Machine learning assisted design of low-carbon aluminosilicate cementitious composites with diverse raw materials and target mechanical strength

The synergistic design of cementitious composites with low carbon emission and high performance has been the continuous pursuit in the field. Despite great efforts have been drawn for decades on exploring various low-carbon raw materials, there remain significant challenges in the universal and precise design of cementitious composites satisfying desired performance without abundant trial-and-error. This study proposes a machine learning assisted design framework for the aluminosilicate cementitious composites based on the data extracted from hundreds of relevant literatures in recent five years. The well-trained strength prediction model is applicable for almost all supplementary cementitious materials incorporated composites regardless of the diversity of material origin, as long as the proportion of each material is provided with its key chemical constituents. Based on this, several cementitious composites with target strength value and carbon emission are successfully designed based on the available raw materials which are even not included in the training dataset. Furthermore, the influences of various factors on the strength of composites are systematically analyzed through detailed interpretation of machine learning model. This research provides an effective alternative for the development of cementitious composites with diverse calcium/silicon/aluminum-rich materials and multi-objective performances. • The proposed framework combines a prediction model (R² > 0.99) and a reverse design strategy (<5% error). • Model emphasizes material essence and eliminates compatibility issue from diverse material, applicable to most SCMs. • Diverse composite datasets are used separately for training and testing, boosting model generalization and robustness. • Synergistic optimization of low carbon and high strength is achieved with raw materials even absent from the dataset.