Multi-task deep learning with transfer learning for UHPC compressive and flexural strength prediction and inverse mix design

This study develops a domain-adapted multi-task learning (DA-MTL) framework for jointly predicting compressive strength (CS) and flexural strength (FS) of ultra-high-performance concrete (UHPC) and enabling inverse mix design under limited local experimental data. A literature-derived source-domain database comprising 170 UHPC mixtures was constructed under consistent mixture design and testing conditions, while a target-domain laboratory dataset of 25 locally optimized UHPC mixtures was established using MAA-based mix design and laboratory mechanical testing. The proposed DA-MTL model integrates shared feature extraction, dynamic weighted averaging (DWA), and a staged transfer learning strategy with CORAL-based domain adaptation to mitigate distribution mismatch between the databases. Forward prediction results demonstrate substantial improvements in accuracy, stability, and generalization performance for both CS and FS compared with single-task baseline models. In addition, a physics-consistent inverse design module is integrated to enable efficient performance-driven UHPC mixture proportioning under particle packing and material feasibility constraints, and is verified through laboratory testing of selected mixtures. Overall, the proposed DA-MTL framework provides a reliable and data-efficient approach for UHPC strength prediction and inverse mix design, bridging literature-derived knowledge and local engineering practice. • A domain-adaptive multi-task learning framework is proposed for UHPC strength prediction. • CORAL-based domain adaptation improves generalization under limited local experimental data. • Joint learning captures shared and task-specific mechanisms governing compressive and flexural strength. • SHAP analysis provides physically interpretable feature contributions. • A physics-consistent inverse design module enables performance-driven UHPC mix proportioning.