Engineering the Peel Strength of Biobased Polycarbonate Diol-Derived Polyurethane Adhesives via Machine Learning with a Small Data Set

Abstract

Biobased polycarbonate diol (BPCDL)-derived polyurethane adhesives (PUAs) are promising sustainable materials. Among various properties, the peel strength is a key indicator for evaluating the adhesion reliability of adhesives. However, the peel strength is governed by coupled formulation and processing variables, limiting its accurate prediction and regulation. To address this issue, a small-data-driven and interpretable machine-learning-assisted prediction framework was developed to engineer the peel strength of BPCDL-based PUAs. A full factorial data set covering five formulation and processing variables, including soft segment ratio (SSR), cross-linking level (CL), R value, hot-pressing temperature, and hot-pressing time, was constructed to sample the multidimensional design space. Extremely Randomized Trees (Extra Trees) and Gaussian Process Regression (GPR) were benchmarked against advanced statistical baselines under repeated and structured validation, while analysis of variance, Extra Trees-derived SHapley Additive exPlanations, and GPR-derived response surfaces were integrated to distinguish main effects, interactions, and processing modulation. The results identify R value as the principal formulation factor, reveal a pronounced interaction between SSR and R value, and highlight hot-pressing temperature as a key processing modulator. The model indicated that high peel strength could be achieved through multiple formulation and processing combinations, including low SSR and low R value, high SSR and high R value, or moderate R value with sufficient hot-pressing temperature. Increasing CL, however, does not necessarily enhance the peel strength and may even suppress it. Molecular dynamics simulations further suggest that increasing temperature weakens cohesive interactions within the polyurethane while enhancing interfacial interactions with the substrate, thereby balancing chain mobility, interfacial adhesion, and bulk cohesion. Beyond forward prediction, a multicriteria inverse design strategy was proposed by integrating milder hot-pressing conditions, predictive uncertainty, and target-oriented constraints, enabling uncertainty-informed screening of promising candidates. This work provides an interpretable small-data strategy for inverse design and performance optimization of advanced biobased adhesives.