From molecular descriptors to data-driven design: Multiscale modelling and machine learning for shape memory polymers

Shape memory polymers (SMPs) exhibit coupled structural and mechanical responses spanning molecular, mesoscale, and continuum levels, making predictive design challenging because information is fragmented across modelling scales. This review presents a descriptor-driven framework for integrating multiscale modelling and machine learning (ML), in which physically interpretable descriptors provide a common representation that links electronic structure, molecular interactions, phase morphology, constitutive behaviour, and macroscopic shape memory performance. The review systematically examines quantum-mechanical calculations, molecular dynamics, coarse-grained and dissipative particle dynamics simulations, continuum constitutive modelling, and their connections through quantitative descriptor mapping across modelling hierarchies. It further evaluates ML methods that support forward prediction, surrogate modelling, inverse design, physics-informed learning, adaptive optimisation, and uncertainty-aware modelling in multiscale polymer design. Descriptor transfer between modelling scales is also given attention, as is the preservation of physical interpretability during information mapping, and the complementary roles of simulations, experiments, and ML in establishing predictive workflows. Current limitations, including descriptor standardisation, heterogeneous datasets, uncertainty propagation, and incomplete coupling across modelling levels, are critically discussed alongside emerging opportunities offered by integrated simulation-learning frameworks. By organising recent advances around descriptor continuity and cross-scale information transfer, this review provides a comprehensive perspective on the current state of multiscale SMP modelling and identifies crucial challenges and opportunities for developing more predictive, scalable, and application-oriented computational design strategies.