El uso potencial de policloruro de aluminio (PAC) en morteros y hormigones a base de cemento: Predicción de las propiedades mecánicas mediante modelos de aprendizaje automático y análisis del impacto ambiental

Abstract

CONDITION FOR PROJECT PUBLICATION. The concrete industry faces dual pressure to reduce its significant carbon footprint, associated with cement production, and to sustainably manage increasing volumes of industrial waste. This research addresses both challenges by evaluating Polyaluminum Chloride (PAC) residual sludge, a byproduct of water treatment, as a supplementary cementitious material (SCM). The primary objective was to develop and validate an artificial intelligence predictive model to optimize PAC dosage in concrete and to analyze its technical, economic, and environmental viability. This research utilized a mixed-methods approach, integrating an exhaustive literature review with the analysis of a reference dataset (UCI Concrete Compressive Strength dataset) with the development of a Multilayer Perceptron (MLP) neural network model. A comparative analysis of machine learning algorithms was performed to validate the predictive model. The results demonstrated that while ensemble models like Random Forest (R²=0.954) and the Multilayer Perceptron (MLP) (R²=0.952) showed outstanding and very similar predictive performance, the neural network model was chosen as the model for getting the lowest Mean Absolute Error (2.48 MPa), confirming its robustness and superiority for this engineering application. The model was trained on a robust, unified database of over 3000 samples, including real-world reference data and PAC-specific synthetic data. The results demonstrate the model's high predictive capability, obtaining a Coefficient of Determination (R2) of 0.952 and a Mean Absolute Error (MAE) of 2.48 MPa of compressive strength. Variable importance analysis confirmed the significant role of PAC and its chemical activation in compressive strength. The model identified an optimized mixture with a 27.8% binder substitution by PAC, which not only surpasses the control mix's strength by 34% at 90 days but also improves economic efficiency (MPa/$) by 23%. The life cycle assessment demonstrated that conventional concrete has a total GWP of 350.00 kg CO₂ eq/m³, while the optimized mixture with PAC sludge resulted in a GWP of 375.25 kg CO₂ eq/m³. This thesis concludes that PAC sludge is a viable and high-potential SCM, and that machine learning models are indispensable tools for accelerating its adoption, facilitating the design of sustainable, high-performance concrete. Keywords: polyaluminum chloride, sludge, water treatment, circular economy, neural networks, concrete, cement