Self-training AI for geopolymer strength prediction from microstructure
Year: 2026
Type: article
Abstract: Establishing quantitative link between microstructural features and macroscopic properties remains a challenge in cementitious materials, particularly for reducing reliance on destructive large-scale testing and coring. Recent advances in machine learning show improved predictive capability; however, their effectiveness is constrained by data scarcity, exhaustive labelling and limited interpretability. This study introduces a self-training framework for achieving high-performance strength prediction in data-constrained scenarios on a model metakaolin-based geopolymer system. The framework employs a two-stage approach: first, a robust segmentation model is achieved through self-training, complemented by an adversarially generated synthetic dataset. Second, high dimensional quantitative microstructural descriptors, including phase fractions and particle-level geometric features, were extracted from the microstructure to train a classification model. The proposed framework achieved a classification accuracy of over 95% for predicting the geopolymer strength. By bridging the gap between microstructural imaging and data-driven predictive modelling, the framework enables scalable and quantitative linkage between microstructural descriptors and macroscopic performance, establishing a rigorous basis for microstructure-informed design and evaluation of cementitious materials.
Source: Cement and Concrete Research
Institutions: Technical University of Denmark, Riga Technical University
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $5,180
OpenAlex ID: W7233580671
