What question did this study set out to answer?

To develop an explainable machine learning framework for predicting drying shrinkage in treated recycled aggregate concrete.

April 1, 2026

Explainable machine learning framework for drying shrinkage prediction in treated recycled aggregate concrete

Key Points

To develop an explainable machine learning framework for predicting drying shrinkage in treated recycled aggregate concrete.
Compiled a database of 460 concrete mixes from peer-reviewed studies.
Optimized five machine learning algorithms integrated with explainable techniques.
Conducted explainability analyses to assess factors influencing shrinkage.
Developed a graphical user interface for real-time scenario analysis.
Extreme Gradient Boosting achieved R² = 0.974 and reduced prediction errors by 55%.
Curing age and natural aggregate content explained over 85% of shrinkage variance.
Mechanical and thermal treatments are effective up to ~400 kg/m³ replacement levels.
Chemical treatments are recommended to be limited to <700 kg/m³.
Framework validated against international design standards for immediate implementation.

Abstract

Abstract This study presents the first explainable machine learning (ML) framework for predicting drying shrinkage in treated recycled aggregate concrete, addressing a critical knowledge gap hindering structural applications. A comprehensive database of 460 concrete mixes from 30 peer‐reviewed studies (2015–2024) was compiled, encompassing seven recycled concrete aggregates (RCA) treatment strategies and replacement levels of 25%–100%. Five ML algorithms were optimized and integrated with explainable Absorption Index techniques (SHapley Additive Explanations, Local Interpretable Model‐Agnostic Explanations, partial dependence plots) to provide both predictive accuracy and mechanistic transparency. Extreme Gradient Boosting achieved exceptional performance ( R 2 = 0.974, root mean square error = 12.25 με, mean absolute error = 9.29 με), reducing prediction errors by 55% compared to classical formulations (ACI 209R, B3, Comité Euro‐International du Béton – Fédération Internationale de la Précontrainte CEB‐FIP) while maintaining errors below experimental measurement uncertainty (±20–30 με). Explainability analyses revealed that curing age and natural aggregate content dominate shrinkage evolution (>85% variance), while treatment‐specific thresholds were identified: mechanical and thermal treatments remain effective up to ~400 kg/m 3 (~50% replacement), whereas chemical treatments should be limited to <700 kg/m 3 . A graphical user interface enables real‐time scenario analysis for performance‐based mix optimization. Unlike traditional empirical codes, this framework provides instance‐level explanations of how specific parameters influence shrinkage, transforming treated recycled aggregates from waste substitutes into engineered materials. The system achieves shrinkage prediction within ±5 με of experimental observations for 50% RCA replacement mixtures, demonstrating that treated recycled aggregates can match conventional concrete performance while reducing embodied CO 2 by 12%–18% and diverting 400–600 million tons of construction waste annually if adopted globally. The framework is validated against international design standards (ACI 209R, Eurocode 2, IS 1343), enabling immediate implementation in structural engineering practice.

Perguntar à IA

Bookmark