Scholarlink

<p>Predicting rare failures in high-dimensional engineering systems remains challenging: existing reliability methods either incur prohibitive computational costs or fail to capture irregular, multimodal failure boundaries. Here we introduce an automated importance sampling framework using mixture von Mises–Fisher–Nakagami distributions that naturally captures directional statistics in high-dimensional spaces. The method employs subset simulation to map failure regions, constructs locally adaptive probability densities without geometric assumptions, and integrates Bayesian machine learning to handle expensive computational models. Across benchmark problems and practical applications in aerospace and semiconductor packaging, the approach achieves 10- to 100-fold efficiency gains over current techniques while eliminating manual parameter tuning. By enabling robust quantification of rare-event probabilities in previously intractable problems, this framework advances reliability-informed design for next-generation engineering systems operating under uncertainty.</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

Adaptive von mises–fisher–nakagami importance sampling for high-dimensional rare event simulation

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Yuming Zhang,
  • Xiaojing Zheng,
  • Juan Ma

摘要

Predicting rare failures in high-dimensional engineering systems remains challenging: existing reliability methods either incur prohibitive computational costs or fail to capture irregular, multimodal failure boundaries. Here we introduce an automated importance sampling framework using mixture von Mises–Fisher–Nakagami distributions that naturally captures directional statistics in high-dimensional spaces. The method employs subset simulation to map failure regions, constructs locally adaptive probability densities without geometric assumptions, and integrates Bayesian machine learning to handle expensive computational models. Across benchmark problems and practical applications in aerospace and semiconductor packaging, the approach achieves 10- to 100-fold efficiency gains over current techniques while eliminating manual parameter tuning. By enabling robust quantification of rare-event probabilities in previously intractable problems, this framework advances reliability-informed design for next-generation engineering systems operating under uncertainty.