Use of Machine Learning to Predict the Performance of Tile Adhesive Mortars

Tile adhesive mortars are industrialized products used for installing ceramic coverings and are classified according to the Brazilian standard ABNT NBR 14081/2012 on the basis of tensile adhesion performance under different curing conditions. Their formulation directly affects both technical performance and manufacturing competitiveness, while conventional product development remains slow, costly and strongly dependent on trial-and-error laboratory testing. This study evaluates whether historical industrial formulation data can support the retrospective prediction of approval or failure of tile adhesive mortars under ambient, oven, immersed and open-time curing conditions. A dataset comprising 6031 individual pull-off observations collected between 2021 and 2023 by a European multinational company in the construction materials sector was used to train and compare Logistic Regression, Random Forest, Boosted Decision Tree and Support Vector Machine models in R and Azure. The study was designed as an industrial-data modelling investigation rather than as a prospective optimization experiment. The results show that ensemble tree-based models, particularly Boosted Decision Tree and Random Forest, achieved the strongest predictive performance, whereas Logistic Regression remained more suitable for inferential interpretation of formulation variables. Model performance was uneven across curing conditions: prediction was more reliable for oven and immersed curing, whereas ambient curing and open time were affected by strong class imbalance and low failure prevalence. The findings indicate that Machine Learning can support formulation screening and quality-oriented decision-making for tile adhesive mortars, provided that its use remains restricted to the formulation ranges represented in the historical dataset and is complemented by prospective experimental validation before deployment in new product development.