Reincarnate is pleased to share the publication of the paper “Meta-Learning for Adaptive Mix Design of Alkali-Activated Concrete”, issued in the Proceedings of the RILEM Spring Convention and Conference 2025 (RSCC 2025) and included in the RILEM Bookseries, Volume 66.
Authored by Ghezal Ahmad Jan Zia, Benjamín Moreno Torres, Ahmad Rashid Hazem, Ravi Patel, and Sabine Kruschwitz, the paper presents a novel meta-learning approach designed to overcome one of the central challenges in developing Alkali-Activated Concrete (AAC): generating accurate compressive-strength predictions when experimental data are limited.
Conventional machine learning techniques depend heavily on large datasets and often fail to generalize across AAC mixes that use different precursors, activators, and curing conditions. This study introduces a technique based on Model-Agnostic Meta-Learning (MAML) and Reptile, enabling predictive models that can rapidly adapt to new AAC formulations using only a few available samples.
By defining tasks around AAC material properties and curing regimes, the meta-learning process identifies an optimal initialization that supports few-shot learning and efficient fine-tuning. The results demonstrate that the approach significantly improves adaptability and generalization, offering a viable pathway for real-time strength prediction in both research and industrial practice. This can reduce laboratory workload, accelerate mix-design optimization, and support more sustainable material development workflows.
This research aligns with Reincarnate’s mission to advance data-driven circular construction practices, enabling more efficient use of resources and accelerating innovation in low-carbon materials like AAC.