Active learning for data-scarce multi-objective polymer design: Robust strategies with experimental validation
Journal
Results in Engineering
Journal Volume
32
Start Page
111974
ISSN
25901230
Date Issued
2026-12
Author(s)
Wei, Yu-Chieh
Chang, Wei-Che
Tsai, Po-An
Bai, Chen-Yu
Chu, Ruei-Jing
Mu, Chia-Wen
Chen, Chin-Wen
Abstract
Data-driven polymer design is fundamentally limited by severe experimental data scarcity and intrinsic trade-offs between competing material properties, such as mechanical stiffness and network mobility. While active learning (AL) offers a promising pathway for navigating high-dimensional design spaces, its practical deployment in experimental polymer systems remains hindered by the lack of robust strategies under strict data constraints. In this work, we establish an experimentally grounded active learning framework for data-efficient multi-objective polymer optimization. By systematically evaluating combinations of structured initial sampling and acquisition strategies within a Gaussian Process framework, we identify that the coupling between initialization and acquisition governs optimization efficiency under data scarcity. The framework is experimentally validated on a polyester-based vitrimer system with a highly combinatorial formulation space. With only 28 samples, including 12 initial points and 4 active-learning iterations, the approach rapidly expands the stiffness-mobility Pareto frontier, identifying non-obvious formulations that achieve a 34.5% reduction in stress-relaxation time and a 3.3% increase in Young’s modulus. These results demonstrate that well-designed active learning strategies can extract meaningful Pareto improvements from minimal experimental data and provide practical guidance to accelerate polymer discovery under severe data scarcity. The present validation spans two systems with relatively continuous formulation–property landscapes, a literature-derived polyurethane benchmark, and the polyester-based vitrimer system; extension to highly constrained or high-dimensional design spaces remains to be established.
Subjects
Active learning
Data scarcity
Multi-objective optimization
Polymer design
Vitrimer
Publisher
Elsevier B.V.
Type
journal article