Optimized Physics-informed Machine Learning Modeling with GAIN Data Imputation for Sustainable Pozzolanic Concrete Strength Prediction

Abstract

The reliable design of sustainable concrete demands predictive models integrating data-driven insights with fundamental physical laws. This paper presents a new framework to predict the 28-day compressive and tensile strengths of pozzolanic concrete with metakaolin, silica fume, and zeolite. This study collected 440 experimental samples from published studies, using 18 input variables related to mix composition, materials, and specimen geometry. Unlike previous research, this study included several influential factors, such as cement grade, the active solid mass of superplasticizer, pozzolanic particle size, and specimen dimensions for developing models. To address missing data, a Generative Adversarial Imputation Network (GAIN) was developed and tested its reliability by comparing it to XGBoost-based Multivariate Imputation by Chained Equations (MICE) with missing rates up to 90%. This paper proposes a new Physics-Informed Machine Learning (PIML) framework that combines an XGBoost predictor with fundamental concrete mechanics constraints, such as porosity–strength relationships, size-effect laws, and tensile–compressive strength correlations, which, to the best of our knowledge, have not been studied together before. Model hyperparameters and coefficients were optimized using the Improved Hybrid Growth Optimizer (IHGO) algorithm. The proposed PIML model achieved reliable predictive performance, with coefficients of determination (R²) of 0.9612 and 0.8671 on an independent testing set for compressive and tensile strengths, respectively. Analysis using PDP and SHAP confirmed that pozzolanic fineness, cement quality, water content, and superplasticizer activity are the most important factors. This framework offers an accurate, understandable, and physically based tool for designing sustainable pozzolanic concrete.