EUVIMEDDescubrir EUVIMED
Crossref

A mechanics-informed deep learning constitutive model for sequential prediction of strain rate-dependent behavior and microstructural evolution

Wei Dai, Xiaochuan Sun, Kecheng Zhou, Yongkai Wang, Huamiao Wang, Yue Liu

Microstructures · 2026-06-12 · 6 (4)

DOI: 10.20517/microstructures.2026.17 ↗

At a glance

Journal article in Microstructures, published online on 2026-06-12. Written by Wei Dai and 5 co-authors. 49 references.

ResumenDatos bibliográficosCiteRelated studies

Abstract (original language)

Classical constitutive models explicitly couple macroscopic mechanical responses with underlying microstructural evolution, which is crucial for capturing complex deformation mechanisms across varying strain rates. However, current deep learning (DL) constitutive models predominantly focus on macroscopic stress-strain mapping, often neglecting these critical microstructural transitions. To bridge this gap, this work proposes a mechanics-informed deep learning constitutive model (MIDLCM) that integrates gated recurrent units and multi-head attention with a mechanics-informed layer and a mechanics-informed loss, enabling simultaneous prediction of stress response and microstructural descriptors. Trained on a CrFeNi FCC alloy dataset spanning strain rates from 10-4 to 5,000 s-1, MIDLCM accurately reproduces strain-rate-dependent stress-strain behavior and captures the associated evolution of dislocation density and twin volume fraction. Crucially, the model successfully represents the distinct dislocation accumulation regimes and the dynamic transition of plasticity mechanisms - from dislocation-dominated to twinning-assisted - across extreme dynamic loading, consistent with experimental trends and crystal-plasticity-based references. Ablation studies show that attention-based temporal encoding and mechanics-informed constraints contribute complementary improvements while preserving inference efficiency. By explicitly tracking these internal state variables, the proposed framework provides a mechanism-level interpretable and computationally efficient microstructure-mechanics coupled alternative for rate-dependent constitutive modeling and is readily extendable to other alloy systems and loading paths.

Cite

APA

Dai, W., Sun, X., Zhou, K., Wang, Y., Wang, H., & Liu, Y. (2026). A mechanics-informed deep learning constitutive model for sequential prediction of strain rate-dependent behavior and microstructural evolution. Microstructures, 6(4). https://doi.org/10.20517/microstructures.2026.17

Vancouver

Dai W, Sun X, Zhou K, Wang Y, Wang H, Liu Y. A mechanics-informed deep learning constitutive model for sequential prediction of strain rate-dependent behavior and microstructural evolution. Microstructures. 2026;6(4). doi:10.20517/microstructures.2026.17

BibTeX
@article{Dai2026,
  author = {Dai, Wei and Sun, Xiaochuan and Zhou, Kecheng and Wang, Yongkai and Wang, Huamiao and Liu, Yue},
  title = {A mechanics-informed deep learning constitutive model for sequential prediction of strain rate-dependent behavior and microstructural evolution},
  journal = {Microstructures},
  year = {2026},
  volume = {6},
  number = {4},
  doi = {10.20517/microstructures.2026.17}
}
Start for free →