Data-driven prediction of strength development of alkali-activated concrete with uncertainty quantification using interpretable amortised Gaussian process

This paper presents a domain-informed amortised Gaussian process (AGP) framework for predicting the strength development of alkali-activated concrete across multiple ages, delivering both accurate forecasts and calibrated uncertainty estimates. The model integrates prior predictions based on input features with posterior updates that refine later-age forecasts using early-age observations. Compared to multi-output extreme gradient boosting models, the AGP approach provides strong predictive performance and well-calibrated uncertainty intervals, whereas conventional methods typically offer only point predictions. Uncertainty decomposition shows that epistemic sources, stemming from the Gaussian process and neural network dropout, dominate, with aleatoric noise playing a minor role. Incorporating early-age strength measurements significantly improves later-age predictions, increases prediction interval coverage probability, and reduces interval width. Feature importance analysis reveals key factors consistent with established experimental knowledge, indicating alignment with known physical behaviour. This work provides a robust, interpretable tool for multi-stage prediction in heterogeneous materials, with potential applications in mix design.