Predicting Fast Crack Propagation in Welded Steel Plates Using Machine Learning
- Halid Can YILDIRIM
- Aug 8
- 2 min read
Updated: Aug 14
A hybrid computational framework integrating finite element analysis and recurrent neural networks for fracture prediction.
Overview
This repository presents an innovative framework combining Finite Element Analysis (FEA) and Recurrent Neural Networks (RNNs) to predict fast crack propagation in welded steel plates with random flaws. The methodology addresses material inhomogeneity, residual stresses, and mixed-mode loading conditions, offering a data-driven approach to enhance structural integrity assessments.
By leveraging Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models, this framework achieves high-accuracy predictions of stress redistribution and crack growth, outperforming traditional analytical methods. The integration of SHAP analysis further quantifies the influence of key parameters, supporting predictive maintenance and weld design optimisation.
Key Challenges Addressed
Material variability: Random flaws and residual stresses in welded joints complicate crack path prediction.
Computational cost: High-fidelity FEA simulations are resource-intensive; machine learning reduces runtime by 40%.
Mixed-mode loading: Traditional models struggle with complex stress interactions; RNNs capture dynamic crack behaviour.
Key Features
Hybrid Physics-Informed ML: Combines FEA with LSTM/GRU networks for accurate, efficient predictions.
SHAP-Driven Interpretability: Quantifies feature importance (e.g., displacements account for 75% of stress variance).
Residual Stress Integration: Incorporates neutron diffraction data to improve crack path accuracy in heat-affected zones.
Scalable Dataset: Includes 30,000 simulated crack scenarios for benchmarking. The welded joint consists of four distinct material layers or zones:
• Base Material (Parent Material): The original steel plates being joined through welding.
• Weld Material: This layer consists of the weld metal, which has been melted and subsequently
solidified during the welding process.
• Heat-Affected Zone (HAZ): This area is not melted during welding but undergoes changes in
microstructure and properties due to the thermal effects of the welding process.
• High-Frequency Mechanical Impact (HFMI) Material: This layer represents the enhanced
material properties resulting from HFMI treatment.
Repository Structure
/data/: FEA simulation results, experimental crack propagation data, and residual stress measurements.
/models/: Pretrained LSTM/GRU architectures and SHAP analysis scripts.
/results/: Performance metrics, stress-field predictions, and comparative analyses.
Usage
Preprocess FEA data using scripts in /data/.
Train/evaluate RNN models from /models/ for crack growth prediction.
Interpret results via SHAP analysis and visualise stress/displacement fields.
Applications
Predictive maintenance: Early detection of critical cracks in offshore/platform welds.
Weld design optimisation: Evaluate flaw tolerance during structural planning.
Industrial standards: Data-driven acceptance criteria for weld flaws.
Contributions
Contributions to model optimisation, experimental validation, or industrial case studies are welcome. Contact for collaboration.
Citation
If this work supports your research, please cite the original paper: Yıldırım, H.C. (2025). "Predicting Fast Crack Propagation in Welded Steel Plates with Random Flaws Using Recurrent Neural Networks." Advanced Engineering Informatics.[https://doi.org/10.1016/j.aei.2025.103671]. @article{YILDIRIM2025103671,
title = {Predicting fast crack propagation in welded steel plates with random flaws using Recurrent Neural Networks},
journal = {Advanced Engineering Informatics},
volume = {68},
pages = {103671},
year = {2025},
issn = {1474-0346},
doi = {https://doi.org/10.1016/j.aei.2025.103671},
url = {https://www.sciencedirect.com/science/article/pii/S1474034625005646},
author = {Halid Can Yıldırım},
keywords = {Crack propagation, Welded steel plates, Weld improvement, Neutron diffraction, Finite element analysis (FEA), Recurrent neural networks (RNN)}
}
Comments