A Hybrid Polynomial Chaos Expansion-Based Surrogate Framework for Reliability and Interpretability of Concrete Compressive Strength

Abstract

Probabilistic analysis of concrete strength is essential for reliability-based design, but it requires surrogate models that are both accurate and analytically tractable. Classical statistical surrogates such as polynomial chaos expansion (PCE) are well suited for global sensitivity and reliability analysis, yet they can struggle when fitted directly to sparse, noisy experimental data. In contrast, machine learning models such as gradient-boosted trees (GBTs) achieve high predictive accuracy but are difficult to use directly for uncertainty quantification (UQ). This study develops and validates a hybrid framework that combines these strengths. First, a GBT model is trained on a curated experimental database of concrete mixes and used as a high-fidelity “signal extractor,” learning the underlying nonlinear mapping from mix composition and curing age to compressive strength. Second, a PCE is constructed as a surrogate of this deterministic GBT response using Latin hypercube sampling. The resulting hybrid GBT-PCE model enables efficient global sensitivity analysis (Sobol and Kucherenko indices) and reliability assessment. Its performance is critically compared with two benchmarks: a standalone PCE trained directly on the data and an advanced polynomial chaos–Kriging (PCK) surrogate. Results show that the hybrid model attains predictive accuracy comparable to the best purely data-driven surrogates while providing numerically stable global sensitivity indices. Interpretability in the proposed framework is provided by global sensitivity indices and reliability metrics obtained from the PCE surrogate. The framework therefore offers a practical tool for risk-informed design and mix optimization, especially when experimental data are noisy and multisource.