Fiber-reinforced ultra-high-performance concrete (FR-UHPC) has matured over the past 30 years from niche lab mixes to critical responses for high-performance infrastructure, but capabilities of the community continue lag behind expressed multi-crosscut needs in robust vascular prediction of mechanical performance after exposure to fire across modern concrete families. Extension from the historical empirical and mechanistic modeling to more recent data-driven endeavours along with contributions herein, this study presents and validates a full machine learning-based, interpretable framework for the residual flexural strength (RFS) of FR-UHPC following exposure to elevated temperatures. Utilizing a wide experimental database of 800 three-point bending tests, covering binder chemistry, contents of silica-fume and fly-ash, various fiber types (PVA, steel, basalt), curing regimes, and thermal paths, nine state-of-the-art algorithms including the novel hybrid AegisFusion architecture were benchmarked. AegisFusion is a dual-track neural/tree fusion with MetaSwarm tuning and Bayesian calibration. It achieved superior accuracy and calibrated uncertainty in both hold-out (R² = 0.98, RMSE ≈ 0.52 MPa, VAF = 0.98), and 5-fold (R² = 0.93–0.98; RMSE = 0.28‒0.57, VAF = 0.93‒0.98) testing phases. Rigorous non-parametric testing including Friedman and Nemenyi critical difference analysis, ranked AegisFusion best. Model explainability via SHAP and distance-correlation analysis identified exposure maximum temperature (EMT) and fiber volume content (FV) as dominant drivers of RFS, revealing thresholded and nonlinear interactions that align with known thermo-mechanical degradation mechanisms. This study thus forges a data-centric bridge between the historical experimental foundations of FR-UHPC research and contemporary, uncertainty-aware artificial intelligence tools for resilient infrastructure design.
Kewalramani et al. (Mon,) studied this question.