Introduction

Engineered cementitious composites (ECCs) represent a class of ultra-high-performance cementitious composites that have significantly transformed the landscape of civil engineering. Since its inception by Li and colleagues in 1992, ECC has garnered substantial interest owing to its notable deformability and superior crack control capabilities1,2. Numerous experimental studies on mechanical properties of ECC have been conducted worldwide, such as uniaxial compression3, uniaxial tension4, flexural5, shear6, fatigue7 and impact8, as well as a large number of structural or engineering applications9,10,11,12,13,14,15. The importance of comprehending and forecasting the mechanical characteristics of ECCs cannot be overstressed, as they directly influence the design and functionality of civil engineering structures.

Prior studies have often necessitated numerous experimental trials to attain an optimal ECC mix ratio. This process not only consumes considerable time but also demands substantial financial resources. Li introduced strength and energy criteria as a means to ensure ECC materials achieve a stable state of cracking, thereby mitigating potential errors in testing16. Based on numerous test results, empirical ranges for the strength and energy criteria have been established for various types of fibers16,17. The double criteria are generally obtained by means of a three-point flexural test and a single-crack tensile test18,19. However, the double criterion can only be used to determine whether a given mixture ratio can attain multiple cracks and cannot predict the mechanical properties of the mixture ratio.

Machine learning belongs to a sub-direction of artificial intelligence, including supervised and unsupervised20. Machine learning has a wide range of applications in the strength prediction of civil engineering materials. Hu used six machine learning models to predict the compressive strength of FRP-concrete confined column and compared them with prevailing mathematical or empirical models. SHAP analysis was utilized to ascertain feature importance and model sensitivity. The findings indicate that the XGB model exhibited superior robustness and accuracy21. At the same time, the problem of small experimental datasets was solved by using an innovative method of data augmentation, facilitating the achievement of a higher prediction accuracy for the punching shear strength of steel fiber-reinforced concrete slabs22. Jayasinghe used eight machine-learning models to develop a framework for predicting the shear capacity of recycled aggregate concrete (RAC) beams, and the results revealed that the XGBoost model had the best prediction effect, with the coefficient of determination R2 on the test set reaching 0.95 (slender beam) and 0.78 (deep beam)23. You utilized 48 datasets and three machine learning models to predict the bond strength between UHPC and deformed reinforcing bars. Among these models, the Adaptive Neuro-Fuzzy Interface System (ANFIS) emerged as the most accurate prediction model. The coefficient of determination R2 on the test set reached 0.9824. Similarly, Zhang predicted FRP and concrete interfacial bond strength with six machine learning models. Here, the XGBoost model performed best, and the root mean square error was reduced by 73% compared with the original empirical formula25. Liu predicted the creep compliance of concrete with four machine-learning models and found that the XGBoost was the optimal prediction model, with the R2 on the test set reaching 92%26. Mai predicted the compressive strength of fiber-reinforced self-compacting concrete with three machine-learning models. Among the three models, the categorical gradient boosting (CGB) model was the best, and the coefficient of determination was 0.986 for the test dataset27. Zhu used an artificial neural network (ANN) and support vector machine (SVM) to predict the 7-day compressive strength of ultrahigh-performance concrete (UHPC). The findings indicate that the ANN model yields superior prediction results compared to the SVM model28.

Several scholars have recently used machine learning to predict ECC strength. For instance, Shanmugasundaram used the ANN model to predict the compressive strength of ECC mixed with fly ash and GGBS, respectively. Here, the minimum and maximum differences between the predicted and experimental values were 0.317% and 11.4%, respectively29. Mahjoubi employed three machine learning models to predict the compressive strength, tensile strength, and ductility of SHCC. The results showed that XGBoost performed best, and the coefficients of determination on the test set were 0.95, 0.97 and 0.93, respectively30. Meng employed an artificial neural network algorithm to forecast both the compressive strength and tensile strength of PVA-ECC. The coefficients of determination on the test set were notably high at 0.98 for compressive strength and 0.99 for tensile strength, indicating significant prediction accuracy. However, the dataset in the described study was relatively small, with compressive strength ranging from 21.3 to 75.2 MPa and tensile strength ranging from 1.56 to 5.81 MPa31.

Moreover, the dataset used for machine learning is limited in scope, as it does not encompass high-strength and ultra-high-strength datasets. The algorithms with better results are XGBoost and artificial neural networks. However, artificial neural networks are prone to overfitting, and their training process is intricate. Despite this, there is a noticeable scarcity of predictions for the tensile strain capacity and flexural strength of ECC materials. Existing studies primarily focus on predicting compressive and tensile strength, predominantly within the PVA-ECC domain. The preparation of ECC may use PE fibers, PP fibers, and PVA fibers, which are also divided into oil-film PVA fibers and oil-free PVA fibers. The influence of specimen size on ECC mechanical properties remains unexplored in machine learning literature, indicating a gap in comprehensive understanding within prior research.

Based on a more extensive range of data sets, the aim of the present study was to predict critical mechanical properties of ECC, including compressive strength, flexural strength, tensile strength, and tensile strain capacity. At the same time, the importance of the characteristics affecting the mechanical properties of ECC was explored. Finally, a design strategy for adjusting ECC components was proposed according to the specified mechanical properties.

Data processing

Data collection and preprocessing

In the present study, the dataset uniformly sets the cement ratio as 1, while expressing the proportions of other components relative to cement. To use machine learning to predict the mechanical properties of ECC, a total of 1532 sets of data were collected from the existing literature3,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68, including 501 sets of compressive strength, 224 sets of flexural strength, 404 sets of tensile strength, and 403 sets of tensile strain capacity. The experimental group containing CGP was completed by the present author. Owing to space limitations, the experiment details are not provided, but the data can be published, and the relevant experiments are to be published in other articles. There were a total of 13 input features and four output labels. The selection of input features was based on the research conclusions of the literature referenced in the database. The input features were water-cement ratio (W), silica fume (SF), slag (SG), fly ash (FA), coal gangue powder (CGP), silica sand (SS), fiber content (FCO), water reducer (WR), fiber class (FC), fiber aspect ratio (FAR), compressive specimen size (CSS), cement grade (CG), and flexural specimen size (FSS). The output labels were compressive strength (CS: 7.4–166 MPa), flexural strength (FS: 2.6–34.6 MPa), tensile strength (TS: 0.95–18 MPa), and tensile strain capacity (TSC: 0.002% ~ 12.3%). Among them, FC was divided into surface oil-coating polyvinyl alcohol fiber (PVA-O), surface oil-free polyvinyl alcohol fiber (PVA-UO), ultra-high molecular weight polyethylene fiber (PE), fiber-free (NF) (There was a test group without fiber to analyze the effect of fiber), and polypropylene fiber (PP); CG was divided into PO.42.5 and PO.52.5; CSS was divided into 50 mm, 70 mm, and 100 mm; FSS was divided into three cases: four-point flexural plate (PLATE), three-point flexural beam (3-XL), and four-point flexural beam (4-DL). Therefore, 19 input features were included in the present study. The test conditions from which the data set was derived needed to be specified. In the present study, only the datasets under 28-day standard curing conditions were considered, utilizing variable speed mixing methods. However, the input features did not encompass properties of raw materials, such as particle size distribution, particle shape, and chemical composition, which could potentially have influenced the mechanical properties of ECC. Nevertheless, these factors were omitted due to the challenge of quantitatively describing and simplifying them within the machine learning model.

The machine learning data set was derived from the reference literature, and the mix ratios with missing values were excluded when collecting data. Before training the model, the classification features were processed by means of one-hot encoding, and the input features of the non-tree model were normalized using min–max scaling.

Data analysis

Figure 1 shows the relationship of all continuous characteristic parameters and compressive strength. Figure 2 shows the correlation heat map between all features (including classification features). However, the main focus was primarily on the correlation between each feature and the mechanical properties. Correlations with coefficients between −0.4 and 0.4 were disregarded as they were considered insignificant for the analysis26. According to Figs. 1 and 2, there was a negative correlation between (W, FA) and CS and a positive correlation between (CG, SF) and CS; a negative correlation between FAR and FS; a positive correlation between W and FS; a negative correlation between W and TS; and a positive correlation between (FAR, FC, CG, SF) and TS. There was no negative correlation between feature parameters and TSC, but a positive correlation between (FAR, FC) and TSC.

Figure 1
figure 1figure 1

The relationship between features and mechanical properties [(a) CS, (b) FS, (c) TS, (d) TSC].

Full size image
Figure 2
figure 2

Correlation heat map between all features [(a) CS, (b) FS, (c) TS, (d) TSC].

Full size image

Methodology

The technology route of the present study can mainly be divided into three steps: the first step is data collection and division. The training set accounted for 80%, and the test set accounted for 20%. In the second step, the tenfold cross-validation was used to optimize the hyperparameters and reduce overfitting, as shown in Fig. 3. The third step involved predicting the test set and analyzing the importance of features. Figure 4 shows the specific technology route.

Figure 3
figure 3

10-fold cross-validation.

Full size image
Figure 4
figure 4

Machine learning technology route for the prediction of ECC mechanical properties.

Full size image

Machine learning

Machine learning models

Numerous machine learning algorithms are available for predicting the mechanical properties of concrete. In the present study, four representative algorithms were employed to predict the mechanical properties of ECC: linear regression (LR), K nearest neighbors (KNN), random forest (RF), and extreme gradient boosting (XGB).

LR model

LR is the simplest and most basic supervised learning algorithm in machine learning. It has good interpretability and is widely used in engineering. The calculating function can be expressed as follows:

hθ(x1, x2, …, xn) = θ0 + θ1x + … + θnxn, hθ(x1, x2, …, xn) is a mechanical indicator, θ0, θ1, ... θn are regression parameters, and x1, x2, …, xn are characteristic parameters.

KNN model

KNN is also a simple algorithm in machine learning. It calculates the K nearest sample points based on Euclidean distance and derives the final result by averaging these K sample points.

Supposing there are two points P and Q, P = {p1, p2, … pn}, Q = {q1, q2, … qn}, n = 1, 2, 3, …, the Euclidean distance between P and Q is \(d = \sqrt {(p_{1} - q_{1} )^{2} + (p_{2} - q_{1} )^{2} + \cdots + (p_{n} - q_{n} )^{2} }\).

RF model

RF is an ensemble algorithm belonging to the tree model. A random forest comprises multiple decision trees, and there is no correlation between each decision tree in the forest69. The main advantages of random forests lie in their ability to handle high-dimensional data without the need for dimensionality reduction or feature selection. Additionally, they can provide insight into feature importance. However, one disadvantage is that random forest models are generally less powerful for regression problems compared to classification problems.

The working mechanism of the RF algorithm is as follows: firstly, M samples (M ≤ N) are randomly drawn from the training set with a size of N for each tree; if a total of K times are randomly selected, then the randomly selected training set is put into K trees for separate training; in the final prediction, the final output of the test set is determined by each decision tree in the forest. For the regression problem, the mean output values of all decision trees are used as the final output. Figure 5 shows the learning flow chart of RF.

Figure 5
figure 5

Process of random forest algorithm.

Full size image

XGB model

An essential difference between XGB and RF lies in how their decision trees are constructed. In XGB, the decision trees are interrelated, offering certain advantages. The XGB algorithm boasts high accuracy and rapid training speed, and it exhibits robustness to missing values. Nevertheless, it comes with the drawback of high space complexity, necessitating more storage memory70.

The working mechanism of the XGB algorithm is as follows: firstly, a weak learner W1 is trained on the training set by using the initial weight, and the weight of the training sample is updated according to the residual error left by the weak learner 1. The weight is increased when the residual is large, and the weight is reduced when the residual is small. After updating the weights based on the residuals left by the weak learner 1, the new adjusted weights are used to train the weak learner 2. The described process is repeated until the number of weak learners reaches the preset value. Finally, all weak learners are integrated through a set strategy to obtain the final strong learner. Figure 6 shows the learning flow chart of XGB.

Figure 6
figure 6

Process of XGB algorithm.

Full size image

Evaluation indicators

In the present study, the four evaluation indications most widely used in machine learning were adopted to compare the prediction accuracy of different models: mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). The detailed expressions are shown in Eqs. (1–4), representing the predicted value, actual value, and the average of the actual value.

$$ MSE = \frac{1}{m}\sum\limits_{i = 1}^{m} {(y_{i} - \hat{y}_{i} )^{2} } $$
(1)
$$ RMSE = \sqrt {\frac{1}{m}\sum\limits_{i = 1}^{m} {(y_{i} - \hat{y}_{i} )^{2} } } $$
(2)
$$ MAE = \frac{1}{m}\sum\limits_{i = 1}^{m} {\left| {(y_{i} - \hat{y}_{i} )} \right|} $$
(3)
$$ R^{2} = 1 - \frac{{\sum (\hat{y}_{i} - y_{i} )}}{{\sum (\hat{y}_{i} - \overline{y}_{i} )}} $$
(4)

Results and discussion

Compressive strength

Table 1 lists the optimal parameter values of the model after hyperparameter optimization, and the definition of each parameter is also shown, in which the LR model has no hyperparameter. Table 2 shows the three evaluation indicator values of the four models after hyperparameter optimization on the training and test sets. An observation can be made that the R2 values of the LR model on the train and test set were 0.72 and 0.70, respectively, which was the lowest among the four models, while the RMSE and MAE of the LR model on the training and test set were the largest. These indications suggest that the LR model is experiencing underfitting. The R2 of the KNN model was significantly higher than that of the LR model, and the RMSE and MAE values were also significantly reduced.

Table 1 The optimal parameter values of the model (CS).
Full size table
Table 2 Three evaluation indicators on the training set and the test set (CS).
Full size table

The RF and XGB models were superior to the LR and KNN models. The RF model had the best generalization ability, and its R2 was the highest among the four models, at 0.99 and 0.92 on the training and test sets, respectively. The XGB model exhibited signs of overfitting, evident from the R2 of 1.00 on the training set and R2 of 0.89 on the test set. As such, the XGB model was deemed inappropriate under the division of the compressive strength data set in the present study.

From the R2 comparison of the aforementioned models on the training set and the test set, it can be found (Table 2) that the R2 values of the LR and KNN models on the two data sets were small. Both models exhibited underfitting in compressive strength prediction. The LR model reached its limit state as it lacked hyperparameters to optimize, and thus was unable to enhance its accuracy in predicting the compressive strength of ECC further. While the KNN model's prediction accuracy for compressive strength may not have matched that of XGB and RF models, it offered simplicity with only one hyperparameter, resulting in faster training speed. Despite this, the KNN model demonstrated acceptable performance on the test set. Overall, in terms of compressive strength prediction accuracy, the RF model emerged as the top performer.

Figure 7 shows the predicted and actual compressive strength values on the test set. The predicted values of the LR model, KNN model, XGB model, and RF model gradually became more aggregated on both sides of the straight line, indicating that the RF model has the best generalization level on the compressive strength test set.

Figure 7
figure 7

Comparison between predicted value and actual value of CS [(a) LR, (b) KNN, (c) RF, (d) XGB].

Full size image

While ensemble learning methods often exhibit exceptional accuracy, they lack transparency in explaining their output, leading to what is known as the “black box” of machine learning. To address this challenge, the SHapley Additive exPlanations (SHAP) approach was introduced. SHAP is a mathematical framework developed to elucidate the underlying prediction mechanisms of machine learning models. This methodology originated from Shapley game theory and was initially proposed by Lundberg and Lee71.

Figure 8 shows the results of the feature importance analysis of the RF model (analyzing the optimal model). The SHAP value of a feature indicates its contribution to the target value. When the SHAP value is closer to 0, it suggests that the feature has a smaller impact on the target value. Conversely, when the SHAP value is farther from 0, it indicates a more significant contribution of the feature to the target value. An observation can be made from Fig. 8 that the water-cement ratio, silica fume, and water reducer contributed the most to the prediction of compressive strength. In contrast, the size of the compressive specimen, PVA fiber, and CGP contributed the least to the prediction of compressive strength. Although the feature importance of Fig. 8 can illustrate the contribution of feature parameters, it cannot indicate whether the feature had a positive or negative impact on the prediction results. However, the Global SHAP value in Fig. 9 can demonstrate how each feature positively or negatively affected the compressive strength. In Fig. 9, each point represents a sample, with red representing a high eigenvalue and blue representing a low eigenvalue. Taking 'W' as an example, the high eigenvalue (red) made a negative contribution to the model output, and the low eigenvalue (blue) made a positive contribution to the model output. A lower eigenvalue (blue) would cause a higher SHAP value, indicating that the water-cement ratio and compressive strength had a negative impact. Contrarily, the high eigenvalue (red) positively contributed to the model output, and the low eigenvalue (blue) negatively contributed to the model output. A higher eigenvalue (red) would cause a higher SHAP value, indicating that silica fume positively affected compressive strength. This pattern applies to other features analyzed similarly.

Figure 8
figure 8

Feature importance of RF modeling.

Full size image
Figure 9
figure 9

Global SHAP values of RF modeling.

Full size image

SHAP offers the capability to observe both local (individual) interpretability as well as global interpretability, a feature not achieved by traditional variable importance algorithms. In Fig. 10, the SHAP values of a randomly selected single sample from the training model are depicted. The colors, blue and red, represent negative and positive contributions, respectively. Each feature is represented by a bar, with the length indicating its contribution to the prediction result. Red bars signify features that increase the prediction value, while blue bars indicate features that decrease the prediction value. Notably, in this specific example, quartz sand exhibited the longest red bar, indicating its significant contribution to increasing compressive strength, while silica fume had the longest blue bar, suggesting its substantial role in decreasing compressive strength. Additionally, SHAP can illustrate partial dependence plots, which differ from traditional ones by using SHAP values on the y-axis rather than the target value. Figure 11 shows the SHAP partial dependency plots. As shown, the compressive strength was positively correlated with silica fume and water reducing agent and negatively correlated with water cement ratio. In addition, the relationship between silica fume and compressive strength was almost linear. In contrast, the relationship between the water-cement ratio of the water-reducing agent and compressive strength was virtually logarithmic. From Fig. 11b, it can be observed from Fig. 11b that when W < 0.5, the SHAP value decreased with the increase in W, and the compressive strength was highly sensitive to the change of W. When W > 0.5, although W was still inversely correlated with compressive strength, the sensitivity was significantly weaker than W < 0.5. According to Fig. 11c, it can be found that when WR < 0.1, the SHAP value increased with the increase in W, and the compressive strength was highly sensitive to the change of WR; when WR > 0.1, the sensitivity of compressive strength to WR was significantly weakened.

Figure 10
figure 10

SHAP values of an individual sample.

Full size image
Figure 11
figure 11

SHAP partial dependency plots for the CS.

Full size image

Flexural strength

Table 3 shows the optimal parameter values of the model after hyperparameter optimization. Table 4 shows the evaluation indicators of the four models on the training and test sets. An observation can be made that the R2 values of the LR model and KNN model on the training set were smaller than those on the test set, which suggests the presence of data leakage, potentially leading to overly optimistic results72. The RF and XGB models performed better, with R2 of 0.97 on the training set and R2 of 0.91 on the test set. The RMSE and MAE of the RF model on the training set were 1.16 and 0.80, and the RMSE and MAE on the test set were 2.32 and 1.62, smaller than the XGB.

Table 3 The optimal parameter values of the model(FS).
Full size table
Table 4 Three evaluation indicators on the training set and the test set (FS).
Full size table

Figure 12 shows the predicted and actual flexural strength values on the test set. The four models had good prediction results on the test set, but the LR and KNN models exhibited data leakage phenomena that could not be used to make accurate predictions. Thus, only the generalization ability of the RF and XGB models on the test set was deemed practical, and the RF model was the most superior in terms of predicting flexural strength.

Figure 12
figure 12

Comparison between predicted value and actual value of FS [(a) LR, (b) KNN, (c) RF, (d) XGB].

Full size image

In summary, under the premise of the present study’s flexural strength data set, the LR and KNN models exhibited data leakage, and thus were not feasible for ECC flexural strength prediction. The XGB and RF models demonstrated exceptional prediction accuracy for the flexural strength of ECC, effectively addressing a gap in the literature where previous studies did not predict the flexural strength of ECC.

Figure 13 shows the importance of analyzing RF model characteristics of flexural strength. It indicates that PE fiber, water-cement ratio, and fiber aspect ratio contributed the most to the prediction of flexural strength. In contrast, other fiber types contributed the least to the prediction of compressive strength.

Figure 13
figure 13

Feature importance of RF modeling.

Full size image

Figure 14 shows that PE fiber positively affected flexural strength, while the water-cement ratio had a negative effect. Other characteristics can also be analyzed according to the Global SHAP value. Figure 15 shows the SHAP values of a random single sample of the training model. The most extensive red bar was coal gangue powder, which would increase the flexural strength for the specific sample of coal gangue powder. The most extensive blue bar was fiber content, which would reduce the flexural strength in this particular sample of fiber content.

Figure 14
figure 14

Global SHAP values of RF modeling.

Full size image
Figure 15
figure 15

SHAP values of an individual sample.

Full size image

In Fig. 16, the SHAP dependency partial plots reveal several key observations regarding the relationship between input features and flexural strength. Notably, flexural strength exhibited a positive correlation with fiber aspect ratio and the presence of PE fiber, while showing a negative correlation with the water-cement ratio. Moreover, the relationship between fiber aspect ratio and flexural strength appeared to be roughly linear, indicating that flexural strength increased with higher fiber aspect ratios. Conversely, the relationship between water-cement ratio and flexural strength appeared to be almost logarithmic. It can be found from Fig. 16b that when W < 0.75, the SHAP value decreased rapidly with the increase in W, indicating that the flexural strength was highly sensitive to the change of W in this range. When W > 0.75, although W was still inversely correlated with flexural strength, the sensitivity had almost disappeared.

Figure 16
figure 16

SHAP partial dependency plots for the FS.

Full size image

Tensile strength

Table 5 lists the optimal parameter values of the model after hyperparameter optimization. Table 6 shows the evaluation indicator values of the four models on the training and test sets. The R2 values of the LR model on the training and test sets were only 0.71 and 0.63, which shows that the LR model could not predict the tensile strength well. The R2 values of the KNN model on the training and test sets were 0.89 and 0.75, which indicates overfitting. The R2 values of the RF model on the training and test sets were 0.96 and 0.84; although RF exhibited slight overfitting, the prediction accuracy on the test set was relatively good. The R2 values of the XGB model on the training and test sets were 0.97 and 0.87. R2 was the largest among the four models, and the RMSE and MAE on the test set were only 1.04 and 0.68, the smallest among the four models.

Table 5 The optimal parameter values of the model(TS).
Full size table
Table 6 Three evaluation indicators on the training set and the test set (TS).
Full size table

From the collected data sets, it is evident that the variability of tensile strength in ECC surpassed that of compressive strength. In Table 6, it is apparent that both LR and KNN models exhibited underfitting. In the dataset used in the present study, the LR and KNN models failed to accurately predict the tensile strength of ECC. This could be attributed to the models' simplicity and the considerable variability in tensile strength test values. As such, machine learning struggled to fully capture the underlying relationships within the data, necessitating improvements in model complexity. This is why RF and XGB models achieved more accurate predictions. Despite their increased complexity, tree models yielded superior prediction results.

Figure 17 shows the predicted and actual tensile strength values on the test set. The RF and XGB data were considerably aggregated near the line, indicating exceptional generalization ability of the two models for tensile strength. Comparatively, the XGB model was superior.

Figure 17
figure 17

Comparison between predicted value and actual value of TS [(a) LR, (b) KNN, (c) RF, (d) XGB].

Full size image

In the dataset used in the present study, the LR and KNN models failed to predict the tensile strength of ECC accurately. This could be attributed to the models' simplicity and the considerable variability in tensile strength test values. As such, machine learning struggled to fully capture the underlying relationships within the data, necessitating improvements in model complexity. This is why RF and XGB models achieved more accurate predictions. Despite their increased complexity, tree models yielded superior prediction results.

Figure 18 shows analysis of the XGB model's feature importance for tensile strength. PE fiber, water-cement ratio, and water reducer contributed the most to predicting tensile strength, while other fibers contributed minimally.

Figure 18
figure 18

Feature importance of XGB modeling.

Full size image

Figure 19 shows that PE fiber positively affected tensile strength, while the water-cement ratio had a negative effect. Other characteristics could also be analyzed according to the Global SHAP value. Figure 20 shows the SHAP value of a random single sample of the training model. The most extensive red bar was slag. For this specific sample, the slag would increase the tensile strength. The most extensive blue bar was PE fiber. In this particular sample, PE fiber would reduce the tensile strength.

Figure 19
figure 19

Global SHAP values of XGB modeling.

Full size image
Figure 20
figure 20

SHAP values of the individual sample.

Full size image

Figure 21 shows the SHAP dependency partial plots. An observation can be made that the tensile strength was positively correlated with the water-reducing agent and whether PE fiber was present or negatively correlated with the water-cement ratio. In addition, the water-cement ratio and water-reducing agent were almost logarithmically related to the tensile strength. Figure 21a shows that when W < 0.5, the SHAP value decreased rapidly with the increase in W, indicating that the tensile strength was highly sensitive to the change of W in this range. When W > 0.5, although W was still inversely related to the tensile strength, the sensitivity became considerably weak. An observation can be made from Fig. 21b that when WR < 2, the SHAP value increased rapidly with the increase in WR, indicating that the tensile strength was highly sensitive to the change of WR in this range. When WR > 2, the sensitivity of tensile strength to WR changes in this range became significantly weaker.

Figure 21
figure 21

SHAP partial dependency plots for the TS.

Full size image

Tensile strain capacity

Table 7 lists the optimal parameter value of the model after hyperparameter optimization. Table 8 shows the evaluation indicator values of the four models on the training and test sets. The LR model was the weakest since the R2 values were only 0.63 and 0.61 on the training and test sets. The R2 values for the KNN model were 0.85 and 0.70 on the training and test sets, respectively, indicating a slight degree of overfitting. However, the model's generalization ability on the test set remained moderate. The R2 of the RF model reached 0.92 on the test set, but was only 0.72 on the test set, indicating overfitting of the model. The XGB model was the best, although slight overfitting existed. The R2 values of the XGB model were 0.95 and 0.80 on the training and test set. As shown in previous research73, the variation range of tensile strain capacity in the test was [± 0.56%, ± 1.77%]. However, the MAE on the test set of the XGB model in the present study was only 0.84%, which is entirely within a reasonable error range.

Table 7 The optimal parameter values of the model.
Full size table
Table 8 Three evaluation indicators on the training set and the test set.
Full size table

Figure 22 shows the predicted values and the actual values of tensile strain capacity on the test set. Among them, the XGB model prediction data was relatively more aggregated near the line and had the best effect on the prediction of tensile strain capacity. Notably, the prediction of tensile strain capacity could not be better than the other three mechanical indicators. This discrepancy can be attributed to the close relationship between tensile strain capacity and the uniform defects within the matrix of ECC. Many test groups do not introduce uniform artificial defects, leading to significant variability in tensile strain capacity test values74,75,76.

Figure 22
figure 22

Comparison between predicted value and actual value of TSC [(a) LR, (b) KNN, (c) RF, (d) XGB].

Full size image

In Fig. 23, the analysis of the importance of features for the XGB model reveals that PE fiber, fiber content, and oil film PVA fiber made the most significant contributions to predicting tensile strain capacity, while other fibers contributed minimally. Figure 24 shows that PE fiber, fiber content, and oil film PVA fiber positively impacted tensile stress capacity. Additionally, other characteristics can be further analyzed based on Global SHAP values.

Figure 23
figure 23

Feature importance of XGB modeling.

Full size image
Figure 24
figure 24

Global SHAP value of XGB modeling.

Full size image

Figure 25 shows the SHAP values of a random single sample of the training model. The most extensive red bar was fly ash. For this specific sample, fly ash would increase the tensile strain capacity. The most extensive blue bar was the fiber content. In this particular sample, the fiber content would reduce the tensile strain capacity. Figure 26 shows the SHAP dependency partial plots. An observation can be made that the tensile strain capacity was positively correlated with the water-cement ratio, fiber content, and whether there was PE fiber. In addition, the three features were roughly linear with the tensile strain capacity, and the sensitivity between tensile strain capacity and input features was moderate and stable.

Figure 25
figure 25

SHAP values of the individual sample.

Full size image
Figure 26
figure 26

SHAP partial dependency plots for the TSC.

Full size image

Control strategy of ECC mechanical indicator

According to the feature importance and SHAP analysis, there are different strategies to improve each mechanical indicator of ECC.

Reducing the water-cement ratio, increasing silica fume content, and selecting PE fiber are the most effective strategies for compressive strength. For flexural strength and tensile strength, selecting PE fiber, reducing the water-cement ratio, and increasing the fiber aspect ratio are the most effective methods. For tensile strain capacity, increasing the fiber aspect ratio, selecting PE fiber, and increasing fiber content will be the most effective methods. The control methods of ECC mechanical properties are shown in Table 9. An observation can be made that there are effective methods to improve the four mechanical properties, such as the use of PE fibers and the reduction of the water-cement ratio, thereby providing effective guidance for the design of ECC.

Table 9 Mechanical performance control methods of ECC.
Full size table

Conclusion

Based on the analysis using four machine-learning models, the four most critical mechanical properties of ECC were predicted, encompassing a wide range from ordinary to high compressive strength. Leveraging SHAP analysis technology, the feature importance of the optimal model was investigated and the sensitivity of feature parameters was thoroughly examined. Finally, based on the results of machine learning research, control suggestions for ECC mechanical properties were established. These findings are directly applicable to enhancing the design of ECC, facilitating its real-world application. The main conclusions are as follows:

  • The RF model performed best for compressive and flexural strength prediction, followed by the XGB model. For compressive strength, the R2 of the RF for the training set was 0.99, and the R2 and MAE for the test set were 0.92 and 5.40. For the flexural strength, the R2 of the RF for the training set was 0.97, and the R2 and MAE for the test set were 0.91 and 1.62. Therefore, RF is the most recommended model for predicting ECC's compressive and flexural strength.

  • In predicting tensile strength and tensile strain capacity, the XGB model performed best, followed by the random forest model. For the tensile strength, the R2 of the XGB model for the training set was 0.97, and the R2 and MAE for the test set were 0.87 and 0.68. For the tensile strain capacity, the R2 of the XGB model for the training set was 0.95, and the R2 and MAE for the test set were 0.80 and 0.84. Therefore, the XGB model is the most recommended model for predicting the tensile strength and tensile strain capacity of ECC.

  • Compressive strength was positively correlated with silica fume and water reducing agent and negatively correlated with water cement ratio. PE fiber positively affected flexural strength, and the water-cement ratio had a negative effect on flexural strength. PE fiber positively affected tensile strength, while the water-cement ratio negatively affected tensile strength. PE fiber, fiber content, and oil film PVA fiber positively impacted tensile stress capacity.

  • Compressive strength and tensile strength were highly sensitive to the change of W and WR, and the flexural strength was highly sensitive to the change of W. The sensitivity between tensile strain capacity and input features was moderate and stable.

From the perspective of the ECC mechanical index control strategy, the application of PE fiber is a very effective method to improve each mechanical index, and the second is to reduce the parameter of the cement ratio. In the design of ECC in practical engineering, these two characteristic parameters should be adjusted first to obtain the maximum effect with the least test. It avoids the repeated adjustment of each parameter in the mix ratio in the previous ECC design, which greatly reduces the test and time cost. At the same time, this paper only obtains the potential relationship between the characteristic parameters and the target values from many test data sets. In the future, we can consider the test data set and collect enough microstructure parameters of the fiber and matrix as the characteristic parameters to expect higher prediction accuracy.