Skip to main navigation Skip to search Skip to main content

Implicit integration of a Mixture of Experts neural network surrogate elasto-viscoplastic constitutive model of HT9 steel in a finite element framework: Verification, comparative assessment, and validation

Research output: Contribution to journal › Article › peer-review

Abstract

Microstructure-informed and mechanism-based constitutive models that track evolving internal state variables inherently possess an increased ability to be used for extrapolation over traditional empirical descriptions of high-temperature creep. However, their consistent and efficient integration within finite element (FE) frameworks remains a non-trivial algorithmic challenge. This work presents a fully implicit FE integration framework for a Mixture-of-Experts (MoE) machine-learning surrogate describing the elasto-viscoplastic response of HT9 ferritic–martensitic steel under combined thermal and irradiation-assisted creep conditions. Dislocation cell and wall densities are treated as microstructural internal state variables governing creep kinetics; the accuracy of the MoE surrogate relative to the underlying viscoplastic self-consistent (VPSC) model is documented in a companion study. Two radial-return algorithms are developed: a univariate scheme (UNR-RR), which solves implicitly only for the plastic multiplier while updating dislocation densities explicitly, and a multivariate scheme (MNR-RR), which simultaneously solves for the plastic multiplier and dislocation density increments within a single Newton iteration. Both schemes are implemented in the MOOSE FE framework via the NEML2 material library, enabling batched quadrature-point evaluation, efficient vectorization with optional GPU acceleration, and automatic differentiation for consistent algorithmic Jacobians, offering a scalable alternative to MOOSE's native automatic differentiation. Verification studies demonstrate excellent agreement with reference solutions within the reliable operating domain of the surrogate model. The multivariate MNR-RR scheme exhibits substantially improved accuracy and efficiency relative to UNR-RR, enabling larger time steps and reduced computational cost, with increasing advantage for long-term creep simulations. The framework is further validated against thermal creep experiments on pressurized HT9 cladding tubes, accurately reproducing both macroscopic deformation and underlying microstructural trends. These results establish a robust framework for integrating machine-learning-based microstructure-informed constitutive models into FE simulations and support future extensions to broader operating regimes.

Original languageEnglish
Article number104623
JournalFinite Elements in Analysis and Design
Volume261
Early online dateAug 20 2026
DOIs
StatePublished - Oct 2026

Keywords

  • HT9 steel
  • Implicit return mapping
  • Mixture-of-Experts surrogate
  • MOOSE
  • NEML2
  • Thermal and irradiation creep

INL Publication Number

  • INL/JOU-26-90120
  • 212153

Fingerprint

Dive into the research topics of 'Implicit integration of a Mixture of Experts neural network surrogate elasto-viscoplastic constitutive model of HT9 steel in a finite element framework: Verification, comparative assessment, and validation'. Together they form a unique fingerprint.

Cite this