The rapid urbanization in China significantly increases a demand for concrete, driving the development and application of recycled aggregate concrete (RAC) to reduce natural resource consumption and mitigate land pollution from construction waste. The RAC is produced via crushing, screening, and reusing demolished concrete as partial or full replacements for natural aggregates. However, its creep behavior exhibits higher deformation than conventional concrete due to the attached old mortar on recycled aggregates, which weakens interfacial transition zones (ITZs) and increases porosity. This review represents the influencing factors, mechanisms, and prediction models of RAC creep, aiming to guide its engineering applications.

In this review, the creep characteristics and influencing factors of RAC are summarized and discussed from three aspects (i.e., material composition, mix design and environmental factors). The material composition consists of recycled aggregate quality, binder system and parent concrete. Residual mortar content (20%–40%), water absorption (3%–8%), and microcracks in aggregates are exacerbated by 30%–50%. Some techniques like mechanical shaping (e.g., high-speed self-impact to remove weak mortar) and chemical strengthening (e.g., polymers, CO₂ treatment) can reduced by 26.7%. For a binder system, mineral admixtures (e.g., fly ash, GGBS, silica fume) optimize hydration products and ITZs. Fly ash (20%–30%) reduces a long-term creep by 0.25–0.35 in coefficient, while excessive dosage (>50%) increases it. A lower parent concrete strength (e.g., water-cement ratio >0.6) increases RAC creep by 33%. The carbonation of parent concrete alters pore structures, accelerating moisture migration at the RH of <60%. For a mix design, the water-cement ratio, recycled aggregate replacement ratio and mixing methods are discussed. The existing research indicates that saturated surface-dry or additional water method compensates aggregate absorption mitigating creep. Conventional recycled aggregate replacement methods increase a creep by 6.5%, whereas the “equivalent mortar volume” method reduces it by 23.3% via balancing mortar content. The two-stage mixing (TSMA) and ultrasonic vibration enhance the ITZ density, lowering a creep by 4.7%. The environmental factors on the RAC creep of humidity, temperature, CO₂ curing and carbonation effects are summarized. The RH reduction from 70% to 30% increases a creep by 37.7% due to moisture loss in porous aggregates. Elevated temperature accelerates the ITZ microcrack propagation, especially in freeze–thaw cycles. CO₂ curing densifies ITZs via CaCO₃ formation, reducing creep. Finally, the maximum information coefficient (MIC) analysis is conducted, ranking water-cement ratio (MIC=0.476), aggregate absorption (0.454), and loading age (0.441) as dominant effects, and the cement type (0.193) and specimen size (0.213) are less significant.

This review also discusses the microscale creep mechanisms. The RAC’s creep stems from viscous flow, dual ITZs, moisture diffusion and elastic modulus mismatch according to the results of the existing research. The viscous flow is caused by the stress-induced water migration in C-S-H gels of old mortar. More types of ITZ exist in RAC micro-structures, and weak zones between old/new mortar and aggregates promote microcracking (The SEM images reveal 2–3 times higher porosity in ITZs). A high water absorption (3%–10%) of recycled aggregate creates internal humidity gradients, also coupling shrinkage and creep. Recycled aggregates present a lower stiffness (i.e., 20%–40% reduction compared to natural aggregates), which transfers stress to the matrix, amplifying deformation.

The predicted models consisting of empirical models, classical models, code-oriented models, theoretical aging models and machine learning algorithms prediction are summarized. The conventional prediction models, such as ACI 209 and CEB-FIP MC90, rely on regression-based equations, but struggle to account for the complexity of recycled aggregate concrete (RAC). The theoretical models like B3/B4 incorporate aging theory, but still require a recalibration for RAC, such as introducing residual mortar coefficients. These models have limitations, including narrow parameter coverage (3–5 factors), reliance on linear assumptions, and poor long-term prediction accuracy (R²~0.7). In contrast, machine learning (ML) models (e.g., XGBoost, CNN, LGBM) achieve significantly a higher performance, with R² > 0.95, compared to 0.6–0.8 for the conventional models. The XGBoost demonstrates the optimum accuracy (i.e., R²=0.989, RMSE=0.117). The ML key advantages include handling nonlinear interactions among 10+ parameters (e.g., the SHAP analysis) and improving stability through CNN’s residual encoder. However, some challenges remain, such as high data dependency (requiring 23300+ samples), overfitting risks, and limited interpretability, due to their “black-box” nature.

This review highlights the complexities of creep in RAC and proposes key strategies for optimization and innovation. Material optimization can be achieved via controlling residual mortar (<20%), incorporating supplementary cementitious materials (SCMs) like 30% GGBS, and adopting CO₂ curing to mitigate creep. Multiscale modeling, combining molecular dynamics with FEM, can simulate interfacial transition zone (ITZ) effects. Machine learning (ML) advancements should focus on expanding databases to include load history, environmental cycles, and admixtures, while hybrid models (e.g., physics-informed neural networks) can enhance both accuracy and interpretability. Sustainable solutions like bio-chemical ITZ self-healing and digital twins for lifecycle prediction offer promising future directions. Overall, integrating ML-enhanced, data-driven approaches enable safer, large-scale RAC applications, supporting greener construction practices.

Ultra-high performance concrete is an advanced cement-based material with high strength and high durability. Due to the low water-binder ratio and high binder consumption of UHPC, the early shrinkage of UHPC during the setting and hardening process is larger than that of conventional concrete and high-strength concrete, which easily leads to the cracking of engineering structures and affects the performance of structures. Therefore, it is of great significance to study the early shrinkage characteristics of UHPC for the optimization of the material and the prediction of early cracking. In this paper, the effects of binder-sand ratio, water-binder ratio, different fiber types and contents and curing environment on the early shrinkage performance of UHPC were investigated, and the early shrinkage model of UHPC was established by combining BPNN and WOA-BPNN neural network.

In this study, a total of seven groups of specimens were set up, considering four control factors, i.e., the cement-sand ratio, water-binder ratio, curing environment and polypropylene fiber volume content. Each group consisted of four specimens with size of 25 mm×25 mm×280 mm (including three specimens for drying-shrinkage tests and one specimen for autogenous-shrinkage tests). The specimens were cured in a curing box with a temperature of (25±2) ℃ and a relative humidity of (98%±2%). At the same time, a control group was set up, which was cured in a natural environment with a temperature of (28±5) ℃ and a relative humidity of (60%±15%). The shrinkage deformation of the specimens was measured by a specific length meter at 0, 1, 2, 3, 4, 6, 8, 10, 12 h and 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 5.0, 6.0, 7.0, 9.0, 11.0, 14.0, 18.0, 21.0, 25.0, 28.0 d. Combined with BPNN and WOA-BPNN machine learning models, the measured data were trained with small samples, and finally a neural network model that can be used to predict the early drying and autogenous shrinkage performance of UHPC was obtained.

The early drying-shrinkage and autogenous-shrinkage of UHPC increased with an increase of the cement-sand ratio, and the MIC values of drying shrinkage and cement-sand ratio also increased. The early drying-shrinkage and autogenous-shrinkage of specimen with a cement-sand ratio of 1.2 increased by 108% and 60%, respectively, compared with specimen with a cement-sand ratio of 0.8. The early drying-shrinkage and autogenous-shrinkage of UHPC increased with a decrease of the water-binder ratio. The appropriate amount of polypropylene fiber and steel fiber mixture would limit the early shrinkage of UHPC. With an increase of polypropylene fiber content, the inhibitory effect on both the early drying-shrinkage and autogenous-shrinkage grew firstly and then weakened. The specimens with fiber content of 0.10 % exhibit the best inhibitory effect. The drying-shrinkage expressed a great correlation with the curing method, and the MIC value was 0.56. The correlation between the autogenous-shrinkage and curing method is small, and the MIC value is 0.27. The drying-shrinkage rate and self-shrinkage rate under dry curing conditions were 2.5 times and 1.2 times that under standard curing conditions, respectively.

The two machine learning (ML) algorithms were used in the shrinkage prediction and exhibit good accuracy. The WOA-BPNN algorithm expressed delightful predicted accuracy, whose R², RMSE and MAE for drying-shrinkage model are 0.959, 0.050 and 0.040, respectively, and R², RMSE and MAE for autogenous-shrinkage model are 0.896, 0.076 and 0.053, respectively. The predicted results indicated that the whale optimization algorithm could improve the ML model effectively.

The main conclusions of this paper are given as follows: 1) The fine aggregate has an inhibitory effect on the early drying-shrinkage and autogenous-shrinkage of UHPC. When the cement-sand ratio decreases from 1.2 to 0.8, the early drying-shrinkage and autogenous-shrinkage increase by 108% and 60%, respectively. 2) The early drying-shrinkage and autogenous-shrinkage of UHPC increase with a decrease of water-binder ratio, and the decrease of water-binder ratio leads to the advance of UHPC self-drying phenomenon. 3) The research found that the polypropylene fiber volume content of 0.10% exhibit the best inhibitory effect on the UHPC shrinkage. 4) The curing method had a great influence on its early drying-shrinkage and autogenous-shrinkage. The drying-shrinkage rate of specimens under dry curing condition is 2.5 times that under standard curing condition, and the autogenous-shrinkage rate of specimens under dry curing condition is 1.2 times that under standard curing condition. The early drying-shrinkage and autogenous-shrinkage predicted results of UHPC based on the WOA-BPNN neural network show nice accuracy and robustness compared to that of BPNN.