Session: 21-06 Design Methods

Submission Number: 176783

Workflow-Based Software Applications That Employ Ai-Driven Assessment to Develop, Manage and Deploy Material Models for Industrial Use 

Germany possesses a unique pool of experimental data and long-standing expertise in advanced visco-plastic and creep-fatigue material modelling. Yet model formulations, data formats and software implementations still differ noticeably from one research group or industrial partner to another, impeding the seamless transfer of knowledge into engineering practice. The BMFTR funded “DigitalModeling” project, launched in early 2024, therefore pursues a unifying standard and open interface that will (i) describe constitutive equation systems in a solver-independent way, (ii) enable objective, largely automated parameter identification, and (iii) maintain real-time-capable, application-specific material models for computer-aided engineering (CAE).

As a showcase for this strategy, we present a workflow for setting up a mechanism-based creep-fatigue model for heat-resistant steels that is tailored to a given load case. Model parameters are extracted directly from heterogeneous raw experiments by means of artificial neural networks implemented in a Python script, thereby illustrating the envisaged automated, data-driven identification process. In case of incomplete experimental data, data mining models based on artificial neural networks enable synthetical generation of batch-specific material data on strength behavior and creep behavior. The calibrated model is transferred to commercial finite-element environments via a user subroutine; users compile the user material subroutine with or without additional libraries. By a wrapper the subroutine can be offered for different commercial finite element programs. Verification and validation include isothermal and thermo-mechanical fatigue tests over a wide stress/temperature range, a benchmark test and full-scale analyses of an industrial component. A pyiron-workflow is provided to integrate the entire process of material parameter identification into a GUI environment. In the GUI, each step of the process is represented by corresponding nodes, enabling users to implement the problem more conveniently, and ensuring high levels of reproducibility and transparency.

Furthermore, the calibrated material cards (parameter sets) are integrated into an ontology-based knowledge graph and made accessible via a Digital Product Passport (DPP) interface. This ensures end-to-end provenance (from raw experiments and data preprocessing through neural-network-based parameter identification, verification/validation, and deployment), enabling users to trace the entire calibration history at a glance. This reinforces transparency, trustworthiness, and reusability across partners.

By coupling a harmonized, solver-independent description of constitutive behavior with machine-learning-based parameter fitting and ready-to-use CAE subroutines, this work exemplifies how “DigitalModeling” will streamline the development, dissemination and industrial deployment of state-of-the-art visco-plastic material models.

Keywords: steam turbine, visco-plastic material models, creep, fatigue, UMAT, neural networks, FEA, simulation workflows, ontologies

Presenting Author: Yevgen Kostenko Siemens Energy Global GmbH & Co. KG

Presenting Author Biography: Advisory Engineer at Siemens Energy Global GmbH & Co. KG for mechanical integrity.

Authors:

Yevgen Kostenko Siemens Energy Global GmbH & Co. KG
Konstantin Naumenko Otto-von-Guericke University Magdeburg
Alexander Jahnke Otto-von-Guericke University Magdeburg
Felix Koelzow MPA and Institute of Materials Technology TU Darmstadt
Ehsan Borzabadi Farahani Federal Institute for Materials Research and Testing Berlin
Bernard Fedelich Federal Institute for Materials Research and Testing Berlin
Elena Garcia Trelles Fraunhofer Institute for Mechanics of Materials Freiburg
Min Huang MPA and University of Stuttgart
Christoph Schweizer Fraunhofer Institute for Mechanics of Materials Freiburg