Unified prediction model for concrete strength based on progressive feature fusion and knowledge transfer

Abstract

Existing concrete strength prediction models are typically trained on single material systems, limiting their generalization to diverse concrete types. This study proposes a unified prediction model (UPM) based on progressive feature fusion and knowledge transfer. The UPM introduces two core mechanisms. First, a progressive feature fusion framework that dynamically aligns and expands features across datasets via zero-padding and global feature indexing, resolving feature incompatibility among different material systems. Second, a dynamic sample weighting mechanism based on prediction error that enhances knowledge transfer by intensifying the model’s focus on high-error samples during incremental training. Validated on five multi-source datasets covering standard, geopolymer, phosphogypsum, lightweight high-strength, and recycled aggregate concrete, the model achieves an average R² of 0.88, with predictive performance on small-sample and special material tasks improved by over 30% compared to baseline models. SHAP interpretability analysis confirms that the model retains knowledge of universal governing laws while precisely capturing the nonlinear influence of special components on strength, providing a generalized and accurate solution for concrete strength prediction across diverse material systems.