Levenberg-Marquardt Based Prediction Models for Slabs with Magnesium Sacrificial Anodes Subjected to Chloride Ingress

This work focused on developing prediction models using ANN, in order to forecast the long-term performance of reinforcements in concrete slabs containing pure Mg anodes, and subjected to Cl- ingress. The experimental set-up consisted of two built RCC slabs with 1000 x 1000 x 100 mm. Slab #1 was cast with 3.5% NaCl by cement weight. Slab #2 was cast without NaCl. Considered input parameters were the distance of the anode from the point under consideration in x and y-axes, T, RH and concrete age in days. HCP values concerning SCE were considered output. Experimental values consisted of 80 HCP per slab/day. These were collected for 270 days, in order to generate the prediction model. The learning heuristic used LM supervised learning in feed-forward. A two-layer feed-forward network, with ten hidden sigmoid neurons and trained linear output neurons, was employed. The network architecture [5-10-1] and ten neurons in the hidden layer were used for all prediction models. The accuracy level of results obtained with LM was above 97%.