Scholarlink

<p>This study proposes a hybrid model based on ResNet34-MLP cross-modal fusion for predicting the effective elastic properties of unidirectional composites. Using the Random Sequential Expansion algorithm, 1,200 representative volume element (RVE) samples with varying fiber volume fractions (30%, 40%, 50%, 60%) were generated, with 300 samples per fraction. These were divided into training, validation, and test sets at an 8:1:1 ratio. An additional three independent sets of 100 RVEs with fiber volume fractions of 35%, 45%, and 55% were generated exclusively for generalization evaluation (excluded from training). The equivalent elastic parameters of all samples were obtained via ABAQUS finite element simulations. The model processes dual-modal inputs: (1) RVE microstructure images and (2) intrinsic material parameters of fibers/matrix. To address initial performance limitations, three optimization strategies were implemented: Physics-constrained loss function, Cross-modal attention mechanism and Image preprocessing. Through ablation studies on eight model variants, two configurations achieved comparable performance: using only the physics-constrained loss function and using all three strategies together. The former achieved a mean relative error (MRE) of 3.03% on the test set, while the latter achieved 3.54%. On the independent validation sets (35%, 45%, and 55% volume fractions), the MREs were 6.58% and 6.73%, respectively.</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

Fusion of microstructural images and constituent properties for elastic property prediction in unidirectional composites: a hybrid ResNet34-MLP approach

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Tao Peng,
  • Tongyu Liu,
  • Junping Liu,
  • Xinrong Hu,
  • Li Li,
  • Zili Zhang

摘要

This study proposes a hybrid model based on ResNet34-MLP cross-modal fusion for predicting the effective elastic properties of unidirectional composites. Using the Random Sequential Expansion algorithm, 1,200 representative volume element (RVE) samples with varying fiber volume fractions (30%, 40%, 50%, 60%) were generated, with 300 samples per fraction. These were divided into training, validation, and test sets at an 8:1:1 ratio. An additional three independent sets of 100 RVEs with fiber volume fractions of 35%, 45%, and 55% were generated exclusively for generalization evaluation (excluded from training). The equivalent elastic parameters of all samples were obtained via ABAQUS finite element simulations. The model processes dual-modal inputs: (1) RVE microstructure images and (2) intrinsic material parameters of fibers/matrix. To address initial performance limitations, three optimization strategies were implemented: Physics-constrained loss function, Cross-modal attention mechanism and Image preprocessing. Through ablation studies on eight model variants, two configurations achieved comparable performance: using only the physics-constrained loss function and using all three strategies together. The former achieved a mean relative error (MRE) of 3.03% on the test set, while the latter achieved 3.54%. On the independent validation sets (35%, 45%, and 55% volume fractions), the MREs were 6.58% and 6.73%, respectively.