An intelligent hybrid machine learning framework for compressive strength prediction of alkali-activated binders based on fly ash characteristics

The inconsistent physicochemical properties of coal combustion fly ash limit its reliable utilization in alkali-activated binders, despite its potential as a sustainable precursor. To overcome this challenge, this work proposes a hybrid machine learning framework that incorporates dataset optimization for the prediction of alkali-activated binder performance based on the characteristics of fly ash (e.g., chemical composition, particles state, and leaching capacity). The framework is comprised of three key segments: data-optimization, data-preparation module, as well as the training module. The framework addresses data scarcity through synthetic sample generation via attention-enhanced generative adversarial networks, followed by anomaly removal using isolation forest algorithms. Subsequently, an optimized database related to compressive strength was creation to analyzed the performances of six models, in which the transformer model shows the best ability, with testing determination coefficient of the transformer model increased from 0.89 to 0.97 following the implementation of the framework. The generalization of the model was evaluated via microstructural analysis and previous calculated model.