Enhanced Constitutive Creep-Fatigue Model for Industry Using Neural Networks-Based Parameter Identification and Implementation in FE-Code

Abstract Advanced constitutive material models, and specifically creep-fatigue constitutive models as well as corresponding subroutines for various CAE products have been proposed for years by many research groups. However, the need for development of mechanisms-based models for each material individually, limited availability of experimental data and time-consuming identification procedures restrict their applications in the industry. This work focuses on revising a mechanism-based constitutive model for 9–12% Cr heat-resistant steels, identifying material parameters using smoothed experimental data through artificial neural networks (NNs), and developing a user material subroutine for Simcenter NX Nastran software. Work involving NNs is carried out in a Python environment using the machine learning libraries Keras and TensorFlow. The constitutive model is implemented in an NXUMAT subroutine. Multiple calculation tests are run for different stresses and temperatures, and the results are compared to the experimental data. To test the subroutine’s ability to work under variable mechanical and thermal loads and handle complex geometries, a few LCF and TMF tests are conducted. The steam turbine rotor and inner casing are analyzed to evaluate the behavior of real components. Examples of benchmark tests and analyses of power plant components are presented.