features. Chemical composition and operational temperature of the material serve to determine the complex state of equilibrium. This information supplies additional input parameters for the neural network. The created model allows precise prediction and adjustment of the long-term behaviour of 9-12 % Cr-steels by means of application of optimal heat treatment and material composition. In addition, relationships and dependencies between thermodynamically determined states and the creep rupture behaviour of ferritic 9-12 % Cr-steels are characterised. One aim of the model developed on the basis of neural networks in combination with the method minimizing the Gibb's energy in complex systems is to make the prediction of creep rupture behaviour possible and reproducible within the available data volume. Processes occurring near thermodynamic equilibrium can be described sufficiently by means of classical thermodynamics. Problems occurring with prediction of states and processes are caused by a more or less considerable deviation from the state of equilibrium. An indirect consideration of these deviations from the state of equilibrium can be achieved by joining complex thermodynamic approaches with neural networks. This allows the description of metastable and non-eqilibrium phases, i.e. dynamic systems. This includes modelling of time-dependent processes in materials, the creep rupture strength of 9-12 % Cr-steels being determined on the basis of experimental and thermodynamic data. The main target of the symbiosis of neural networks and complex thermodynamics is to combine the advantages of these methods, to compensate their deficiencies and to predict the creep rupture behaviour of materials. steel research 72 (2001) NO.9 observations '-, :L J . -_ 207 steel research 72 (2001) NO.9

Neural Networks and Information in Materials Science

“…Examples include work in the context of creep deformation [38,64] and phase transformation theory [65,66].…”

“…Attempts have also been made to incorporate the outputs of thermodynamic models as additional data in the creation of neural networks on the basis of experimental datasets. Examples include work in the context of creep deformation 38, 64 and phase transformation theory 65, 66.…”

Mathematical models in materials science

“…Thus, after publication, Brun et al 66 made some 2700 sets of creep data, each with 32 variables available on MAP. 18 These were then used by Meyer and Gutte 19 to create a hybrid creep model which includes the thermodynamics of such steels.…”

“…It is not surprising therefore that numerous attempts have been made to use neural network modelling to deal with complex properties such as fatigue, [56][57][58][59] toughness, [60][61][62][63] corrosion resistance 64 and creep rupture life. 19,[65][66][67][68] A neural network is a general method of regression analysis. [69][70][71][72] A few of the advantages of the network over conventional regression can be listed as follows:…”