.png){kind=icon}
.png){kind=icon}
Return
A machine learning bridged concurrent multi-scale computational framework for microstructure related macro-cracking
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jmps.2025.106469.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The concurrent multi-scale methods for microstructure related macro-cracking face challenges in both physical fidelity and computational efficiency. The physical fidelity issue arises from the fact that few models can simultaneously simulate the spatial-temporal evolution of microstructures (e.g. dislocations, multi-phase) and macro-cracking. The computational efficiency issue stems from the mismatch in scale: spatially each grid of macro-simulation corresponds to the whole domain of a micro-simulation and temporally each time step of macro-simulation may encompass many time steps of micro-simulation. This disparity often results in substantial computational expense. In the present work, we significantly accelerate such simulations by developing a machine learning bridged concurrent multi-scale framework for microstructure-related macro-cracking, while preserving main micro-features. First, we establish a phase-field model to simulate the spatial-temporal co-evolution of microstructures under various stress boundary conditions. These simulations generate the data for machine learning models prior to the micro-macro concurrent multi-scale simulations. Subsequently, the well-established machine learning models efficiently provides micro-information to each macro-grid at every time step of macro-cracking, significantly reducing the computational cost. This enables a bidirectional coupling: the macro-cracking behavior is influenced by local microstructures, while the microstructures are continuously updated as macro-cracking progresses. The framework accommodates arbitrary stress-, strain-, and energy-based macro-cracking criteria. We preliminarily validate its accuracy and effectiveness by simulating microstructure-related macro-cracking during 2D high-temperature deformation of film-hole-structured single-crystal superalloys. Under the complex stress states induced by the film holes, the simulated spatial-temporal microstructure evolution and the resulting macro-cracking behavior exhibit good agreement with experimental observations. The present work highlights the possibility of machine learning to accelerate concurrent multi-scale simulations, while maintaining physical fidelity.
Journal
IF:
6
Papers:
5.2K
Citations:
3.0W