Explainable creep prediction and design optimization of SBS/CR modified asphalt: a spectral information coupling encoding and physical data-driven framework

• The spectral information coupling encoding combining Gramian Angular Fields transformation and AutoEncoder can effectively extract FTIR micro-component features and improve prediction accuracy by 9.4%. • A physically driven, robust, and interpretable framework is proposed for predicting the low-temperature creep performance of SBS/CRMA; • SHapley Additive exPlanations and two-dimensional gradient attribution techniques provide quantitative multi-scale insights into the influence of both macroscopic and microscopic features. • The design optimization method enables SBS/CRMA formulation design under target service conditions, achieving consistency with experimental results within 7%. Styrene-butadiene-styrene copolymer and crumb rubber modified asphalt (SBS/CRMA) is widely used in infrastructure structure. However, its durability is significantly influenced by creep behavior, particularly under low-temperature conditions and freeze–thaw cycles, making accurate and explainable prediction essential for design and application. This study explores a predictive framework for low-temperature creep behavior that integrates spectral information coupling encoding with physics-informed neural networks (PINNs). By combining Gramian Angular Field (GAF) transformation with AutoEncoders (AE), the framework develops a feature extractor capable of coupled FTIR spectral information and capture micro-component inputs. Two PINNs with distinct tasks and architectures were developed for prediction. To advance interpretability, SHapley Additive exPlanations (SHAP) and a developed interpretation strategy for the encoded spectral are employed, providing multi-scale insights into the influence of both macroscopic and microscopic features. The results demonstrate that the feature extraction successfully captures functional group coupling information, improves prediction accuracy by 9.4%, and the PINNs models exhibit clear advantages over conventional approaches. Furthermore, the framework was integrated into a material design optimization workflow, and its feasibility was verified through application in a specialized engineering project. This study advances intelligent characterization and design of SBS/CRMA, offering a novel perspective for modeling and design of complex materials.