Comparative analysis of artificial neural networks models for predicting mortar properties with diatomite incorporation

This study investigated the predictive capabilities of Artificial Neural Networks (ANNs) for the physical, thermal, and mechanical properties of diatomite-based mortars, a promising alternative to conventional sand-based mortars. We compared the performance of various ANN architectures, including single-layer ANN and multilayer ANN models with different activation functions to predict porosity, water absorption, thermal conductivity, diffusivity, compressive strength, and flexural strength. The models were evaluated using metrics such as the Mean Squared Error (MSE), Mean Absolute Error (MAE), coefficient of determination (R2), and maximum error. The results demonstrated that the multi-layer ANN model with the Leaky ReLU activation function exhibited superior predictive accuracy for the studied properties across varying diatomite replacement levels. Notably, the average MSE across all the properties for the best-performing model was below 0.0092. Sobol sensitivity analysis revealed that the diatomite content has a substantial influence on the properties of the novel mortar. Diatomite not only directly affects these properties, but also interacts synergistically with other components to improve thermal insulation and increase physical property values, despite the potential expense of mechanical strength.