Cluster-based adaptive sampling methodology for systems modeling

doi:10.1016/j.compchemeng.2025.109527

Cluster-based adaptive sampling methodology for systems modeling

2026-01-21

0

PRE

AI

DOI：10.1016/j.compchemeng.2025.109527

Original

Original request for help

Save

Abstract

En 中文

Modeling real-world (experimental) or simulated (computational) systems using data-driven surrogate models involves selecting a sampling technique to generate the input-output data for training and selecting a surrogate form. In this work, we present a novel sampling technique, Cluster-based Adaptive Sampling, that generates training data smartly and adaptively for developing surrogate models over a given input domain. CAS iteratively clusters sampled points, defines Voronoi tessellation of cluster centroids, and approximates the tessellations using simple hypercubes. It then searches locally and globally over the domain at each iteration to identify nonlinear and under-explored regions respectively, where it samples two new points using a distance-based metric. CAS is agnostic to surrogate form and terminates automatically based on a surrogate quality metric. We assessed CAS against two existing sampling techniques on 40 diverse test functions using six surrogate forms. CAS outperformed both techniques in developing more accurate surrogates for a given computational effort and required lower computational effort for a specified accuracy across most test functions and forms. We highlight the practical applicability of CAS in modeling two pharmaceutical processes and showcase its superior performance over the two techniques.

Keywords:

Systems modeling

Design of experiments

Adaptive sampling

Active learning

Surrogate modeling

Machine learning

Cluster-based adaptive sampling methodology for systems modeling

Abstract

Journal

Organization