In the context of global low-carbon and carbon-neutrality targets, chemico-activated aluminosilicate cementitious materials that are represented by geopolymers become one of the most promising green alternatives to ordinary Portland cement. Geopolymers are synthesized via the chemico-activation of reactive aluminosilicate solid wastes with an alkaline activator. Their manufacturing process is featured by low energy consumption and reduced carbon emissions through the valorization of industrial solid waste or low-energy natural mineral precursors. Furthermore, the reaction products are endowed with high early-age mechanical strength, excellent durability and fire resistance. However, the complexity of raw material composition, unclear polymerization mechanisms and intricate interdependencies among multiple variables result in low efficiency, poor reproducibility and accuracy of conventional trial-and-error, seriously restricting the multi-objective performance optimization and large-scale application of geopolymers.

Based on the materials initiative and rapid development of artificial intelligence, the data-driven paradigm initiates a fundamental shift in materials science research. Establishing the relationships between high-dimensional data (e.g., materials compositions, structure, physical and chemical properties) and the material performance through learning massive multi-source data can effectively address the limitations of the conventional methods, i.e., high cost, long time and numerous resources. It therefore provides intelligent and accelerated routes to revolutionize the design and investigation of chemically activated aluminosilicate cementitious materials, leading a paradigm shift from experimental-driven to data-driven approach. This review represents research progress on data-driven approaches in geopolymer, with emphasis on three aspects, i.e., 1) machine learning methodologies for performance prediction; 2) application of machine learning in targeted performance prediction; and 3) metaheuristic algorithms-based composition optimization.

This review proposes a complete data-driven workflow for geopolymer research, including target definition, dataset collection, data processing, feature engineering, model construction and application. Based on the application scenarios, the existing models can be divided into four categories, i.e., neural networks, ensemble learning, traditional regression and statistical learning, and symbolic regression combined with neuro-fuzzy systems. Each category offers distinct advantages and disadvantages. Neural networks capture nonlinear relationships, but require large dataset and lack interpretability. Ensemble learning handles high-dimensional mixed data with robustness, but risks overfitting with small samples. Traditional regression combined with neuro-fuzzy systems works for small to medium dataset, but faces an exponential rule growth under high-dimensional inputs. In terms of targeted performance prediction, machine learning is effectively applied to early-age properties, mechanical properties, and durability. For early-age behaviors, machine learning models can accurately predict geopolymerization heat flow, peak time, setting time and slump. For mechanical properties, compressive strength prediction is the most mature predicted performance, and the prediction is extended to flexural strength, splitting tensile strength, elastic modulus, etc.. For durability, the initial research mainly focuses on drying shrinkage, autogenous shrinkage, carbonation, while some studies on chloride penetration, sulfate attack and long-term durability are relatively scarce.

Metaheuristic algorithms play a significant role on geopolymer composition design. Firstly, metaheuristic algorithms serve to optimize hyperparameters and improve prediction accuracy. Subsequently, the well-trained predictive models can be leveraged to facilitate both forward guidance and inverse mix design. Based on the trained model, the importance of material composition, preparation parameters, and curing conditions to prediction accuracy can be quantitatively assessed. Furthermore, metaheuristic algorithms can also apply to reverse geopolymer mixture design targeting single or multiple objectives. For instance, a Pareto frontier is constructed via leveraging NSGA-Ⅱ to balance high strengths, low carbon and low cost, and the optimal solution can be further selected via TOPSIS. This approach pioneers an effective pathway for the inverse design of geopolymers and establishes a versatile decision-support tool that is poised to become a new standard for the accelerated development of high-performance and eco-efficient cementitious materials.