Prediction of unconfined compressive strength of cement paste with pure metal compound additions

Neural network analysis was used to construct models of unconfined compressive strength (UCS) as a function of mix composition using existing data from literature studies of pure compound additions to Portland cement paste. The models were able to represent the known nonlinear dependency of UCS on age and water content, and generalised from the literature data to find relationships between UCS and contaminant concentrations, resulting in the following ranking of the UCS values predicted for addition of the contaminants, on an equimolar basis: at 7 days, Cl≈Cr(III)>NO₃⁻≈Cd>control>Zn≥Ni>Pb>Cu?Ba; at 28 days, Cl>Cr(III)>NO₃⁻≈control≥Zn≥Cd>Ni>Pb>Cu?Ba. Application of the best neural network to other data suggested that Cs is a retarder and Cr(VI) has no effect. No trends could be discerned for Hg, K, Mn, Na and SO₄²⁻. The root-mean-square error for the best neural network seems to be an estimate of the interlaboratory error for UCS.