Scholarlink

<p>This study proposes a hybrid physics-guided neural network framework for estimating forming limit in incremental sheet forming. A two-stage surrogate modeling framework has been introduced. At first stage, the network predicts the principal strain state at fracture while enforcing plastic incompressibility through thickness-based regularization. The network at stage II estimates fracture depth by incorporating physically motivated constraints on process-paramete'r monotonicity and damage accumulation. A comprehensive dataset was generated using validated Abaqus/Explicit simulations that incorporated a ductile damage model and systematically varied tool diameter, wall angle, and step depth. Baseline deep neural networks were developed for comparison and subsequently upgraded to physics-guided formulations. The results show that the proposed models improve training stability and reduce prediction scatter relative to purely data-driven networks. The model predicts in-plane major and minor strains and the fracture depth with less than 7% error compared to the finite element analysis.</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

Hybrid Physics-Guided Neural Network Model for Forming Limit Estimation in Incremental Sheet Forming

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Abhay Kumar Dubey,
  • Stuti Govil,
  • Harshal Y. Shahare,
  • R. Seetharam,
  • Puneet Tandon

摘要

This study proposes a hybrid physics-guided neural network framework for estimating forming limit in incremental sheet forming. A two-stage surrogate modeling framework has been introduced. At first stage, the network predicts the principal strain state at fracture while enforcing plastic incompressibility through thickness-based regularization. The network at stage II estimates fracture depth by incorporating physically motivated constraints on process-paramete'r monotonicity and damage accumulation. A comprehensive dataset was generated using validated Abaqus/Explicit simulations that incorporated a ductile damage model and systematically varied tool diameter, wall angle, and step depth. Baseline deep neural networks were developed for comparison and subsequently upgraded to physics-guided formulations. The results show that the proposed models improve training stability and reduce prediction scatter relative to purely data-driven networks. The model predicts in-plane major and minor strains and the fracture depth with less than 7% error compared to the finite element analysis.