Machine Learning-Enabled Real-Time Prediction of Drying Shrinkage in Fly Ash-Modified Cementitious Materials Authors: Chi-tathon Kupwiwat, Lapyote Prasittisopin Published: 2026 ABSTRACT Drying shrinkage remains a critical durability challenge in cementitious materials, where conventional prediction methods are limited by low temporal resolution and lack of interpretability. This study presents an interpretable machine learning (ML) framework for real-time prediction of drying shrinkage in fly ash-containing concrete, integrating high-frequency experimental measurements with physically meaningful input features. A data set comprising 79,008 hly observations across 16 mix designs and different curing conditions was used to train and evaluate various ML models. Ensemble and nonlinear models, including Random Forest, Extra Trees, K-Nearest Neighbors, and Multi-Layer Perceptron, achieved high predictive accuracy (R 2 ≈ 0.99) on the test set with low error margins under both hourly and daily temporal representations. Besides, SHapley Additive exPlanations (SHAP) analysis reveals that multiscale temporal variables govern shrinkage evolution, with short-term capillary moisture loss and long-term pore-scale desorption captured through distinct temporal features. Material parameters, particularly fly ash incorporation, exhibit secondary but consistent influence through microstructural refinement mechanisms based on C–S–H phase nucleation and formation. The proposed framework enables a transition from periodic testing toward continuous, data-driven monitoring, offering a pathway for real-time quality control and reduced physical testing of sustainable concrete infrastructures. Introduction Background Cement production contributes ∼7% of global CO2 emissions. Drying shrinkage is a fundamental time-dependent deformation in cementitious materials that directly affects cracking susceptibility, long-term durability, and serviceability of concrete infrastructures. Many concrete structures have functioned in moist environments for many years, with the humidity levels within these concrete structures exposed to the atmosphere fluctuating as a result of rainfall and seepage. , The moisture content in concrete structures plays a crucial role in determining their mechanical and functional properties, including strength, , young’s modulus, thermal and electrical conductivity, − as well as their overall durability. Furthermore, fluctuations in the humidity of concrete structures lead to shrinkage, fatigue, , carbonization, and freeze–thaw cycles, which compromise their integrity and diminish their performance. The movement of water within concrete leads to diffusion of chloride ions , and corrosion of steel reinforcement. , The drying shrinkage process of concrete is characterized by nonhomogeneous and anisotropic behavior, resulting in significant structural adaptation. This phenomenon generates internal stresses within the material, which can lead to the formation of internal cracks. The cracking process engages with the reduced-volume mechanisms present in drying concrete. The strain has a tendency to mitigate internal stresses generated during drying, subsequently leading to a reduction in crack propagation. Conversely, the occurrence of cracks results in a redistribution of the overall stress state, including both internal stresses and those arising from the applied mechanical loads. It is governed by coupled processes including capillary pressure development, moisture diffusion, and microstructural evolution of hydration products such as calcium silicate hydrate (C–S–H). These processes operate across multiple temporal scales, from rapid early-age moisture loss to long-term equilibrium desorption. Their implication indicates that structural effects concurrently influence the measurements of shrinkage in any experimental setup. Real-time shrinkage prediction is important because shrinkage occurs continuously from the curing stage, and if it is neither monitored nor anticipated in that manner, cracking and long-term deformation can develop before any corrective action is possible. Predictive monitoring therefore enables early intervention rather than postdamage assessment, reducing structural risk and inspection effort. However, there is currently no experimental procedure that facilitates the straightforward separation of intrinsic behaviors from structural effects in the context of shrinkage. The process of identifying constitutive laws from experimental measurements is complicated. A robust mechanical model for cracking is essential to accurately compute the structural component. The inherent characteristics of drying can thus be inferred through a form of reverse analysis by concentrating on drying shrinkage. For the microscale level, the movement of water under sustained load enhances the drying process. For the macro-scale level, this results in drying strain. Such negative impacts not only give the poor appearance of concrete buildings and infrastructure from surfaces’ cracking but also lower the structural performance in the long run. Assessing drying shrinkage from tubes or prisms that are allowed to deform is valuable for understanding shrinkage at a material level; however, it is important to recognize that concrete frequently encounters constraints in actual civil engineering applications. , Many standards for assessing drying shrinkage, such as ASTM C596, ASTM C157, and BS EN 12390, periodically evaluate the concrete specimens. The destructive testing conditions may not serve as effective and low-cost predictors for assessing a real concrete infrastructure. , A nondestructive testing (NDT) method for in situ monitoring utilizing sensors presents an alternative approach to forecast concrete shrinkage in both laboratory and field environments. The sensor gauge embedded in the hardened concrete matrix can offer more precise and effective outcomes. Artificial intelligence (AI) and data-driven approaches for forecasting concrete properties and examining concrete images have attracted significant attention in the construction sector. − Typically, the process of testing concrete is characterized by lengthy and expensive experimental methods, frequently necessitating considerable laboratory efforts. Machine learning (ML) algorithms are currently being developed. ML has the capability to analyze complex data, automate processes, and personalize experiences in many industries. In the concrete industry, ML is capable of examining extensive data sets that include diverse concrete mix designs, − material compositions, , and environmental conditions , to uncover intricate patterns and relationships. These ML models demonstrate a high level of accuracy in predicting critical concrete properties, including compressive strength, ,− fresh properties, , creep, and durability. Few studies for ML prediction in concrete drying shrinkage were explored thoroughly. In ref , four ML models including extreme gradient boosting (XGBoost), light gradient boosting (LGB), random forest (RF), and extra tree regressor (ETR) were tested and trained, concluding that ETR can offer the most effective model with R-squared (R 2) = 0.957. In ref , backpropagation neural network (BPNN), support vector machine (SVM), RF, XGBoost, Gaussian process regression, k-nearest neighbor (KNN), linear regression, and long short-term memory were performed, and findings revealed that XGBoost was the most accurate model (R 2 = 0.980). In ref , 12 ML models were run and showed promising results with R 2 up to 99%; however, the work only measured the drying shrinkage for 60 days. Therefore, research should be done on any work that develops an efficient ML model that can accurately forecast drying shrinkage values for both standard and real-time monitoring (NDT) tests over an extended period of time. For other related past literature regarding the ML for predicting drying shrinkage of various cements and concretes, examples include magnesium silicate hydrate cement, geopolymer, low-carbon cement, and ultrahigh-performance concrete. , Recent studies have increasingly incorporated explainable and optimized ML frameworks for predicting the mechanical and durability properties of cementitious materials, including interpretable ML for shrinkage prediction, optimization-driven mixture design, and explainable AI approaches for sustainable concrete systems. − Research Gap, Objective, and Contribution The primary contribution of this work is the development of an interpretable ML framework for high-resolution shrinkage prediction using automated hourly monitoring data. In addition to predictive performance evaluation, the framework investigates interpretable feature relationships associated with temporal evolution, material composition, and environmental conditions under controlled experimental settings. To the best of our knowledge, this is the first study to integrate real-time shrinkage sensing with interpretable ML-based prediction. The novelty of this work is to deliver a continuous, interpretable, and deployable framework for shrinkage monitoring. This represents a shift from conventional discrete testing toward data-driven infrastructure management. Two notable research gaps in this field include:1.The key engineering innovation lies in enabling continuous shrinkage prediction at an hourly resolution, which is not achievable using existing standard methods and empirical models.2.None of the previous studies have addressed real-time monitoring of shrinkage of sustainable concrete using ML predictors, compared to the low-frequency, manual measurements. 1.The key engineering innovation lies in enabling continuous shrinkage prediction at an hourly resolution, which is not achievable using existing standard methods and empirical models. The key engineering innovation lies in enabling continuous shrinkage prediction at an hourly resolution, which is not achievable using existing standard methods and empirical models. 2.None of the previous studies have addressed real-time monitoring of shrinkage of sustainable concrete using ML predictors, compared to the low-frequency, manual measurements. None of the previous studies have addressed real-time monitoring of shrinkage of sustainable concrete using ML predictors, compared to the low-frequency, manual measurements. Conventional manual testing standards require a destructive, labor-intensive, and periodic specimen removal. This real-time framework mitigates the mentioned issues, leading to reduced material waste and lab energy use over a monitoring campaign. It also enables high-resolution predictive modeling of shrinkage behavior and provides mechanism-informed interpretation using explainable ML. Improved shrinkage prediction using AI enables early mitigation, reducing cracking-induced maintenance and associated lifecycle emissions. In this research, the input features used in this study include temporal variables (hour and day), material-related parameters (aggregate type and fly ash condition), environmental conditions (curing temperature and relative humidity), and statistical temporal descriptors (mean, standard deviation, minimum, and maximum shrinkage values), while the output variable corresponds to the measured drying shrinkage values. Methodology The research methods consist of two parts, as shown in the flowchart in Figure : Methodology flowchart. (1) In this work, the term ‘real-time’ refers to automated hourly resolution monitoring using embedded strain-gauge sensing and data logging, representing near real-time monitoring conditions. The real-time monitoring data collection process involved measuring the drying shrinkage of concrete mixtures containing four different aggregates, fly ash, curing temperatures and relative humidity. The ASTM standard testing of four different concrete mixtures was also performed for comparison. It should be noted that the samples’ sizes in both measurements are different. Typically, rectangular sample shape per ASTM standard has higher deformation than the circular shape. This can cause deformation difference on particular sides of the samples, subsequently leading to more unstable deformation than the circular shape. (2) The development and testing process of several ML algorithms of single- and ensemble-based models for predicting drying shrinkage properties. In this part, a correlation analysis was performed to investigate the relationship between features and shrinkage. The input parameters were chosen on the basis of their established physical influence on drying shrinkage. Temporal variables (i.e., hour and day) capture the time-dependent progression of shrinkage, while material-related parameters (i.e., aggregate type, fly ash content, basalt, and gravel) reflect the role of the mix composition. Environmental conditions (curing temperature and relative humidity) were included because of their strong effect on moisture loss and shrinkage development. These features were directly available from our experimental program and consistent with established factors influencing shrinkage reported in the literature. Next, the data were divided into test-train sets, and the models were trained and evaluated. A hyperparameter grid search with 5-fold cross-validation and temporal feature comparison was also performed in this process. After training and evaluation, high-performing models were further analyzed for explainability using Shapley additive exPlanations (SHAP) analysis. Real-Time Data Collection The real-time data collection for drying shrinkage of concrete with different mix designs is depicted in Figure a. The drying shrinkage testing per ASTM C157 is illustrated in Figure b. The mix designs and testing methods performed in this work have the designed compressive strength at 28 days of 24 MPa, as shown in Table . The cement procured is Portland cement Type I [21.2% SiO2; 4.95% Al2O3; CaO 62.8%; Fe2O3 2.8%; MgO 4.0%; SO3 2.63%; Other 1.6%]. Fly ash, an industrial byproduct, was sourced from a coal-fired power plant in northern Thailand. Their specific gravity and fineness were 2.53 and 3200–3600 cm2/g, respectively. The chemical compositions of fly ash are SiO2 (37.3%); Al2O3 (28.6%); CaO (17.9%); Fe2O3 (13.2%); MgO (3.9%); SO3 (3.5%); and other (5.5%). Fly ash was used as a binder replacement at 20% by weight, since it is the optimal range for normal concrete. It is important to note that fresh concrete samples were collected at the mixing stations. The mix designs and the percent replacement of the ready-mix concrete are the common ones for normal strength ready-mix concrete at the batching plants. This 20% replacement level contributes to 0.15–0.20 ton CO2 reduction of 1 ton cement. This is equivalent to 30–40 kg CO2/m3 concrete. Drying shrinkage test: (a) Real-time monitoring and (b) ASTM C157. 1Summary of the Mix Designs and Testing Methods of 24-MPa Concretes Having the Slump at 10–15 cm no.fly ashaggregatecuring condition (°C; %RH)method1no fly ashlimestone23; 50%real time2no fly ashgranite23; 50%real time3no fly ashbasalt23; 50%real time4no fly ashgravel23; 50%real time5with fly ashlimestone23; 50%real time6with fly ashgranite23; 50%real time7with fly ashbasalt23; 50%real time8with fly ashgravel23; 50%real time9no fly ashlimestone30; 60%real time10no fly ashgranite30; 60%real time11no fly ashbasalt30; 60%real time12no fly ashgravel30; 60%real time13with fly ashlimestone30; 60%real time14with fly ashgranite30; 60%real time15with fly ashbasalt30; 60%real time16with fly ashgravel30; 60%real time17no fly ashlimestone23; 50%ASTM18with fly ashlimestone23; 50%ASTM19no fly ashlimestone30; 60%ASTM20with fly ashlimestone30; 60%ASTMaNote: ASTM represents the ASTM C157 testing method. Summary of the Mix Designs and Testing Methods of 24-MPa Concretes Having the Slump at 10–15 cm Note: ASTM represents the ASTM C157 testing method. The strain gauge (P series Polyester wire strain gauge; Tokyo Measuring Instruments Laboratory) for measuring the strains of the samples was first installed inside the cylindrical aluminum molds. This strain gauge has been widely used for concrete applications. It is placed in the middle of the mold. The cables were connected to the data logger. After fresh concrete samples were cast and cured for 24 h, the samples were removed from the aluminum molds. Consequently, the samples were stored in the curing chambers at two different curing conditions for approximately 6 months at 23 °C and 50% RH as well as 30 °C and 60% RH. It is noted that both curing conditions are designed to perform in the laboratory condition as specified by the ASTM standard and in the hot/humid (tropical) climate. The wet environment (e.g., 80–90% RH) is not assessed here since the drying shrinkage of the wet condition is minimal. The drying shrinkages of all concrete samples were recorded at a regular interval. Triplicate samples were assessed for each testing condition. Compared to standard shrinkage measurement, which required removing the samples from chambers during the measurement process, this real-time data collection method is more accurate and error-free, resulting in higher-quality data for ML training. Table summarizes the statistical characteristics of the full real-time shrinkage data set. The temporal variables cover the entire monitoring window, with hour ranging from 0 to 4960 h and day ranging from 1 to 209 days, ensuring complete representation of the drying process. All shrinkage measurements exhibit negative strain values, as expected, with mixture-dependent means typically between approximately −360 and −610 microstrains and standard deviations between 80 and 120 microstrains. The minimum values (up to −686 microstrains) exhibit the highest long-term shrinkage, while maximum values are near 0 microstrains, reflecting early-age readings immediately after demolding. The categorical mixture features are consistent with the 16 mix designs and two curing conditions described in Table , ensuring balanced representation across all experimental conditions. These descriptive statistics confirm that the data set spans the full physical range of shrinkage behavior observed during drying and provides sufficient temporal and material variability for robust ML model training and evaluation. 2Statistical Characteristics of the Full Real-Time Shrinkage Datasetfeatureminmeanmaxstdday1.00105.1720960.26Hr.0.002480.9649601435.52limestone with fly ash (23C-50%)–431.00–368.86078.85limestone with fly ash (30C-60%)–501.00–439.51084.45limestone no fly ash (23C-50%)–531.30–456.120103.68limestone no fly ash (30C-60%)–584.70–514.60098.90basalt with fly ash (23C-50%)–531.70–471.84094.09basalt with fly ash (30C-60%)–603.30–538.200100.62basalt no fly ash (23C-50%)–634.70–565.540115.96basalt no fly ash (30C-60%)–736.70–647.000127.76gravel with fly ash (23C-50%)–466.00–400.57085.75gravel with fly ash (30C-60%)–502.70–440.32087.67gravel no fly ash (23C-50%)–563.70–494.390106.05gravel no fly ash (30C-60%)–592.70–509.310104.82granite with fly ash (23C-50%)–515.00–450.38091.68granite with fly ash (30C-60%)–548.30–488.21093.52granite no fly ash (23C-50%)–651.30–567.650121.34granite no fly ash (30C-60%)–686.70–610.640109.79 Statistical Characteristics of the Full Real-Time Shrinkage Dataset ML Model Development and Evaluation This paper investigates the predictive capabilities of various ML models in estimating concrete drying shrinkage based on the material composition and environmental conditions. The procured data are shown in Table as inputs. The data consist of near real-time experimental measurements capturing concrete drying shrinkage under different material compositions, curing temperature conditions, and the presence of fly ash at the cement replacement levels of 20% by weight. This value is ubiquitously utilized in normal ready-mixed concrete mixtures in batching plants. − The features directly encode experimental design parameters (material composition) and not derived from shrinkage. They describe the concrete composition, which is a fundamental determinant of the shrinkage behavior according to materials science principles. The goal is to develop accurate ML models to predict shrinkage behavior measured up to around 4960 h, as shown in Figure . Past research confirmed the results of this work by using the cylindrical concrete specimens for drying shrinkage measurement under typical conditions. It is also found that a similar manner as using ACI 209 prediction for shrinkages (around 580 microstrains at 214 days or 4960 h) is used. Empirical concrete shrinkage data. The drying shrinkage behavior of concrete can be classified into three progressive stages that are primarily governed by moisture migration and microstructural evolution. The first stage, known as the plastic shrinkage, occurs within the first 2–3 h after water is mixed with cement particles, as the hydration reaction progresses. Then, the early-age rapid shrinkage period (0–200 h in Figure ) is characterized by the evaporation of free water from large capillary pores following demolding and initial curing. During this phase, a rapid volumetric contraction occurs due to steep surface moisture gradients, leading to high internal capillary tension. These tensions initiate the development of microcracks, particularly near the paste–aggregate interfaces where stiffness mismatches concentrate stress. The process is strongly influenced by curing temperature, relative humidity gradients, and the degree of pore connectivity, which collectively determine the rate of moisture loss and the extent of early shrinkage deformation. As the concrete transitions into the intermediate steady-decline period (i.e., 200–2000 h), the rate of shrinkage begins to diminish as the moisture gradients within the material stabilize. In this stage, water molecules continue migrating from finer capillary and gel pores, accompanied by internal redistribution processes that relieve localized stress concentrations. The incorporation of fly ash plays a significant role during this phase, as the ongoing pozzolanic reaction promotes the formation of an additional C–S–H gel. This reaction refines the pore structure, reduces the average pore radius, and enhances their internal humidity buffering capacity. Consequently, the overall shrinkage rate decreases, and the internal stress field becomes more balanced, reflecting a progressive stabilization of the microstructural framework. Finally, the long-term stabilization stage (i.e., 2000–4960 h) marks the point where the moisture content of the concrete achieves equilibrium with the surrounding environment. Only the bound water within the finest gel pores continues to desorb gradually, leading to a near-constant strain response. The shrinkage curve flattens toward its asymptotic value, indicating that most of the capillary-driven moisture movement has ceased. Aggregate stiffness and curing temperature continue to exert minor influences on residual microstrains. However, their effects are marginal compared to the earlier phases. Microstructurally, the pore water molecules become stabilized, and no significant additional deformation is observed. This steady-state condition represents the long-term dimensional stability of the concrete, marking the completion of the drying shrinkage process. As mentioned above, the drying shrinkage measurement in a real-time manner shows that the microstain values during some periods are not stable. This is anticipated for the measuring errors. The paper also aims to analyze the influence of input features that affect the real-time drying shrinkage prediction of each ML model. Data Preparation and Correlation Analysis The data that were measured either in real time or per ASTM standard from Section were stored in a preprocessed Comma Separated Values (CSV) file. Feature extraction was performed to identify relevant attributes, including material type (i.e., limestone, basalt, gravel, granite), the presence of fly ash, and temperature conditions (i.e., 23 and 30 °C). One encoding was applied to categorical variables to enable numerical processing by ML models. Statistical aggregates, such as mean, standard deviation, minimum, and maximum shrinkage values, were calculated based on recorded time (day and hr.) to incorporate temporal trends. Statistical aggregates, including mean, standard deviation, minimum, and maximum shrinkage values, were computed from shrinkage trajectories across all concrete mixtures to represent temporal evolution characteristics of the measured data sets. These descriptors were intended to emulate prior engineering knowledge and historical shrinkage characteristics that may also be obtainable from external databases or empirical standards. All input features are standardized to ensure comparability. This process adjusts each variable by centering it around zero and scaling it to have unit variance. By removing differences in magnitude and distribution across features, the method prevents variables with larger numeric ranges from dominating the learning process. A correlation analysis was conducted to assess the relationships between features and shrinkage values. The heatmap visualization in Figure identifies potential multicollinearity and redundant predictors. The analysis indicates that time data have the largest correlation with concrete shrinkage since the shrinkage nature increases over time. Four types of coarse aggregate mixtures, including basalt, granite, gravel, and limestone, have a smaller correlation, compared to fly ash. Fly ash typically reacts with cement to create pozzolanic (C–S–H) reactions over the long run, while aggregate normally acts as an inert material. ,, The curing temperature shows the smallest correlation with the shrinkage. Lastly, the statistical value of the shrinkage shows somewhat high correlation, indicating the possibility of utilizing ML prediction. Correlation matrix. ML Models and Hyperparameter Tuning Figure presents 10 ML models that are implemented and optimized to predict concrete shrinkage using the input features excluding the shrinkage value. The core models include ridge regression (RR), a linear model that applies L2 regularization to reduce overfitting by penalizing large coefficients; lasso regression (LR), which uses L1 regularization to shrink some coefficients to zero, enabling feature selection; and Bayesian ridge regression (BRR), a probabilistic approach that incorporates prior distributions to improve robustness under noisy and limited data. Ensemble methods such as random forest (RF) and extra tree regressor (ETR) are also employed. The RF model builds multiple decision trees on bootstrapped data to enhance prediction stability, while the ETR model increases model diversity through additional randomness in feature selection and split criteria. The XGBoost model corrects previous errors iteratively to reduce bias, whereas the AdaBoost model emphasizes misclassified instances to improve overall performance. The support vector regression (SVR) model is applied to capture complex relationships through kernel-based mapping of input features into higher-dimensional spaces. The K-nearest neighbors (KNN) model predicted shrinkage based on the average of the k most similar training samples, reflecting local data patterns. Lastly, a multilayer perceptron (MLP) model, a type of the feedforward neural network with hidden layers, is used to model nonlinear interactions among features. The rationale for selecting these ten ML models is provided in this section, and all models follow the Scikit-learn algorithmic formulation. Each ML model is trained using an 80/20 train–test split. The training set contains 63,206 samples, and the testing set contains 15,802 samples. To ensure robustness against variability in data splitting, model training and evaluation are repeated across multiple random seeds. The authors acknowledge that, because of the strong temporal continuity of hourly shrinkage measurements, random splitting may introduce partial temporal information leakage between adjacent observations and may contribute to optimistic predictive performance. However, the data set still includes multiple mixture types, aggregate conditions, fly ash conditions, and curing environments, providing meaningful variability across shrinkage behaviors. Variance inflation factor (VIF) analysis was additionally conducted on the predictor variables prior to ML model training to evaluate multicollinearity among input features. Diagrams of predictive ML models. Hyperparameter optimization is subsequently performed by grid search with 5-fold cross-validation. This is a balanced strategy to ensure model generalization while maintaining computational efficiency, particularly suitable for our large hourly concrete shrinkage data set. The hyperparameter search space is summarized in Table . For instance, tree-based models are tuned over the number of estimators, while learning-based models are tuned over the learning rates and architectural depth. The optimal configurations, exhibited in Table , are selected based on validation root mean squared error (RMSE) scores. 3ML HyperparametersmodelparametervaluesRRalpha[0.1, 1, 10]LRalpha[0.1, 1, 10]BRRalpha_1[1e–6, 1e–5]alpha_2[1e–6, 1e–5]RFn_estimators[50, 100, 200]ETRn_estimators[50, 100, 200]XGBoostn_estimators[50, 100]learning_rate[0.01, 0.1]SVRC[1, 10]kernel[″rbf″, ″linear″]AdaBoostn_estimators[50, 100]learning_rate[0.01, 0.1]KNNn_neighbors[3, 5, 10]MLPhidden_layer_sizes[(64, 64), (128,128)]max_iter[1000, 2000, 4000] ML Hyperparameters 4Optimal Hyperparameters Obtained from Grid SearchmodelparametervaluesRRalpha10LRalpha0.1BRRalpha_11e–6 alpha_21e–6 RFn_estimators200ETRn_estimators200XGBoostn_estimators100learning_rate0.1SVRC10kernel“rbf”AdaBoostn_estimators100learning_rate0.1KNNn_neighbors3MLPhidden_layer_sizes(64,64)max_iter2000 Optimal Hyperparameters Obtained from Grid Search Each ML model is trained using an 80/20 train–test split. The training set contains 63,206 samples, and the testing set contains 15,802 samples. Ten random seeds for training and evaluating the experiments are performed. To ensure robustness for variability in data splitting, model training and evaluation are repeated across 10 random seeds. This approach provides a reliable estimate of model performance and variability without additional computational overhead from nested cross-validation. Because the split was random and repeated over multiple seeds, the distributions of temporal and material features in the training and testing sets remained representative of the full data set. To ensure statistical reliability, each experiment was repeated using 10 predefined random seeds [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. For each seed, both the train–test split and the internal stochastic elements of the algorithms (e.g., parameter initialization) were controlled. Model performance metrics were computed independently for each run and then averaged across 10 repetitions to obtain the final reported values. Standardization of numerical features is performed to maintain consistent scaling across all models. To evaluate the effectiveness of the shrinkage prediction of the ML models, two performance metrics including RMSE and R 2 scores are used to evaluate model effectiveness. RMSE provides an absolute measure of prediction errors, while R 2 quantifies the proportion of variance explained by the ML model. It is noted that overfitting is mitigated through cross-validation, ensuring generalization to unseen data. The equations for RMSE and R 2 are defined as eqs and , respectively. RMSE=1n∑i=1n(yi−y^i)2 1 R2=1−∑i=1n(yi−y^i)2∑i=1n(yi−y−)2 2 where n denotes the number of data. y i , ŷ i , and y− denote the actual shrinkage, predicted shrinkage, and mean shrinkage values, respectively. Results and Discussion To address the research gap concerning ML models for predicting drying shrinkage and to explore their use in real-time monitoring, this paper evaluates 10 ML models on a large-scale data set comprising 79,008 hly observations of concrete shrinkage with different mix designs and curing conditions. The data set includes both temporal features and material properties. Each model is evaluated under two distinct temporal conditions:1.Hourly measurement, simulating real-time monitoring.2.Daily measurement, simulating lab-based and infrastructure-level long-term monitoring. 1.Hourly measurement, simulating real-time monitoring. Hourly measurement, simulating real-time monitoring. 2.Daily measurement, simulating lab-based and infrastructure-level long-term monitoring. Daily measurement, simulating lab-based and infrastructure-level long-term monitoring. Performance on Time Data Table summarizes the maximum (max), average (avg), minimum (min), and standard deviation (std) values of RMSE, normalized RMSE (NRMSE), and R 2 obtained from each ML model at different random seeds. In the table, the four best-performing models are RF, ETR, KNN, MLP, and XGBoost, having R 2 > 0.99, which indicates strong shrinkage predictive performance. Among these models, the KNN model has the lowest average RMSE value (1.1533), reflecting an extremely low error margin. As a result, this model can be well-suited within the practical threshold for shrinkage estimation under both laboratory and field conditions. This precision level is competitive with or higher than those of several previously reported ML-based shrinkage prediction studies under comparable experimental conditions. In contrast, traditional linear models such as RR and LR models, while simpler, failed to capture the complexity of the shrinkage behavior. 5MLs’ Performances on Hourly DatamodelvalueRMSENRMSE R 2 modelvalueRMSENRMSE R 2 RRmax69.19761.86600.712XGBoostmax5.69055.08760.998avg68.09960.88450.700avg5.47674.89650.998min66.42759.38950.690min5.26094.70350.998std0.97660.87310.006std0.14470.12940.000LRmax69.53562.16840.709SVRmax33.18429.66810.938avg68.45561.20210.696avg32.13728.73220.933min66.76359.68950.687min30.78527.52330.930std0.99320.88800.007std0.85680.76600.003BRRmax69.19661.86540.712AdaBoostmax44.87240.11790.886avg68.09860.88290.700avg43.07438.51040.882min66.42859.39010.690min41.537.10320.878std0.97490.87160.006std1.07030.95690.002RFmax1.20911.08101.000KNNmax1.20721.07931.000avg1.18791.06201.000avg1.15331.03111.000min1.15441.03211.000min1.11851.00001.000std0.02020.01810.000std0.02890.02580.000ETRmax1.24621.11421.000MLPmax3.38113.02291.000avg1.22711.09711.000avg3.00982.69090.999min1.19441.06791.000min2.65012.36930.999std0.0190.01700.000std0.2480.22170.000 MLs’ Performances on Hourly Data To evaluate the models under conventional testing scenarios where daily readings are more practical (i.e., in standard ASTM lab protocols and in situ monitoring setups), predictions were made using shrinkage data aggregated by day, as shown in Table . Despite the temporal abstraction, the ranking of models remained largely unchanged. Results revealed that ML models that have R 2 > 0.99 including RF, ETR, KNN, MLP, and GB indicate strong shrinkage predictive performances using daily data. Among these models, the ETR model has the lowest average RMSE (3.2656). The developed models in this work successfully offer highly accurate prediction, compared to the existing models. It is noted that both “hour” and “day” data were included because the real-time shrinkage data recorded every hour are much more frequent than the standard ASTM C157 test. The hourly values capture the short-term changes within each day, while the overall drying process still progresses over longer day-to-day time scales. Using both variables, therefore, reflects these two different aspects of the shrinkage behavior. Although SHAP results in Tables and show that “Hour” has a stronger influence, “Day” still helps represent the longer-term trend and contributes to more stable performance in several ML models (e.g., KNN, MLP, and ETR). For these reasons, keeping both temporal features is appropriate and consistent with the multiscale nature of drying shrinkage. It is noted that despite high predictive accuracy, the model is intended to complement rather than replace physics-based approaches, particularly under extrapolative conditions. 6MLs’ Performances on Daily DatamodelvalueRMSENRMSE R 2 modelvalueRMSENRMSE R 2 RRmax75.4331.32470.703XGBoostmax7.11492.95470.998avg72.18729.97800.675avg6.3492.63660.998min68.64328.50610.657min5.75032.38800.997std2.0780.86300.017std0.37040.15380.000LRmax75.4331.32460.703SVRmax74.92531.11490.739avg72.18729.97780.675avg68.99428.65190.703min68.64828.50850.657min64.4426.76080.661std2.07570.86200.017std3.36051.39560.024BRRmax75.43431.32630.703AdaBoostmax44.87218.63450.896avg72.18829.97850.675avg43.07417.88780.884min68.64428.50640.657min41.517.23420.868std2.07820.86300.017std1.07030.44450.009RFmax4.9742.06560.999KNNmax11.2054.65330.998avg4.53831.88470.999avg8.26563.43260.996min4.01611.66780.998min6.32072.62490.992std0.35430.14710.001std1.7030.70720.002ETRmax4.07811.69361MLPmax6.96492.89240.999avg3.26561.35610.999avg5.33782.21670.998min2.4081.00000.999min3.92211.62880.997std0.50820.21100.001std1.0740.44600.001 MLs’ Performances on Daily Data Figure illustrates the scatter plots comparing actual and predicted shrinkage values for each ML model under the hourly input conditions. The diagonal line represents the ideal 1:1 agreement. High-performing models such as KNN, RF, ETR, and MLP demonstrate tight clustering around this reference line, indicating an excellent predictive alignment with actual measurements. In contrast, linear models (e.g., ridge and lasso) show wider scatter, reflecting their limited capacity to capture complex, nonlinear shrinkage behavior. These visualizations reinforce the numerical results by highlighting how different models handle temporal variability in shrinkage and underscore the superior reliability of nonlinear and ensemble-based models for real-time shrinkage prediction, which was evidenced in refs − . Scatter plots comparing predicted and actual shrinkages by MLs with hourly data. A comparison between hourly resolution and daily aggregated inputs reveals that finer temporal granularity significantly enhances model accuracy, particularly in reducing RMSE. KNN achieves an 86.1% RMSE reduction (from 8.27 to 1.15), while RF, ETR, and MLP show 73.8, 62.4, and 43.6% reductions, respectively. These results indicate that models leveraging local temporal patterns, such as instance-based, ensemble, and neural models, benefit substantially from high-frequency data (i.e., hourly data) that capture the short-term shrinkage dynamic. The finding that the performance of hourly data is better than that of daily data is confirmed by ref . In contrast, improvement in R 2 is relatively small, as many models already explain most of the variance using daily data. For example, tree-based models exceed R 2 = 0.99 in both scenarios, suggesting that hourly data provide only limited additional gain in variance explanation. Nonlinear models like SVR and MLP also show significant RMSE improvement with marginal R 2 gains, reinforcing that hourly inputs enhance precision rather than trend fitting. Figure indicates substantial RMSE gains (blue bars) across top models, while R 2 gains (orange line) remain modest. These results highlight that hourly data are critical for applications requiring precise real-time monitoring and early-age shrinkage detection, even if the overall trend explanation saturates with coarser input. Performance gain from hourly over daily data. SHAP-Based Feature Interpretation SHAP assigns each feature a value representing its marginal impact on the model’s prediction. This allows mechanistic discovery for both global model interpretation and local explanation. Figure presents the SHAP plots for the four top-performing models: KNN, RF, ETR, and MLP, using hourly input data. Each plot ranks features by their average impact on model output (mean absolute SHAP value). Color gradients represent the original feature values (i.e., red = high; blue = low). The horizontal spread reflects how much each feature shifts the prediction across the data sets. SHAP analysis of the four best models using hourly data. For the tree-based models (RF and ETR), time-based features such as “hour” and “day” consistently dominate, reinforcing the importance of high-resolution temporal input in capturing shrinkage dynamics. These models also show moderate contributions from material-related features such as “FlyAsh_with Fly ash” and “Material_Basalt,” confirming their sensitivity to concrete mix composition. In the RF model, “hour” consistently dominates, confirming that short-term temporal resolution is the key driver of shrinkage dynamics, while “day” contributes less. By contrast, the ETR model distributes importance more evenly across both “hour” and “day,” representing its tendency to exploit multiple time scales due to its more randomized split selection. In the KNN model, the feature importance distribution is similar, though slightly flatter. KNN leverages the distance between neighboring data points rather than internal parameter learning. For the MLP model, the SHAP plot reveals a broader spread of influential features, with not only “hour” and “day” but also interaction and environmental variables like curing conditions (i.e., “Temperature_30C” and “Temperature_27C”) and “FlyAsh_with Fly ash” exerting strong nonlinear effects. This indicates the model’s flexibility in capturing complex compound relationships, though at the cost of lower interpretability, compared to tree ensembles. Taken together, these four models reveal consistent patterns. Temporal resolution (especially “hour”) emerges as the dominant predictor across all methods, while the material-related features such as fly ash and basalt show secondary but stable contributions. This cross-model agreement enhances confidence in the reliability of the conclusion regarding feature importance. The dominance of hour temporal input across all high-performing models reflects the governing role of early-stage capillary water evaporation. During the initial drying period (0–200 h), moisture loss occurs primarily from larger capillary pores, generating significant capillary tension. This induces rapid volumetric contraction and drives the majority of early-age shrinkage strain. The strong SHAP contribution of hour therefore captures the kinetics of capillary pressure development and short-term moisture gradients. In contrast, the contribution of day input represents the transition toward diffusion-controlled moisture redistribution within finer pore structures. In this regime (200–2000 h and beyond), shrinkage is governed by gradual desorption of physically bound water from gel pores and the relaxation of internal stress fields. The reduced but persistent importance of day indicates that ML models are capturing the asymptotic stabilization behavior associated with long-term equilibrium moisture conditions. These SHAP results serve several key roles in addressing our research objectives. First, although results confirmed that hour and day as one of the most important features since shrinkage is a time-dependent process, temporal resolution is the most critical factor for accurate shrinkage prediction. It is evident that the hourly input drastically improves the model performance. Hour predominantly captures the short-term moisture loss from capillary pores, which drives rapid early-age shrinkage, whereas day represents the long-term desorption from fine gel pores and gradual stress relaxation, corresponding to the slower stabilization observed experimentally. This cannot be explained when the ASTM test is performed on a weekly basis. Moreover, results ensure model transparency, which is essential for practical adoption in civil engineering, where explainability and domain insight are as important as the accuracy. By linking model decisions to meaningful physical parameters, SHAP transforms the ML models from black boxes into interpretable tools, enabling engineers to trust and act upon their predictions in real-world applications, such as shrinkage monitoring, quality control, and early warning systems. Model robustness was evaluated using cross-validation, yielding stable performance with minimal variance across folds. VIF for all predictor variables used in the models was then calculated. VIF quantifies the degree to which a predictor’s variance is inflated due to linear dependence with other predictors. VIF values ≥5 indicate problematic multicollinearities. Table presents the computed VIF for each input variable. The temporal predictors (hour and day) show very high VIF values (39,619 and 39,598), reflecting their deterministic linear dependence. Similarly, the mixture- and curing-related encoded variables yield high VIF values (5073–37,447) due to their mutually exclusive nature. It is seen that while such magnitudes typically indicate severe multicollinearity in the linear regression, they are not problematic in this study because the ML algorithms, including RF, ETR, KNN, MLP, and XGBoost models, used are nonparametric and do not rely on coefficient estimation. Instead, these ML models learn for splitting rules and local patterns, and these are subsequently unaffected by the linear dependencies among features. Accordingly, Table highlights structural feature dependencies rather than instability in the model process. Overall, results indicate strong generalization within the study domain. 7Variance Inflation Factor (VIF) AnalysisfeatureVIFfeatureVIFday39598.36basalt no fly ash (30C-60%)9711.47hr.39619.34gravel with fly ash (23C-50%)5529.99limestone with fly ash (23C-50%)5073.07gravel with fly ash (30C-60%)27878.58limestone with fly ash (30C-60%)13054.33gravel no fly ash (23C-50%)28934.39limestone no fly ash (23C-50%)35530.81gravel no fly ash (30C-60%)33797.25limestone no fly ash (30C-60%)6794.18granite with fly ash (23C-50%)16027.29basalt with fly ash (23C-50%)19281.62granite with fly ash (30C-60%)27237.72basalt with fly ash (30C-60%)27095.27granite no fly ash (23C-50%)37447.95basalt no fly ash (23C-50%)22114.56granite no fly ash (30C-60%)19139.18 Variance Inflation Factor (VIF) Analysis The current SHAP analysis further confirms that hour and day capture distinct temporal signals. Several models (i.e., KNN, ETR, and MLP) demonstrate improved stability or lower variance when both temporal features are retained, supporting their complementary roles. On this combined basis of quantitative VIF assessment and model interpretability through SHAP, both hour and day parameters can be retained as necessary and nonredundant temporal descriptors that enhance predictive performance across multiple ML models. Discussions and Implications Based on all findings from the work, the ML can be utilized for prediction of the real-time drying shrinkage values of fly ash-containing concrete. This is not only applied for laboratory investigation, but also for the quality control process of concrete manufacturing plants. Many works have been performed to reduce labor force in quality control and quality insurance in batching plants. This can reduce labor cost and reduce testing time, while providing more accurate outcomes. , Moreover, it can be implemented as the SHM system in any infrastructure under service with the aims of reducing human workforces and enhancing smart, sustainable, and resilient cities. , It is assumed that, compared to manual shrinkage testing protocols (i.e., ASTM requiring periodic measurements over 28–90 days), the proposed real-time predictive framework reduces experimental frequency by up to 83% (12 times to 1 time), which corresponds to an estimated reduction of 67% in material consumption (3 sample to 1 sample) and associated laboratory energy demand. This reduction arises from minimizing repeated specimen preparation and long-term monitoring requirements. In addition, improved prediction of shrinkage evolution enables earlier mitigation of cracking risk, thereby reducing maintenance frequency and contributing to lower lifecycle CO2 emissions in cementitious systems. The comparison reveals clear differences in predictive capability, particularly in the early-age regime. ACI 209 produces smooth, monotonic shrinkage curves governed by fixed empirical coefficients, which are unable to capture the rapid nonlinear evolution observed within the first 100–200 h. In contrast, the ML models accurately reproduce these transient dynamics due to their ability to learn high-resolution temporal dependencies from hourly data. This distinction is critical, as early-age shrinkage contributes disproportionately to cracking risk and structural performance. At intermediate and long-term durations, ACI 209 provides reasonable trend approximation but exhibits systematic bias in magnitude and the rate of shrinkage development, particularly for fly ash-modified mixtures. This limitation arises from the model’s inability to explicitly account for microstructural evolution associated with pozzolanic reactions and pore refinement. The ML framework, by contrast, implicitly captures these effects through data-driven relationships between material composition and temporal progression. Quantitatively, the ML models achieve significantly higher predictive accuracy (R 2 ≈ 0.99) compared to the ACI 209 formulation, with substantially lower error dispersion across all time scales. More importantly, however, is the difference in temporal resolution and adaptability. While ACI 209 is constrained to predefined functional forms and discrete input assumptions, the ML approach can continuously predict at hourly intervals and can readily incorporate additional variables and sensor inputs without reformulation. This research illustrates that real-time, explainable ML predictions of concrete shrinkage are not only technically feasible but also practically significant. Despite these key contributions, limitations exist. The generalization of these models to full-scale field environments with greater variability needs further validation, as the data set is derived from controlled laboratory conditions with a common mix design and specific curing conditions. Investigation of validating the framework under field-scale variability and environmental fluctuations should be conducted. The study does not yet incorporate uncertainty quantification, which is essential for assessing the confidence bounds of predictions under data variability and environmental fluctuations. Likewise, a comprehensive sensitivity analysis is not conducted to determine the relative influence of individual input parameters and their potential interactions on the model outcomes. Incorporating domain knowledge such as moisture transport dynamics can improve predictive robustness and interpretability. Conclusions This work establishes an interpretable, mechanism-informed ML framework that transforms shrinkage prediction from a manual laboratory assessment into a real-time engineering tool for monitoring and controlling moisture-driven deformation in concrete. Results can be summarized below:1.Fly ash incorporated in simultaneously reducing binder CO2 emissions and improving shrinkage performance, which is accurately validated by ML-driven analysis.2.Hourly input data significantly improved prediction accuracy, particularly in reducing the RMSE. Top-performing models such as KNN, RF, ETR, and MLP achieved R 2 values of 0.99, with RMSE values under 1.2 microstrain. This indicates a high degree of precision suitable for real-time applications in concrete monitoring.3.While gains in R 2 were modest, the reduction in RMSE confirmed that a finer temporal resolution enhanced local accuracy.4.SHAP analysis strengthened the findings by revealing that temporal features consistently dominated model predictions, while material properties, such as fly ash content and aggregate type, played a critical role.5.These insights validated the experimental design and highlighted the potential for integrating interpretable ML models into engineering workflows for transparent and trustworthy decision making. 1.Fly ash incorporated in simultaneously reducing binder CO2 emissions and improving shrinkage performance, which is accurately validated by ML-driven analysis. Fly ash incorporated in simultaneously reducing binder CO2 emissions and improving shrinkage performance, which is accurately validated by ML-driven analysis. 2.Hourly input data significantly improved prediction accuracy, particularly in reducing the RMSE. Top-performing models such as KNN, RF, ETR, and MLP achieved R 2 values of 0.99, with RMSE values under 1.2 microstrain. This indicates a high degree of precision suitable for real-time applications in concrete monitoring. Hourly input data significantly improved prediction accuracy, particularly in reducing the RMSE. Top-performing models such as KNN, RF, ETR, and MLP achieved R 2 values of 0.99, with RMSE values under 1.2 microstrain. This indicates a high degree of precision suitable for real-time applications in concrete monitoring. 3.While gains in R 2 were modest, the reduction in RMSE confirmed that a finer temporal resolution enhanced local accuracy. While gains in R 2 were modest, the reduction in RMSE confirmed that a finer temporal resolution enhanced local accuracy. 4.SHAP analysis strengthened the findings by revealing that temporal features consistently dominated model predictions, while material properties, such as fly ash content and aggregate type, played a critical role. SHAP analysis strengthened the findings by revealing that temporal features consistently dominated model predictions, while material properties, such as fly ash content and aggregate type, played a critical role. 5.These insights validated the experimental design and highlighted the potential for integrating interpretable ML models into engineering workflows for transparent and trustworthy decision making. These insights validated the experimental design and highlighted the potential for integrating interpretable ML models into engineering workflows for transparent and trustworthy decision making. Despite these key contributions, limitations exist. The present data set is derived from controlled laboratory conditions using Portland Type I cement, one fly ash source, 20% fly ash replacement level, and 24-MPa concrete mixtures. Nevertheless, the data set still includes multiple aggregate types, curing environments, and mixture configurations, allowing the ML models to learn across different shrinkage patterns. Future work should extend the framework toward broader concrete classes, additional supplementary cementitious materials, and field-scale validation. Note that the present validation framework is based on repeated random train–test splitting of high-frequency shrinkage observations. Because temporally adjacent measurements from the same specimen may appear in both training and testing data sets, partial temporal information leakage may occur. Future work should investigate stricter temporal and specimen-level validation strategies to further assess the deployment-level generalization. This paper demonstrates that real-time, explainable ML prediction of concrete shrinkage is both technically viable and practically impactful for the pathway toward smart, low-carbon infrastructure management. Future work will focus on extending this framework for deployment for other low- and high-strength as well as special concrete mix designs and integrating it with sensor-based monitoring systems to enable intelligent, adaptive concrete management. Also, uncertainty bounds for predictions, sensitivity analysis, and physics-informed neural networks (PINNs), which rely on physical and chemical mechanisms of concrete, shall be further investigated. Also, in this research, statistical descriptors were incorporated to represent overall temporal evolution characteristics across different mixture configurations and curing conditions while also emulating prior shrinkage knowledge obtainable from historical databases and empirical standards. Nevertheless, because these descriptors were derived from measured shrinkage trajectories within the present data set and combined with random train–test splitting, partial temporal information leakage may occur. Future work should investigate stricter temporal and specimen-level validation together with externally available prior descriptors.