Scholarlink

<p>Although recent advances in deep generative models have transformed product development across manufacturing industries, their application in design evaluation remains limited due to the challenges of ensuring fidelity and reliability required by strict industrial standards. We introduce ProtoDiff, a framework for high-fidelity pseudo-virtual prototyping and design evaluation, demonstrated through the task of generating tire footprint images - a critical element in tire performance analysis. Built upon diffusion-based architectures, the framework integrates three core components: a dual-stream embedding method for processing complex structured tabular data by separating numerical features from semantic information; a physics-aware loss function to enforce consistency with real-world physical characteristics; and a reference-free predictor that enables evaluation of generated outputs without requiring ground truth. Experiments with data collected from real tire manufacturing processes show that ProtoDiff achieves an average contact error of 6.8% and a reference-free predictor <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> score of 0.92. The proposed framework significantly reduces the environmental and financial costs associated with traditional prototyping and demonstrates that deep generative models can move beyond creative design exploration to support accurate, constraint-aware design evaluation for manufacturing contexts.</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

Beyond creative generation: a deep generative framework for high-fidelity pseudo-virtual prototyping and design evaluation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月9日

查看全文
  • Sol Lee,
  • Sungrae Hong,
  • Jisu Shin,
  • Mun Yong Yi

摘要

Although recent advances in deep generative models have transformed product development across manufacturing industries, their application in design evaluation remains limited due to the challenges of ensuring fidelity and reliability required by strict industrial standards. We introduce ProtoDiff, a framework for high-fidelity pseudo-virtual prototyping and design evaluation, demonstrated through the task of generating tire footprint images - a critical element in tire performance analysis. Built upon diffusion-based architectures, the framework integrates three core components: a dual-stream embedding method for processing complex structured tabular data by separating numerical features from semantic information; a physics-aware loss function to enforce consistency with real-world physical characteristics; and a reference-free predictor that enables evaluation of generated outputs without requiring ground truth. Experiments with data collected from real tire manufacturing processes show that ProtoDiff achieves an average contact error of 6.8% and a reference-free predictor \(R^{2}\) R 2 score of 0.92. The proposed framework significantly reduces the environmental and financial costs associated with traditional prototyping and demonstrates that deep generative models can move beyond creative design exploration to support accurate, constraint-aware design evaluation for manufacturing contexts.