Scholarlink

<p>The spatial arrangement of voids plays a critical role in determining metallic ductility, yet currently most theoretical models remain constrained to single void approximations. In this study, we propose an interpretable machine learning (ML) framework integrating a periodic convolutional neural network (PCNN) that directly maps heterogeneous porosity distributions to macroscopic fracture strains. Training datasets are generated through finite element simulations implementing a shear-modified Gurson-Tvergaard-Needleman model. By embedding periodicity constraints via three key components—circular padding, learnable polyphase downsampling, and permutation-invariant layer, the PCNN architecture achieves superior performance compared to conventional CNNs, realizing higher prediction accuracy with superior data efficiency. Moreover, gradient-based interpretability analysis reveals the capability of the model to identify strain localization patterns, providing insight into failure mechanism: ductility is governed not only by local porosity within void coalescence bands but predominantly by the damage disparity between the bands and adjacent regions. This work establishes an interpretable ML framework that directly links microstructural heterogeneity to macroscopic fracture, providing both predictive accuracy improvement and mechanism understanding advancement in ductile failures.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Ductile fracture prediction of spatially heterogeneous porous material by periodic convolutional neural network

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Liujun Wu,
  • Ce Chen,
  • Jiaqi Fu,
  • Chenyang Xin,
  • Wenbin Liu,
  • Pei Wang,
  • Huiling Duan

摘要

The spatial arrangement of voids plays a critical role in determining metallic ductility, yet currently most theoretical models remain constrained to single void approximations. In this study, we propose an interpretable machine learning (ML) framework integrating a periodic convolutional neural network (PCNN) that directly maps heterogeneous porosity distributions to macroscopic fracture strains. Training datasets are generated through finite element simulations implementing a shear-modified Gurson-Tvergaard-Needleman model. By embedding periodicity constraints via three key components—circular padding, learnable polyphase downsampling, and permutation-invariant layer, the PCNN architecture achieves superior performance compared to conventional CNNs, realizing higher prediction accuracy with superior data efficiency. Moreover, gradient-based interpretability analysis reveals the capability of the model to identify strain localization patterns, providing insight into failure mechanism: ductility is governed not only by local porosity within void coalescence bands but predominantly by the damage disparity between the bands and adjacent regions. This work establishes an interpretable ML framework that directly links microstructural heterogeneity to macroscopic fracture, providing both predictive accuracy improvement and mechanism understanding advancement in ductile failures.