Ultra-high performance concrete chloride ion diffusion coefficient prediction method based on machine learning

Abstract

The invention discloses an ultra-high performance concrete chloride ion diffusion coefficient prediction method based on machine learning. According to the method, a prediction model suitable for the chloride ion diffusion coefficient of the ultra-high performance concrete is constructed by comparing various machine learning algorithms, learning a large amount of data and screening evaluation indexes. The method is suitable for chloride ion diffusion coefficient prediction under different mixing ratios and environmental conditions, and provides a reliable basis for durability evaluation and structure life prediction of ultra-high performance concrete.

Description

Technical Field

[0001] This invention relates to a method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete based on machine learning, belonging to the field of concrete performance prediction technology. Background Technology

[0002] Ultra-high performance concrete (UHPC), as a novel building material with excellent mechanical properties and durability, has promoted the sustainable development of the modern construction industry and has been applied in road and bridge engineering in recent years. However, chloride ion diffusion problems bring huge operation and maintenance costs to concrete structures, especially coastal, marine, and offshore structures. Once chloride ions diffuse into the concrete and accumulate to a critical concentration around the reinforcing steel, corrosion of the steel will occur. Therefore, accurately estimating the chloride ion diffusion coefficient of concrete is crucial to improving the durability and service life of structures and developing effective repair plans.

[0003] Chloride ion diffusion coefficient is one of the three major indicators for assessing the durability of reinforced concrete in marine environments. Traditional methods for testing chloride ion diffusion coefficients or methods for measuring chloride ion diffusion within concrete are time-consuming and costly. Therefore, various predictive models have been developed to estimate chloride ion diffusion coefficients or their internal diffusion within concrete. These predictive models include various ion transport models and geochemical models, all of which require chloride ion diffusion coefficients as input. To reduce testing costs and time, several empirical models have been developed to determine the chloride ion diffusion coefficient of concrete, including those using supplementary cementitious materials (SCMs) such as fly ash (FA), abrasive blast furnace slag (GGBFS), and silica fume (SF). However, due to the complexity of diffusion phenomena in concrete, empirical estimation of the chloride ion diffusion coefficient using SCMs becomes more complex, making accurate derivation of relevant formulas difficult.

[0004] Over the past few decades, machine learning models have been widely applied to all aspects of people's lives thanks to the rapid development of artificial intelligence technology. Many methods exist for using machine learning models to predict the chloride ion diffusion coefficient of concrete, such as polygenic and artificial neural network (ANN) models. Numerous studies have predicted the chloride ion diffusion coefficient of ordinary concrete, high-performance concrete, and self-compacting concrete. One study used an ANN model built with 653 samples to predict the chloride ion diffusion coefficient of concrete containing different SCMs (such as slag, fly ash, and silica fume). Currently, there are no reports of using machine learning models to predict the chloride ion diffusion coefficient of ultra-high performance concrete.

[0005] While numerous studies exist, no effective machine learning model has yet been developed to analyze the true impact of various factors (cement, fiber, water-reducing agent, aggregate, and admixtures) on the chloride ion diffusion coefficient. In fact, previous research has shown that using supplemental cementitious materials (SCMs) can significantly reduce the chloride ion diffusion coefficient and prevent steel reinforcement corrosion. Studies also indicate that the amounts of fly ash, water, and cement have a significant impact on the chloride ion diffusion coefficient, but the specific quantitative effects of these factors have not been determined. Therefore, it is necessary to utilize sophisticated machine learning techniques to accurately quantify the influence of each factor on the chloride ion diffusion coefficient. Summary of the Invention

[0006] The purpose of this invention is to provide a machine learning-based method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete. This method can accurately predict the chloride ion diffusion coefficient of ultra-high performance concrete while promoting the research and development of the durability performance of ultra-high performance concrete.

[0007] The technical solution for achieving the objective of this invention is as follows:

[0008] A machine learning-based method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete includes the following steps:

[0009] Step 1: Extract and collect experimental data on ultra-high performance concrete from the literature, covering the main components of ultra-high performance concrete and the chloride ion diffusion coefficient (D). cl The main components of ultra-high performance concrete include water-reducing agents, fibers, water-cement ratio (W / B), cement, glass micron powder, metakaolin, lime powder, fly ash, slag powder, silica fume, coarse aggregate, fine aggregate, water, and nanomaterials, and a dataset is constructed.

[0010] Step 2: Divide the dataset into three independent subsets: training set, validation set, and test set using random sampling.

[0011] Step 3: Construct a machine learning model based on the Extreme Gradient Boosting (XGBoost) algorithm. The input is the main components of ultra-high performance concrete, and the output is the chloride ion diffusion coefficient. The dataset obtained in Step 1 is divided into training and testing sets using five-fold cross-validation for training and testing. The model parameters are then optimized to improve the model's coefficient of determination (R²) on the test set. 2 To improve the prediction accuracy and stability of the model, the model for predicting the chloride ion diffusion coefficient of ultra-high performance concrete is constructed by making the value as close to 1 as possible.

[0012] Step 4: Combine the SHapley additive interpretation to perform feature importance analysis on the prediction model and analyze the influence of each input on the chloride ion diffusion coefficient of ultra-high performance concrete.

[0013] Step 5: Substitute the main components of the ultra-high performance concrete to be predicted into the prediction model constructed in Step 3 to obtain the predicted value of its chloride ion diffusion coefficient.

[0014] Furthermore, in step 2, the training set accounts for 80%, while the validation set and test set each account for 10%.

[0015] Furthermore, in step 2, the training set is used for the basic training of the subsequent model to build the prediction logic framework of the chloride ion diffusion coefficient; the validation set is responsible for debugging the model hyperparameters, evaluating the model's generalization ability, and playing a key role in suppressing model overfitting; the test set is independent of the training and validation process and is used to finally verify the model's prediction accuracy and practical application performance.

[0016] Furthermore, in step 3, the average coefficient of determination and the mean absolute percentage error (MAPE) are used as evaluation metrics for the model.

[0017] Furthermore, in step 3, the model parameters include the learning target parameters and the Booster parameters.

[0018] Further, in step 3, the dataset obtained in step 1 is divided into training and testing sets using five-fold cross-validation for training and testing. The specific steps are as follows: the dataset obtained in step 1 is randomly and equally divided into five non-overlapping subsets. Then, each subset is selected as the test set in turn, and the remaining four subsets are combined as the training set. Five rounds of training and testing are carried out. In each round of testing, the model's coefficient of determination and mean absolute percentage error are generated. The arithmetic mean of the coefficient of determination and mean absolute percentage error is taken to calculate the model's mean coefficient of determination and mean absolute percentage error.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] (1) Traditional models are difficult to adapt to complex and variable environments and differences in material properties. This invention adopts machine learning methods, which can learn from a large amount of data and adapt to the prediction of chloride ion diffusion coefficient under different mix proportions and environmental conditions, thereby providing a reliable basis for the durability assessment and structural life prediction of ultra-high performance concrete.

[0021] (2) By comparing various machine learning algorithms and screening evaluation indicators, this invention constructs an optimal machine learning algorithm suitable for predicting the chloride ion diffusion coefficient of ultra-high performance concrete, and obtains the optimal prediction model for the chloride ion diffusion coefficient of ultra-high performance concrete. Attached Figure Description

[0022] Figure 1 This is a flowchart of a machine learning-based method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete.

[0023] Figure 2 The determination coefficients R for four types of machine learning 2 Box-shaped diagram;

[0024] Figure 3 Box plots of the mean absolute percentage error (MAPE) for four types of machine learning;

[0025] Figure 4 Feature importance analysis diagram;

[0026] Figure 5 A comparison chart of experimental values ​​and LSBoost model predictions of chloride ion diffusion coefficient in ultra-high performance concrete.

[0027] Figure 6 This is a comparison chart of the experimental value and the predicted value of the chloride ion diffusion coefficient of ultra-high performance concrete using the XGBoost model. Detailed Implementation

[0028] To enable those skilled in the art to better understand the prediction method of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments of the present invention are only one example of the present invention, and can be extended to many embodiments. All other embodiments based on the concept of the present invention should be protected by the present invention.

[0029] Example 1

[0030] The specific steps of the machine learning-based method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete are as follows:

[0031] (1) The system retrieved publicly available academic literature related to the durability performance of ultra-high performance concrete from both domestic and international sources, comprehensively summarized the experimental data reported in the literature, and organized them into an Excel file to establish the original database. For cases where the literature only presented data in the form of charts without clearly labeling specific data points, Getdata data extraction software was used for precise data collection to ensure data integrity. The collected raw data underwent targeted screening: samples with insufficient data corresponding to a single input parameter were removed. This is because the training effect of a machine learning model is closely related to the size of the dataset; sufficient data is key to ensuring the model's generalization ability and prediction reliability. Insufficient data will lead to inadequate model training and difficulty in achieving accurate predictions. The specific information of the final dataset constructed in this invention is shown in Table 1.

[0032] Table 1. Parameters of the dataset

[0033] unit Sample size Minimum value Maximum value average value median 25th percentile 75th percentile 1 Water reducing agent <![CDATA[kg / m 3 ]]> 340 0 59 23.1485 24 15.9 31.25 2 fiber <![CDATA[kg / m 3 ]]> 340 0 300 82.04075 76.8 0 156 3 W / B - 340 0.12 0.3 0.185 0.18 0.17 0.1945 4 cement <![CDATA[kg / m 3 ]]> 340 300 1930 786.31 720 620 890 5 Glass micro powder <![CDATA[kg / m 3 ]]> 340 0 550 33.92 0 0 0 6 metakaolin <![CDATA[kg / m 3 ]]> 340 0 366 4.21 0 0 0 7 lime powder <![CDATA[kg / m 3 ]]> 340 0 511 24.103 0 0 0 8 fly ash <![CDATA[kg / m 3 ]]> 340 0 509 81.12 0 0 175 9 Slag powder <![CDATA[kg / m 3 ]]> 340 0 758.8 40.317 0 0 0 10 silica ash <![CDATA[kg / m 3 ]]> 340 0 690 166.8157 159 115 208 11 coarse aggregate <![CDATA[kg / m 3 ]]> 340 0 759 25.9 0 0 0 12 fine aggregate <![CDATA[kg / m 3 ]]> 340 0 1445.7 862.145 943 750.5 1018.5 13 water <![CDATA[kg / m 3 ]]> 340 123.8 396.76 213.47 204.6 182 233 14 Nanomaterials <![CDATA[kg / m 3 ]]> 340 0 88 1.2 0 0 0 15 <![CDATA[D cl ]]> <![CDATA[×10 -12 m 2 / s]]> 340 0.0011 6.83 1.1847 0.615 0.23 2.05

[0034] (2) The ultra-high performance concrete chloride ion diffusion coefficient special dataset constructed in step (1) is divided into three independent subsets: training set, validation set and test set by random sampling. The division ratio and sample size are as follows: the training set accounts for 80% (including 272 samples), the validation set and the test set each account for 10% (each containing 34 samples).

[0035] (3) Four machine learning models, namely XGBoost, Support Vector Machine (SVM), Random Forest (RF), and Boosting Decision Tree (LSBoost), were built respectively. The prediction performance of the above four machine learning models was quantitatively evaluated using the five-fold cross-validation technique. In practice, the dataset constructed in step (1) was first randomly and equally divided into five non-overlapping subsets. Then, each subset was selected as an independent test set according to the cross-validation rules. The remaining four subsets were used as the training set. Five rounds of training and testing were carried out for each machine learning model. After each round of testing, the R of the corresponding model was recorded. 2 And MAPE, finally calculate five R values. 2 The arithmetic mean of MAPE and MAPE is used as the core evaluation criterion for the model's predictive performance.

[0036] (4) Based on the R values ​​of each model obtained from the five-fold cross-validation in step (3), 2 And MAPE, draw box plots as follows Figure 2 and Figure 3 As shown. In model performance evaluation, the average R... 2 A higher MAPE value, and the closer it is to 1, indicates a better fit to the data and higher predictive reliability. MAPE is calculated by subtracting the actual value from the predicted value and dividing by the experimental value. A lower MAPE value indicates a more reliable prediction. Comparing the statistical metrics of the four models, the XGBoost model has the highest average R-value. 2 The XGBoost and LSBoost models both achieved the highest MAPE, with the LSBoost model showing the lowest MAPE, indicating excellent accuracy in predicting the chloride ion diffusion coefficient of ultra-high performance concrete. Furthermore, the box plots reveal no outliers, demonstrating superior prediction stability. Considering both prediction accuracy and stability, the XGBoost and LSBoost models clearly demonstrate the most outstanding overall performance among the four algorithms.

[0037] (5) The parameter system of the XGBoost algorithm is divided into three categories according to function: general parameters, Booster parameters, and learning target parameters. Among them, general parameters mainly regulate the overall operation logic of the model through macroscopic functions, and the default configuration of the algorithm can usually meet the basic requirements; Booster parameters are used to control the performance of the Booster component in each iteration of the model, while learning target parameters directly affect the achievement of the model training objective. Therefore, for the parameter optimization of the XGBoost algorithm, the core adjustment objects are the learning target parameters and Booster parameters. Among them, the optimization of Booster parameters can adopt a step-by-step strategy: first, set all Booster parameters to the default value, and then combine the characteristics of the dataset to gradually optimize key parameters, including the maximum depth of the tree, the upper limit of the number of nodes, and the weight distribution of the tree, so as to adapt to the data distribution characteristics. The role of the learning target parameters is to define the optimization objective of the model and the evaluation criteria of the results of each iteration. Taking the eval_metric parameter as an example, its default value varies with the problem type: MAPE is used by default in regression problems, and error rate is used by default in classification problems. Since the prediction of chloride ion diffusion coefficient in ultra-high performance concrete, which is addressed in this invention, is a regression problem, the default value of the eval_metric parameter is RMSE. The ultimate goal of parameter tuning is to iteratively adjust the XGBoost model to make its coefficient of determination (R²) on the test set as close to 1 as possible, thereby achieving optimal prediction performance.

[0038] (6) Perform feature importance analysis on the XGBoost model. Feature importance analysis is used to explain and understand the contribution of each input variable to the prediction of chloride ion diffusion coefficient by combining the "SHapley additive interpretation" technique of the XGBoost model. Figure 4 The effects of 14 input variables on the chloride ion diffusion coefficient of UHPC are shown, and two key pieces of information are provided by the scatter plot of SHAP values: (1) the effect of each input variable on the chloride ion diffusion coefficient of UHPC represented by the y-axis; and (2) the effect of each input variable on the chloride ion diffusion coefficient of UHPC represented by the x-axis.

[0039] (7) The prediction results of the XGBoost model and the LSBoost model on the test set are as follows: Figure 5 and Figure 6 As shown in the figure, the comparative analysis reveals a good agreement between the experimentally measured values ​​of the chloride ion diffusion coefficient of ultra-high performance concrete and the predicted values ​​of the XGBoost model. This indicates that the model possesses high accuracy and reliability in predicting target performance parameters, outperforming the LSBoost model. In conclusion, the XGBoost model constructed in this invention can be effectively applied to the quantitative prediction of the chloride ion diffusion coefficient of ultra-high performance concrete, providing strong support for related engineering design and performance evaluation.

Claims

1. A method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete based on machine learning, characterized in that, Includes the following steps: Step 1: Extract and collect experimental data on ultra-high performance concrete from the literature, covering the main components of ultra-high performance concrete and chloride ion diffusion coefficient. The main components of ultra-high performance concrete include water-reducing agent, fiber, water-cement ratio, cement, glass micro powder, metakaolin, lime powder, fly ash, slag powder, silica fume, coarse aggregate, fine aggregate, water and nanomaterials, and construct a dataset. Step 2: Divide the dataset into three independent subsets: training set, validation set, and test set using random sampling. Step 3: Construct a machine learning model based on the XGBoost algorithm. The input is the main components of ultra-high performance concrete, and the output is the chloride ion diffusion coefficient. The dataset obtained in Step 1 is divided into training and testing sets using five-fold cross-validation. The model parameters are then optimized to make the coefficient of determination of the model on the testing set as close to 1 as possible, thereby improving the prediction accuracy and stability of the model and constructing a prediction model for the chloride ion diffusion coefficient of ultra-high performance concrete. Step 4: Combine the SHapley additive interpretation to perform feature importance analysis on the prediction model and analyze the influence of each input on the chloride ion diffusion coefficient of ultra-high performance concrete. Step 5: Substitute the main components of the ultra-high performance concrete to be predicted into the prediction model constructed in Step 3 to obtain the predicted value of its chloride ion diffusion coefficient.

2. The method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete according to claim 1, characterized in that, In step 2, the training set accounts for 80%, while the validation set and test set each account for 10%.

3. The method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete according to claim 1, characterized in that, In step 2, the training set is used for the basic training of the subsequent model to build the prediction logic framework of the chloride ion diffusion coefficient; the validation set is responsible for debugging the model hyperparameters, evaluating the model's generalization ability, and playing a key role in suppressing model overfitting. The test set is independent of the training and validation process and is used to ultimately validate the model's prediction accuracy and performance in real-world applications.

4. The method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete according to claim 1, characterized in that, In step 3, the average coefficient of determination and the average absolute percentage error are used as evaluation indicators for the model.

5. The method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete according to claim 1, characterized in that, In step 3, the model parameters include the learning target parameters and the Booster parameters.

6. The method for predicting the chloride ion diffusion coefficient of ultra-high performance concrete according to claim 1, characterized in that, In step 3, the dataset obtained in step 1 is divided into training and testing sets using five-fold cross-validation. The specific steps are as follows: the dataset obtained in step 1 is randomly and equally divided into five non-overlapping subsets. Then, each subset is selected as the test set in turn, and the remaining four subsets are combined as the training set. Five rounds of training and testing are carried out. In each round of testing, the model's coefficient of determination and mean absolute percentage error are generated. The arithmetic mean of the coefficient of determination and mean absolute percentage error is taken to calculate the model's mean coefficient of determination and mean absolute percentage error.

Images