Artificial neural networks test for the prediction of chemical stability of pyroclastic deposits-based AAMs and comparison with conventional mathematical approach (MLR)

The investigation on the reticulation degree of volcanic alkali-activated materials, AAMs, were experimentally determined in terms of chemico-physical properties: weight loss after leaching test in water, ionic conductivity and pH of the leachate and compressive strength. Artificial neural network (ANN) was successfully applied to predict the chemical stability of volcanic alkali-activated materials. Nine input data per each chemico-physical parameter were used to train each ANN. The training series of specific volcanic precursors were tested also for the other one. Excellent correlations between experimental and calculated data of the same precursor type were found reaching values around one. The evidence of strong effect on chemical stability of the alkaline activator SiO2/Na2O molar ratio as well as the Si/Al ratio of precursor mixtures on the reticulation degree of ghiara-based formulation with respect to volcanic ash-based materials is presented. It must be noted that such effect was much less pronounced on the compressive strength values, appearing more insensitive the molar ratio of the alkaline activator. The comparison of the ANN results with more conventional multiple linear regression (MLR) testifies the higher prediction performance of the first method. MLRs results, less significant, are useful to confirm the powerful capacity of ANNs to identify the more suitable formulation using a set of experimental AAMs. This study, as few others, on the correlation between chemical stability and compressive strength of AAMs provide a great contribution in the direction of durability and in-life mechanical performance of these class of materials. [GRAPHICS] .

Automated summary

Experimental investigation and artificial neural network modeling were used to predict the chemical stability of volcanic alkali-activated materials. The SiO2/Na2O molar ratio of the alkaline activator and the Si/Al ratio of precursor mixtures were found to significantly affect the reticulation degree of these materials. However, the compressive strength values were less influenced by these factors. The comparison between ANN and MLR results showed that ANN had higher prediction performance in identifying suitable formulations for AAMs.