سيدعلي افتخارافضلي

Approximately 4 billion tons of CO2 emissions are released annually from cement production industries, making them one of the primary sources contributing to the increase in greenhouse gas levels in the atmosphere. As a result of growing concerns about global warming and climate change in recent years, geopolymers and alkali-activated materials have been introduced as environmentally friendly alternatives to cement-based concrete. These materials can be produced with lower energy consumption and contribute to reducing CO2 emissions from the construction industry. Given the crucial role of geopolymers and alkali-activated materials in sustainable constructions and the lack of reliable models for predicting their mechanical properties and durability, this study aimed to assess, employ, and eva‎luate the capabilities of various machine learning algorithms in modeling and interpreting the nature of these systems. To achieve this, databases were developed by reviewing various sources in the technical literature, including mix design, chemical compositions of precursor materials, as well as curing and pre-curing conditions. Hundreds of models were then developed using machine learning algorithms, including gradient boosting machine, random forest, decision tree, artificial neural network, and support vector machine. In the next stage, the developed models were eva‎luated using various performance metrics, and the predictions made by the best models were interpreted using different methods such as feature importance analysis and Shapley explanations. The effects of variations in different input parameters on the model outputs were also investigated using partial dependence plots. The results demonstrate the outstanding performance of gradient boosting machine model in predicting the compressive strength of metakaolin-based geopolymer concrete. Additionally, it was found that the coarse-to-fine aggregate ratio, water content in the mix design, H2O/Na2O molar ratio, and the amount of sodium hydroxide solution used in the mix design have the greatest influence on the mechanical properties of metakaolin-based geopolymer concrete. Furthermore, in the modeling of the durability of blast furnace/ fly ash-based alkali-activated materials, it was observed that artificial neural networks with a multilayer perceptron architecture perform best in predicting the non-steady state chloride ion migration coefficient. Additionally, the results highlight the significant impact of the atomic ratio of Ca/Si, the content of SiO2 in the alkali activator solution, the total water content in the mix design, and the curing temperature based on feature importance analysis.

يادگيري ماشين , ژئوپليمر , مواد قليايي فعال , مقاومت فشاري , ، ضريب مهاجرت ناپايدار يون كلريد

machine learning , geopolymer , alkali-activated materials , compressive strength , non-steady state chloride migration coefficient

Seyed Ali Eftekhar Afzali

Mohsen Ali Shayanfar