Machine learning-aided design and prediction of cementitious composites containing graphite and slag powder

The electrically conductive cementitious composite (ECCC) offers plenty of advantages such as high conductivity and strain sensitivity. The ECCC can also act as a conductive sensor in a cathodic protection system for structural health monitoring. Before the ECCC application, it is essential to understand and predict the uniaxial compressive stress (UCS) and electrical resistivity. In this study, we produced ECCC with three conductive fillers: graphite powder (GP), waste steel slag (SS) as well as ground granulated blast-furnace slag (GGBS). By changing the content levels of the three conductive fillers, cement and curing ages, we prepared 81 mixture proportions for UCS test and 108 mixture proportions for resistivity test. The results show that although GP improves the conductivity more significantly than the other conductive fillers but it simultaneously has a higher negative influence on UCS. Meanwhile, slag solids (GGBS and SS) enhance the conductive performance but reduce UCS after their replacement ratio is larger than 20%. Compared with GGBS, ECCC containing SS has higher UCS and conductivity. Besides, we proposed a random forest (RF) based machine learning model to predict the UCS and resistivity. The hyperparameters of the RF model were tuned by the beetle antennae search (BAS) algorithm. This hybrid BAS-RF model has high prediction accuracy, as indicated by high correlation coefficients on test sets (0.986 for UCS and 0.98 for resistivity, respectively). We simulated the influence of different conductive fillers on UCS and conductivity using the developed BAS-RF model. The simulation results agree well with the results obtained by laboratory experiments. This study offers a new idea to use waste slags to produce ECCC and paves the way to intelligent construction.