Strength prediction and optimization for ultrahigh-performance concrete with low-carbon cementitious materials – XG boost model and experimental validation

Abstract

In recent years, supplementary cementitious materials (SCMs), as well as other unconventional materials [such as recycled glass powder (GP)], have gained much attention for the formulation of ultra-high-performance concrete (UHPC). Machine learning (ML) techniques have become more prevalent in everyday life in the current era of cutting-edge technology. Their applicability is readily evident in a wide range of areas, including civil engineering applications. This research aimed to develop, validate, and apply the novel XG Boost model to predict the characteristic compressive strength of UHPC containing recycled GP under normal moist curing conditions. Based on 309 datasets, the study developed and trained an ML model using different components, including four supplementary cementitious materials (silica fume (SF), fly ash (FA), quartz powder (QP), and glass powder (GP). The feature importance of the developed ML model was assessed using two methods (the Gini index and the SHAP values). Based on the validation datasets, the predicted-target data was within a 95% accuracy range. The optimized model was helpful tool for making accurate predictions over a wide range of data sets. It was observed from the results obtained that the most significant features were SF, water–binder ratio, water–powder ratio, and virtual packing density. The developed ML model and the tested mixtures showed a mean ratio of tested-to-predicted values of 0.956, with most data lying within a margin of error of 15%. As a result, an intuitive user interface was developed for optimizing UHPC systems, including SF, FA, QP, and GP.