Advancing coal gangue-based alkali-activated materials development through innovative data augmentation and machine learning strategies

Abstract

In the pursuit of sustainable solutions, the incorporation of coal gangue for alkali-activated materials (AAMs) synthesis offers a promising avenue for diminishing energy expenditure and mitigating carbon emissions. The intrinsic variability of coal gangue properties, however, necessitates exhaustive empirical investigations to elucidate its optimal mix proportion, a process often characterized by its inefficiency and associated costs. Addressing this limitation, the present study delineates a novel design methodology underpinned by data augmentation techniques and machine learning (ML) algorithms. Utilizing the capabilities of Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN), we succeeded in augmenting the constrained experimental dataset. The reliability of such augmented data was subsequently evaluated through a custom difference matrix. Leveraging the expanded data, a quintet of ML models were devised and subjected to a secondary assessment. Through an amalgamation of the superior performing GAN and Multi-Layer Perceptron (MLP), we engineered a robust design framework, facilitating the creation of a mix proportion repository for AAMs derived from coal gangue. This framework, boasting a coefficient of determination (R²) valued at 0.959 and an MAE of 2.643 MPa, offers pivotal insights for deriving mix proportions congruent with stipulated strength criteria. This endeavor signifies a notable stride in the realm of AAMs, underscoring the instrumental role of data augmentation and ML in refining mixture design paradigms.