Hybrid Deep Learning with Conformal Prediction for Recycled Aggregate Self-Compacting Concrete Strength Prediction

This study presents a novel hybrid deep learning framework integrating Feature Tokenizer-Transformer (FT-Transformer) with Masked Multi-Layer Perceptron (Masked MLP) for predicting the compressive strength of recycled aggregate self-compacting concrete (RASCC). The framework addresses incomplete data challenges through a missingness-aware fusion strategy and two-stage stacking scheme with Ridge regression. Using a dataset of 289 experimental records with 18 input parameters, the hybrid model achieved robust predictive performance with enhanced generalization stability (Test R2 = 0.940, RMSE = 4.219 MPa) while demonstrating consistent predictions under data missingness conditions up to 25%. SHAP analysis revealed that cement content, water-to-binder ratio, and curing age are the dominant factors influencing RASCC strength. The proposed uncertainty quantification via split conformal prediction provides 90% coverage with average interval width of 8.32 MPa, enabling practical engineering applications with quantified reliability.