Machine learning guided iterative mix design of geopolymer concrete

Abstract

Current mix design methods for Geopolymer concrete (GPC) require substantial efforts of trial-and-error experiments and is applicable only to those formulated by specific precursor materials. In this work, a machine learning (ML) guided mix design method for GPC is proposed, which can considerably reduce the experimental workload and be versatile with a broad range of precursor materials. First, a database for the slump and compressive strength of GPC was established, and ML prediction models were developed with these as objectives. Subsequently, the mix design for GPC was conducted using the particle swarm optimization (PSO) algorithm. The designed mixture proportions underwent experimental validation, and if targets were not met, results were then used for model iteration until desired performance targets were achieved. The results showcase the excellent performance of the established slump and compressive strength prediction models, with an accuracy and coefficient of determination of 94 % and 0.95, respectively. The developed ML guided model effectively generates concrete mix achieving compressive strengths of 20 MPa, 40 MPa, and 60 MPa, while concurrently satisfying slump requirements. The experimentally validated results for the three designated target strengths after iterations were 25.2 MPa, 43.8 MPa, and 66.6 MPa, respectively. These positive results emphasize the ability of the proposed optimized mix design method to meet both the strength and workability targets in GPC via a minimum number of trail experiments.