Scholarlink

<p>We present a novel sequential experimental design framework that combines the interpretability of classical response surface methodology with the adaptability of Bayesian optimization. At each iteration, a second-order polynomial surrogate model is refitted and the next experiment is selected using a variance-weighted curvature (VWC) acquisition function that targets locations where the surrogate is uncertain and/or strongly curved. Sampling in these information-rich regions can improve global model fidelity while still revealing optima. Through three benchmark problems—including a five-dimensional sparse quadratic and a non-quadratic surface—the VWC criterion achieves one to three orders of magnitude lower prediction error than Gaussian process-based Bayesian optimization while requiring significantly less computation. The proposed framework is fast, interpretable, and readily scalable, making it well suited to data-intensive experimentation in chemical engineering and related fields.</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

A Variance-Weighted Curvature Criterion for Sequential Experimental Design

  • 期刊论文

  • 发表日期：

    2025年9月1日

查看全文
  • Boeun Kim,
  • Kyung Hwan Ryu

摘要

We present a novel sequential experimental design framework that combines the interpretability of classical response surface methodology with the adaptability of Bayesian optimization. At each iteration, a second-order polynomial surrogate model is refitted and the next experiment is selected using a variance-weighted curvature (VWC) acquisition function that targets locations where the surrogate is uncertain and/or strongly curved. Sampling in these information-rich regions can improve global model fidelity while still revealing optima. Through three benchmark problems—including a five-dimensional sparse quadratic and a non-quadratic surface—the VWC criterion achieves one to three orders of magnitude lower prediction error than Gaussian process-based Bayesian optimization while requiring significantly less computation. The proposed framework is fast, interpretable, and readily scalable, making it well suited to data-intensive experimentation in chemical engineering and related fields.