Generalized uncertainty in surrogate models for concrete strength prediction

Applied soft computing has been widely used to predict material properties, optimal mixture, and failure modes. This is challenging, especially for the highly nonlinear behavior of brittle materials such as concrete. This paper proposes three surrogate modeling techniques (i.e., polynomial chaos expansion, Kriging, and canonical low-rank approximation) in concrete compressive strength regression analysis. A benchmark database of high-performance concrete is used with over 1,000 samples, and various sources of uncertainties in surrogate modeling are quantified, including meta-modeling assumptions, solvers, and sampling size. Two generalized extreme value distributional models are developed for error metrics using an extensive database of collected data in the literature. Bias and dispersion in the developed surrogate models are empirically compared with those distributions to quantify the overall accuracy and confidence level. Overall, the Kriging-based surrogate models outperform 80%-90% of the existing predictive models, and they illustrate more stable results. The selection of a proper optimization algorithm is the most important factor in surrogate modeling. For any practical purposes, the Kriging regression outperforms the polynomial chaos expansion and the low -rank approximation. The Kriging model is reliable for the pilot database and is less sensitive to modeling uncertainties. Finally, a series of composite error metrics is discussed as a decision-making tool that facilitates the selection of the best surrogate model using multiple criteria.