A machine learning-based go-uhpc bond prediction method and system

Abstract

This invention belongs to the field of materials engineering technology and discloses a GO-UHPC bonding prediction method and system based on machine learning. It constructs a multi-source heterogeneous dataset by integrating multi-dimensional information such as GO microstructure and low-carbon parameters. Through federated learning, it achieves cross-unit data collaborative expansion. Furthermore, through multi-model fusion and a bi-branch attention mechanism, it captures the synergistic effects and temporal evolution patterns of multiple factors. Without complex experimental procedures, it can simultaneously predict multiple core indicators related to bonding performance and low-carbon optimization potential, thereby avoiding various interference factors, reducing experimental costs, and improving engineering design and construction efficiency. A dual-algorithm interpretable analysis module is constructed to present the influence weights and effects of key factors such as GO content and water-cement ratio on the prediction results, and generates a visual report. Simultaneously, the model is encapsulated as a lightweight engineering module, with multi-terminal operation interfaces designed to support batch parameter input, result export, and real-time on-site calibration.

Description

Technical Field

[0001] This invention belongs to the field of materials engineering technology, specifically a GO-UHPC bonding prediction method and system based on machine learning. Background Technology

[0002] Ultra-high performance concrete (UHPC) possesses ultra-high strength, high toughness, and excellent durability, and is widely used in high-end infrastructure and special structures. Its bonding performance with steel bars and precast components is crucial to ensuring the structural integrity and service life. Graphene oxide (GO) has a surface rich in oxygen-containing functional groups, which can form a good interfacial reaction with UHPC cementitious materials, effectively optimizing its microstructure and improving its mechanical and bonding properties.

[0003] Currently, the prediction methods for the bonding performance of GO-UHPC mainly rely on traditional testing methods. Traditional testing methods (such as pull-out tests and shear tests) require the preparation of a large number of specimens and a long curing period. This not only results in high testing costs and low efficiency, but is also easily affected by factors such as testing equipment, environmental conditions, and operators, leading to large dispersion of test results. Furthermore, it is impossible to quickly predict the bonding performance under the combined effects of multiple factors such as different GO dosages, UHPC mix ratios, curing conditions, and interface treatment methods, which makes it difficult to meet the high-efficiency requirements of engineering design and construction.

[0004] Machine learning technology, with its strong multi-factor analysis, nonlinear fitting, and high-precision prediction capabilities, is increasingly widely used in the field of material property prediction. However, its application in predicting the bonding performance of GO-UHPC still faces many technical challenges. Existing research often employs a single machine learning model without considering the specific characteristics of GO-UHPC, resulting in insufficient prediction accuracy and weak generalization ability. Feature selection is mostly focused on macroscopic parameters of UHPC, neglecting microscopic parameters of GO, thus failing to fully reflect the influence mechanism of the GO-UHPC interface on bonding performance. Furthermore, most models are black-box models, making it difficult to explain the contribution and mechanism of each influencing factor to bonding performance, which is not easily understood and applied by engineers, thereby limiting its engineering implementation. Summary of the Invention

[0005] The purpose of this invention is to provide a GO-UHPC bonding prediction method and system based on machine learning to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based GO-UHPC adhesion prediction method, comprising:

[0007] Preferably, the data construction stage constructs a multi-source heterogeneous dataset containing GO microstructures, UHPC macrostructures, interface processing parameters, environmental parameters, measured values ​​of bonding performance, and low-carbon parameters. Quantitative parameters of GO microstructures are collected, and GO sheet thickness, number of layers, types and contents of oxygen-containing functional groups, dispersion uniformity, sheet defect density, and quantitative characterization parameters of GO microstructures, including defect area ratio and defect distribution density, are determined using a combination of scanning electron microscopy, X-ray diffraction, and atomic force microscopy.

[0008] The weight allocation adopts a basic weight + dynamic adjustment rule: the basic weight of key features such as GO layer defect density and interface roughness accounts for 40%, the basic weight of low-carbon features such as admixture replacement rate and carbon emissions accounts for 30%, and the basic weight of other features accounts for 30%. The dynamic adjustment part is adjusted according to the real-time correlation between the feature and the predicted target value. For every 0.1 increase in correlation, the corresponding feature weight increases by 2%-5%, and the weight of a single feature does not exceed 60%, ensuring a balance between the dominance of core features and feature diversity.

[0009] Macroscopic mix proportion parameters and low-carbon indicators of UHPC were collected, including water-cement ratio, cement dosage, silica fume content, slag powder and fly ash content, coarse aggregate type, steel fiber content and aspect ratio, as well as low-carbon parameters such as carbon emissions, admixture substitution rate, and cement dosage reduction ratio. Interface treatment parameters were also collected, including roughness parameters Ra and Rz measured by a three-dimensional profilometer, high-temperature aging resistance parameters of the interface agent, salt spray aging resistance parameters, steel bar diameter and surface condition, interface agent type and dosage.

[0010] Environmental maintenance parameters and extreme environment simulation parameters were collected, including conventional maintenance temperature, humidity, and maintenance age, as well as extreme environmental parameters such as the number and intensity of high-temperature exposure, salt spray corrosion, and low-temperature freeze-thaw cycles. Standard pull-out tests combined with microscopic measurements were used to determine the measured values ​​of GO-UHPC bond strength, bond stiffness, slippage, interfacial shear stress, and quantitative indicators of interfacial damage under corresponding conditions. These measured values ​​served as the model's prediction target values, including quantitative indicators of interfacial damage. The data preprocessing process employed both the Laida criterion and the Grubbs criterion to remove outliers. Missing data were filled using a combination of K-nearest neighbor interpolation and linear interpolation. Z-score standardization was used to eliminate the influence of dimensions. The process included data normalization and outlier backtesting steps. A federated learning framework was introduced to enable data sharing and collaborative expansion among multiple laboratories and engineering units. The dataset labeling system labeled each set of data with the scene type, low-carbon level, and GO microscopic defect level.

[0011] Preferably, the feature optimization stage adopts a feature optimization strategy of hierarchical screening, association construction, and dynamic verification to construct a five-dimensional association feature system adapted to GO-UHPC bonding performance prediction. Through feature hierarchical classification, the dataset features are divided into five categories: GO micro-feature layer, UHPC macro-feature layer, interface parameter layer, environmental feature layer, and low-carbon feature layer. The intrinsic association between each layer of features and bonding performance and low-carbon targets is clarified. The preliminary screening uses the mutual information method to calculate the dual correlation degree between each single feature and the bonding performance target value and the low-carbon target value. A dynamic threshold that is adaptively adjusted according to the scene type is set to eliminate redundant features with a correlation degree lower than the dynamic threshold.

[0012] The correlation features are constructed into four core correlation features: correlation features between GO micro properties and UHPC macro parameters; correlation features between GO micro properties and interface parameters; correlation features between environmental parameters and interface parameters; and correlation features between low carbon parameters and bonding performance.

[0013] Feature determination employs a triple screening method combining mutual information, grey relational analysis, and analysis of variance. Single and related features are comprehensively ranked to select features with high correlation, strong significance, and good low-carbon adaptability as model input features. The method includes a feature interaction verification step to verify the effectiveness of related features. A GO-UHPC bonding prediction feature system is constructed. After feature optimization, a dedicated machine learning model adapted to GO-UHPC bonding prediction is built by combining the selected core related features.

[0014] Preferably, the model building stage combines the nonlinearity, multi-factor synergy, temporal evolution and low-carbon synergy characteristics of GO-UHPC bonding performance to build a hybrid machine learning prediction model that combines CNN, LSTM, Attention and residual connection, optimizes the model gradient propagation through residual connection technology, and builds a two-branch attention mechanism.

[0015] A convolutional neural network is used to extract spatial correlation information from the input features. An adaptive-size convolutional kernel is designed to adjust its size according to the GO sheet size and UHPC microstructure to capture the spatial correlation characteristics in the input features. A long short-term memory network is used to capture temporal correlation information, and an attention forgetting gate is added to focus on remembering the temporal correlation of GO doping amount, curing age and low carbon parameters. A dual-branch attention mechanism is introduced to assign weights to the features extracted by CNN and LSTM, focusing on key features such as GO sheet defect density, interface roughness, water-binder ratio and admixture replacement rate, while strengthening the weight adaptation of low carbon features.

[0016] The model initialization sets initial parameters based on the bonding performance and low-carbon characteristics of GO-UHPC. The number of convolutional kernels in the CNN ranges from 16 to 64, the kernel size ranges from 3×3 to 7×7, and the pooling method combines max pooling and average pooling. The number of hidden layer neurons in the LSTM ranges from 32 to 128, and the forget gate threshold ranges from 0.1 to 0.3. Initial values ​​for the weight coefficients of the Attention mechanism are also set. An adaptive parameter initialization algorithm is used. The model output layer adopts a multi-output structure. The model predicts five indicators of GO-UHPC: bonding strength, slippage, interface damage degree, bonding failure warning threshold, and low-carbon optimization potential. The model uses quantization compression technology to simplify redundant parameters and reduce computational complexity.

[0017] Preferably, the model training stage adopts a collaborative training strategy of hierarchical training, dynamic hyperparameter optimization and transfer learning. Hierarchical training divides the model training into three levels: feature extraction layer training, attention layer training and prediction layer training. First, the CNN and LSTM feature extraction layers are trained using the preprocessed dataset. Then, the dual-branch Attention layer is trained to optimize the feature weight allocation. Finally, the extracted features are input into the prediction layer for joint training.

[0018] Hyperparameter optimization employs a combination of Bayesian optimization and reinforcement learning to dynamically optimize the model's hyperparameters. After each round of optimization, the model performance is jointly verified using 10-fold cross-validation and leave-one-out validation until the optimal hyperparameter combination is found. An adaptive hyperparameter adjustment mechanism is used to adjust the hyperparameters in real time based on the training error. Adaptive Dropout layers with different Dropout probabilities and L2 regularization are used for different feature layers. At the same time, an early stopping mechanism and model ensemble strategy are combined to stop training when the prediction error on the validation set does not decrease for several consecutive rounds.

[0019] The basic model is trained using the general UHPC performance dataset, and then the basic model is fine-tuned using the GO-UHPC small sample dataset to achieve model knowledge transfer.

[0020] Preferably, the verification and correction stage constructs a verification system with multiple scenarios, multiple indicators, and multiple dimensions. Combining the engineering application scenarios of GO-UHPC, a dual adaptive error correction strategy of scenario and low carbon is designed. When dividing the dataset, the preprocessed multi-source heterogeneous dataset is divided into training set, test set and verification set in a ratio of 7:2:1. The training set is used for model training and hyperparameter optimization, the test set is used to initially verify the model prediction performance, and the verification set is used to verify the stability and generalization ability of the model. At the same time, a special dataset for extreme scenarios is set up to separately verify the prediction performance of the model in extreme environments.

[0021] Multi-index validation uses four core indicators—coefficient of determination, mean absolute error, root mean square error, and relative error—to evaluate the model's prediction accuracy. Validation standards are set, requiring the model to have R² ≥ 0.97, MAE ≤ 1.5, RMSE ≤ 2.5, and RE ≤ 5% on the test set, and R² ≥ 0.95 on the extreme scenario validation set. Simultaneously, a low-carbon adaptability index is set to evaluate the fit between the model's prediction results and low-carbon goals. Multi-scenario validation selects five typical engineering scenarios: general infrastructure, marine engineering, high-temperature and high-humidity environment, low-temperature freeze-thaw environment, and salt spray corrosion environment. Test data on the GO-UHPC bonding performance and low-carbon data are collected for each scenario to verify the model's generalization ability and low-carbon adaptability in different scenarios.

[0022] Error correction addresses prediction errors under different scenarios and low-carbon levels by constructing a scenario- and low-carbon dual adaptive error correction model. Error sources are analyzed through SHAP, and error correction functions for different scenarios and low-carbon levels are established to dynamically correct the model's prediction results. At the same time, an error feedback mechanism is set up to use the corrected results in reverse for model fine-tuning.

[0023] Preferably, the engineering adaptation phase designs an engineering adaptation module that integrates interpretability, engineering, and low-carbonization. The interpretability module uses SHAP and LIME dual algorithms to perform interpretability analysis on the model's prediction results, calculates the contribution value and influence weight of each input feature on the predicted bonding performance and low-carbon optimization potential, clarifies the influence law and weight ranking of key factors such as GO content, water-cement ratio, interface roughness, and admixture substitution rate, and generates a visual interpretation report containing feature contribution charts and low-carbon optimization suggestion charts.

[0024] The engineering interface encapsulates the trained and optimized hybrid machine learning model into a lightweight engineering module. It features dual interfaces for PC and mobile devices, supporting input of actual engineering parameters such as GO micro parameters, UHPC mix proportion, curing conditions, and low-carbon targets. It outputs prediction results such as bond strength, slippage, interface damage degree, bond failure warning, and low-carbon optimization potential. It also supports batch data input, result export, and printing functions. The parameter recommendation function automatically recommends the optimal GO content, UHPC mix proportion, interface treatment scheme, and low-carbon optimization scheme based on the model's prediction results, feature contribution analysis, and low-carbon targets, for different engineering needs.

[0025] The field calibration module supports inputting actual field measurement data to calibrate the model's prediction results in real time, correcting errors caused by differences between the field environment and the laboratory environment. It also supports automatic uploading of calibration data to the dataset.

[0026] Preferably, the update iteration stage establishes a dynamic update iteration mechanism for data update, model iteration, version management and low-carbon adaptation. The data update interface adopts a three-in-one design of automatic data collection, manual input and federated learning collaborative update, which can access new GO-UHPC bonding performance test data, engineering application data, low-carbon data and extreme environment test data in real time. At the same time, it realizes data collaborative update of multiple laboratories and multiple engineering units through federated learning, expanding the coverage, sample size and low-carbon scenarios of the dataset.

[0027] The model iteration is set with a regular iteration cycle of 6 months and a triggered iteration mechanism. New data and transfer learning are used to retrain the existing model and optimize hyperparameters. The model structure, feature system and hyperparameters are adjusted in combination with changes in new materials, new engineering scenarios and low-carbon goals. The version management establishes a model version management mechanism and a scenario and low-carbon label system to record the model parameters, performance indicators, applicable scenarios, low-carbon adaptation level and improvement content of each iteration.

[0028] The technology integration and iteration regularly incorporates the latest technological achievements in the fields of new materials, artificial intelligence, and low-carbon building materials to upgrade and optimize the model structure, feature system, and engineering adaptation modules.

[0029] This invention also provides a machine learning-based GO-UHPC adhesion prediction system, which, based on the above method, includes:

[0030] The data acquisition and preprocessing module includes a parameter acquisition unit and a data processing unit. It acquires the microscopic characteristic parameters of GO through scanning electron microscope and X-ray diffraction equipment, and acquires the macroscopic mixing ratio, interface treatment, environmental curing parameters and measured values ​​of bonding performance of UHPC through experiments. It completes outlier removal, missing value filling and dimension elimination by using the Laida criterion, K-nearest neighbor interpolation method and Z-score standardization method, and outputs a multi-source heterogeneous dataset.

[0031] The feature processing module adopts a hierarchical screening-coupling construction strategy. It uses mutual information method, grey relational analysis and variance analysis to complete the initial feature screening, coupled feature construction and final feature determination, and outputs a feature set suitable for model training.

[0032] The model building and training module has a built-in CNN-LSTM-Attention hybrid machine learning model. It completes the model structure initialization and hierarchical training, and uses Bayesian optimization and 10-fold cross-validation to achieve dynamic optimization of hyperparameters. It suppresses overfitting through Dropout layer, L2 regularization and early stopping mechanism, and outputs the best prediction model after training.

[0033] The validation and correction module divides the dataset into training, testing and validation sets according to a set ratio. It validates the model performance through multiple metrics such as R², MAE, and RMSE and multiple engineering scenarios. It also has a built-in scene adaptive error correction function to dynamically correct the prediction results.

[0034] The engineering adaptation module includes an interpretability unit, an interface unit, and a parameter recommendation unit. It generates a visual explanation report through the SHAP algorithm, provides input and output interfaces, and can automatically recommend the optimal parameter scheme.

[0035] The dynamic update module is designed with an automatic data update interface and a periodic iteration mechanism. It can integrate new data to complete model retraining and optimization, and establish a version management unit.

[0036] The beneficial effects of this invention are as follows:

[0037] 1. This invention constructs a multi-source heterogeneous dataset by integrating multi-dimensional information such as GO micro-characteristics and low-carbon parameters. It achieves cross-unit data collaborative expansion through federated learning, and captures the synergistic effects and temporal evolution patterns of multiple factors through multi-model fusion and a dual-branch attention mechanism. Without complex experimental procedures, it can simultaneously predict multiple core indicators related to bonding performance and low-carbon optimization potential, thereby avoiding various interference factors, reducing experimental costs, improving engineering design and construction efficiency, and adapting to complex prediction needs under multi-factor coupling.

[0038] 2. This invention employs a dual-algorithm to construct an interpretable analysis module, presenting the influence weights and patterns of key factors such as GO content and water-cement ratio on the prediction results, and generating intuitive visualization reports. Simultaneously, the model is encapsulated as a lightweight engineering module, with multi-terminal operation interfaces designed to support batch parameter input, result export, and real-time on-site calibration. It can also automatically recommend optimal GO content, UHPC mix ratio, and interface processing solutions based on engineering needs and low-carbon goals, enabling engineers to clearly understand the model logic, simplifying the operation process, adapting to actual on-site applications, and lowering the threshold for technology implementation.

[0039] 3. This invention improves prediction stability in complex scenarios by combining federated learning with hierarchical training and transfer learning; it establishes a dynamic mechanism for data updates, model iterations, and version management, enabling real-time access to various new data, adjusting model structure and parameters according to changes in new materials, new scenarios, and low-carbon goals, and regularly integrating the latest technological achievements from multiple fields for optimization and upgrading; it covers multiple scenarios such as general infrastructure and extreme environmental engineering, and continuously ensures the accuracy of prediction results under different environments through scenario-based and low-carbon dual adaptive error correction. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of the method of the present invention;

[0041] Figure 2 This is a flowchart illustrating the construction process of the multi-source heterogeneous dataset of this invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] like Figures 1 to 2 As shown, this embodiment of the invention provides a machine learning-based GO-UHPC adhesion prediction method, including:

[0044] The data construction phase involves building a multi-source heterogeneous dataset that includes GO microstructures, UHPC macroscopic parameters, interface processing parameters, environmental parameters, measured values ​​of bonding performance, and low-carbon parameters. Quantitative parameters of GO microstructures are collected, and GO sheet thickness, number of layers, types and contents of oxygen-containing functional groups (hydroxyl, epoxy, and carboxyl groups), dispersion uniformity, sheet defect density, and quantitative characterization parameters of GO microstructures, including defect area ratio and defect distribution density, are determined using a combination of scanning electron microscopy (SEM), X-ray diffraction (XRD), and atomic force microscopy (AFM).

[0045] Macroscopic mix proportion parameters and low-carbon indicators of UHPC were collected, including water-cement ratio, cement dosage, silica fume content, slag powder and fly ash content, coarse aggregate type (granite, sandstone, quartzite, etc.), steel fiber content and aspect ratio, as well as low-carbon parameters such as carbon emissions, admixture substitution rate, and cement dosage reduction ratio. Interface treatment parameters were also collected, including roughness parameters Ra and Rz measured by a three-dimensional profilometer, high-temperature aging resistance parameters of the interface agent, salt spray aging resistance parameters, steel bar diameter and surface condition (sandblasted and non-sandblasted, rust removal grade), interface agent type and dosage.

[0046] Environmental curing parameters and extreme environment simulation parameters were collected, including conventional curing temperature, humidity, and curing age (3d, 7d, 28d, 90d), as well as extreme environmental parameters such as the number and intensity of high temperature exposure, salt spray corrosion, and low temperature freeze-thaw cycles. Standard pull-out tests combined with microscopic measurements were used, and SEM was employed to observe the microstructure of the interface bonding. Measured values ​​of the bond strength, bond stiffness, slip, interface shear stress, and quantitative indicators of interface damage degree of GO-UHPC under corresponding conditions were determined as target values ​​for model prediction, including quantitative indicators of interface damage degree. The data preprocessing process employed both the Laida criterion and the Grubbs criterion to remove outliers, thereby improving the accuracy of outlier identification. Specifically, the Laida criterion was used for initial screening with a significance level of α=0.05, followed by secondary verification using the Grubbs criterion. The critical value Gα was adaptively determined based on the data sample size n. When n < 50, the critical value table of the Grubbs criterion was referenced. After double screening, data samples conforming to statistical laws were retained to ensure the consistency and repeatability of outlier removal.

[0047] Missing data is filled by combining K-nearest neighbor interpolation and linear interpolation. The Z-score standardization method is used to eliminate the influence of dimensions. The process includes data normalization and outlier backtesting. A federated learning framework is introduced to enable data sharing and collaborative expansion among multiple laboratories and engineering units. The dataset labeling system labels each set of data with scene type, low-carbon level, and GO micro-defect level. Scene types include general infrastructure, marine engineering, cold region engineering, etc.

[0048] Federated learning adopts a horizontal federated learning collaborative model, in which each participating unit only uploads intermediate parameters of model training (gradients, weight update values) and does not disclose the original data. Data security is ensured through encrypted transmission (using AES-256 encryption protocol), parameter aggregation after local model training (federated averaging algorithm), and abnormal parameter monitoring (setting parameter fluctuation threshold ±30%), ensuring that data privacy is not leaked during cross-unit data collaboration, while ensuring the validity of the aggregated model parameters.

[0049] The low-carbon level is divided into three levels based on carbon emissions and blending material replacement rate: Level A (high-low carbon) has carbon emissions ≤400kg / m³ and blending material replacement rate ≥40%; Level B (medium-low carbon) has carbon emissions 401-550kg / m³ and blending material replacement rate 25%-39%; Level C (basic low carbon) has carbon emissions 551-700kg / m³ and blending material replacement rate 15%-24%, which facilitates data classification management and low-carbon adaptation training of models.

[0050] The feature optimization stage employs a hierarchical screening, association construction, and dynamic verification strategy to construct a five-dimensional association feature system adapted to GO-UHPC bonding performance prediction. Through feature hierarchical classification, the dataset features are divided into five categories: GO micro-feature layer (including defect features), UHPC macro-feature layer, interface parameter layer, environmental feature layer (including extreme environments), and low-carbon feature layer. The intrinsic association between each layer's features and bonding performance and low-carbon targets is clarified. The initial screening uses the mutual information method to calculate the dual correlation degree between each single feature and the bonding performance target value and the low-carbon target value. A dynamic threshold that is adaptively adjusted according to the scene type is set to eliminate redundant features with a correlation degree lower than the dynamic threshold, thereby reducing the computational load of the model while retaining features that are strongly associated with the low-carbon target.

[0051] The dynamic threshold is calculated based on the distribution of the correlation between features and target values ​​in a scenario. The median of the correlation between all single features in the scenario is multiplied by the scenario adaptation coefficient. The coefficient is 0.6 for ordinary infrastructure scenarios and 0.4 for extreme scenarios such as marine engineering and cold-region engineering. This forms the feature screening threshold for the scenario, which eliminates redundant features while retaining key low-carbon related features.

[0052] The correlation features are structured into four core types: correlation features between GO microstructures and UHPC macrostructures, such as the synergistic term between GO lamellar defect density and slag powder content, and the interaction term between GO oxygen-containing functional group content and water-cement ratio; correlation features between GO microstructures and interface parameters, such as the interaction term between GO dispersion uniformity and interface roughness; correlation features between environmental parameters and interface parameters, such as the synergistic term between salt spray corrosion cycles and interface agent dosage; and correlation features between low-carbon parameters and bonding performance, such as the correlation between admixture replacement rate and bond strength, and the interaction term between carbon emissions and interface shear stress.

[0053] Feature determination employs a triple screening method: mutual information method, grey relational analysis, and analysis of variance. The triple screening is performed in the following order: initial screening using mutual information method → ​​secondary screening using grey relational analysis → final screening using analysis of variance. Grey relational analysis focuses on calculating the correlation coefficient between features and bonding performance and low-carbon targets, with a value range of 0-1, retaining features with a coefficient ≥0.6. Analysis of variance determines the significance of the feature's impact on the target value, setting a significance level α=0.05, eliminating non-significant features with a p-value >0.05, and finally selecting core input features that have both high correlation and strong significance.

[0054] The single features and related features are comprehensively ranked, and features with high correlation, strong significance, and good low-carbon adaptability are selected as model input features. The feature interaction verification step is included to verify the effectiveness of the related features. A GO-UHPC bonding prediction feature system that takes into account completeness, simplicity and low carbon emissions is constructed. After feature optimization, a dedicated machine learning model adapted to GO-UHPC bonding prediction is constructed by combining the selected core related features.

[0055] Feature interaction verification adopts a three-step process of splitting, verifying, and integrating: The first step is to split the dataset into a verification subset and a test subset in an 8:2 ratio; the second step is to train a temporary model on the verification subset, calculate the standard deviation of the contribution value of each associated feature, and retain stable associated features with a standard deviation <0.15; the third step is to substitute the stable associated features into the test subset for verification, remove weak associated features on the test subset whose feature contribution value is less than 30% of that on the verification subset, and finally retain the core associated features that have both stability and effectiveness.

[0056] In the model building stage, the nonlinear, multi-factor synergy, temporal evolution and low-carbon synergy characteristics of GO-UHPC bonding performance are combined to build a hybrid machine learning prediction model that combines CNN, LSTM, Attention and residual connection. The model gradient propagation is optimized through residual connection technology, and a dual-branch attention mechanism is built, which includes micro feature branches and macro and low-carbon feature branches.

[0057] The residual connection is implemented using identity mapping and feature fusion. For the CNN feature extraction layer, the residual block consists of a convolutional layer → batch normalization → activation function. The shortcut connection skips two convolutional layers and adds the input features to the output features. For the LSTM layer, the residual connection adds the hidden layer output of the LSTM to the input sequence features element-wise and then passes it to the next layer. The activation function is uniformly ReLU to avoid gradient vanishing or gradient exploding problems.

[0058] Convolutional neural networks (CNNs) are used to extract spatial correlation information from input features, such as the spatial correlation between GO micro-defects and UHPC microstructure, and the spatial correlation between interface roughness and GO dispersion uniformity. Adaptive-size convolutional kernels that adjust their size according to GO layer size and UHPC microstructure are designed to capture the spatial correlation characteristics in input features.

[0059] Long Short-Term Memory (LSTM) networks are used to capture temporal correlation information, such as the dynamic evolution of adhesion performance by maintenance age and number of extreme environment cycles. An attention forgetting gate is added to focus on the temporal correlation of GO content, maintenance age and low carbon parameters.

[0060] A dual-branch attention mechanism is introduced to assign weights to the features extracted by CNN and LSTM, focusing on key features such as GO layer defect density, interface roughness, water-gluoride ratio, and admixture substitution rate. At the same time, the weight adaptation of low-carbon features is strengthened to improve the model's prediction accuracy and low-carbon adaptability.

[0061] The model initialization combines the bonding performance of GO-UHPC and the low-carbon characteristics to set initial parameters. The number of convolutional kernels in the CNN is 16 to 64, the kernel size is 3×3 to 7×7, and the pooling method is a combination of max pooling and average pooling. The number of hidden layer neurons in the LSTM is 32 to 128, and the forget gate threshold is 0.1 to 0.3. At the same time, the initial values ​​of the weight coefficients of the Attention mechanism are set. An adaptive parameter initialization algorithm is adopted to avoid the model getting trapped in local optima. The algorithm first determines the range of initialization parameters by statistically analyzing the feature dimension and variance of the data set, and then dynamically adjusts them according to the GO micro-defect level and the low-carbon target weight: when the GO micro-defect level is low (defect density ≤ 3%) and the low-carbon target weight is ≥ 0.4, the initialization parameters are set to the low to medium value of the range to reduce the initial complexity of the model; when the GO micro-defect level is high (defect density > 5%) or the low-carbon target weight is < 0.2, the initialization parameters are set to the high to enhance the model's feature capture ability and ensure that the initialization parameters are adapted to the data characteristics.

[0062] The number and size of convolutional kernels are determined based on the complexity of GO micro-features and the dimension of input features: when the GO micro-defect density is >5% or the feature dimension is ≥30, select 48-64 convolutional kernels, paired with 5×5-7×7 convolutional kernels; when the GO microstructure is uniform (defect density ≤5%) and the feature dimension is <30, select 16-32 convolutional kernels, paired with 3×3-5×5 convolutional kernels, to balance feature extraction accuracy and computational efficiency.

[0063] The model output layer adopts a multi-output structure, which simultaneously predicts five indicators of GO-UHPC: bond strength, slip, interface damage, bond failure warning threshold, and low-carbon optimization potential. This meets the dual requirements of multi-indicator prediction of bond performance, failure warning, and low-carbon optimization in engineering design. The model uses quantization compression technology to simplify redundant parameters and reduce computational complexity, ensuring that it can adapt to the lightweight deployment requirements of engineering sites.

[0064] The early warning threshold for bond failure is based on the bond strength attenuation rate and the degree of interface damage: when the bond strength attenuation rate is ≥30% and the quantitative value of interface damage is ≥0.7, it is judged as near failure, and the early warning threshold is set at 70% of the peak bond strength under this condition; when the bond strength attenuation rate is <30% but the quantitative value of interface damage is ≥0.5, the early warning threshold is set at 85% of the peak bond strength to ensure the timeliness and accuracy of the early warning.

[0065] The model training phase employs a collaborative training strategy of hierarchical training, dynamic hyperparameter optimization, and transfer learning to improve the model's training effect, stability, and generalization ability. Hierarchical training divides the model training into three levels: feature extraction layer training, attention layer training, and prediction layer training. First, the CNN and LSTM feature extraction layers are trained using a preprocessed dataset to optimize feature extraction capabilities. Then, the dual-branch Attention layer is trained to optimize feature weight allocation. Finally, the extracted features are input into the prediction layer for joint training to improve training accuracy.

[0066] Hyperparameter optimization employs a combination of Bayesian optimization and reinforcement learning to dynamically optimize the model's hyperparameters, including CNN convolutional kernel parameters, the number of neurons in the LSTM hidden layer, learning rate, regularization coefficient, and Dropout probability. After each round of optimization, the model performance is jointly verified using 10-fold cross-validation and leave-one-out validation until the optimal hyperparameter combination is found. An adaptive hyperparameter adjustment mechanism is then used to adjust the hyperparameters in real time based on the training error, thereby improving the model's generalization ability and prediction accuracy.

[0067] Hyperparameter tuning is based on the rate of change of prediction error on the validation set: when the error decrease rate is ≥10% for three consecutive training rounds, the current learning rate is kept unchanged; when the error decrease rate is between 3% and 10%, the learning rate is adjusted to 0.8 times the original value; when the error decrease rate is <3%, the learning rate is adjusted to 0.5 times the original value, and the Dropout probability is increased by 0.05; when the error increase rate is ≥5%, hyperparameter backtracking is triggered to restore the optimal hyperparameter combination of the previous three rounds to ensure the stability of model training.

[0068] An adaptive Dropout layer with different Dropout probabilities and L2 regularization are adopted with different feature layers. At the same time, an early stopping mechanism and model ensemble strategy are combined to integrate the prediction results of three models with different initial parameters. Training is stopped when the prediction error of the validation set does not decrease for several consecutive rounds, which effectively suppresses model overfitting and ensures the prediction performance of the model in new samples and complex scenarios.

[0069] The integrated approach uses a combination of weighted average and majority voting. For continuous indicators such as bond strength and slip, weights are allocated according to the proportion of the validation set R² of the three models. The higher the R², the greater the weight, and the total weight is 1. The weighted average result is then calculated. For categorical indicators such as bond failure warning, the majority voting principle is adopted. If two or more of the three models are determined to be close to failure, a warning result is output, thereby improving the stability and reliability of the prediction results.

[0070] To address the issue of insufficient samples in complex scenarios such as marine engineering and cold-region engineering, transfer learning uses a general UHPC performance dataset to train a basic model, and then uses the GO-UHPC small sample dataset to fine-tune the basic model to achieve model knowledge transfer and improve prediction accuracy in complex scenarios.

[0071] The general UHPC performance dataset should contain ≥5000 samples, covering common mixing ratios, maintenance conditions, and measured bonding performance values. The learning rate is set to 0.001 during the basic model training, and the training epochs are ≥200. In the fine-tuning stage, a strategy of freezing the feature extraction layer and fine-tuning the prediction layer is adopted, the learning rate is reduced to 0.0001, and the training epochs are 50-80. The stopping condition is the convergence of the prediction error of the GO-UHPC dataset (error fluctuation ≤0.5% for 10 consecutive epochs) to ensure the effectiveness of knowledge transfer.

[0072] The verification and correction phase constructs a verification system with multiple scenarios, multiple indicators, and multiple dimensions. Combining the engineering application scenarios of GO-UHPC, a dual adaptive error correction strategy of scenario and low carbon is designed. When dividing the dataset, the preprocessed multi-source heterogeneous dataset is divided into training set, test set and verification set in a ratio of 7:2:1. The training set is used for model training and hyperparameter optimization, the test set is used to initially verify the model prediction performance, and the verification set is used to verify the stability and generalization ability of the model. At the same time, a special dataset for extreme scenarios is set up to separately verify the prediction performance of the model in extreme environments.

[0073] Multi-index validation uses four core indicators—coefficient of determination (R²), mean absolute error (MAE), root mean square error (RMSE), and relative error (RE)—to evaluate the model's prediction accuracy. Validation standards are set, requiring the model to have R² ≥ 0.97, MAE ≤ 1.5, RMSE ≤ 2.5, and RE ≤ 5% on the test set, and R² ≥ 0.95 on the extreme scenario validation set. Simultaneously, a low-carbon adaptability index is set to evaluate the fit between the model's prediction results and low-carbon goals. Multi-scenario validation selects five typical engineering scenarios: general infrastructure, marine engineering, high-temperature and high-humidity environment, low-temperature freeze-thaw environment, and salt spray corrosion environment. Test data on the GO-UHPC bonding performance and low-carbon data are collected for each scenario to verify the model's generalization ability and low-carbon adaptability in different scenarios.

[0074] The low-carbon compatibility index is defined as a comprehensive score of 'adhesion performance compliance rate - low-carbon cost coefficient', with a full score of 100 points. The adhesion performance compliance rate (accounting for 60 points) is based on a test set R² ≥ 0.97 as the full score benchmark, with 5 points deducted for every 0.01 decrease. The low-carbon cost coefficient (accounting for 40 points) is based on carbon emissions ≤ 500 kg / m³ and admixture replacement rate ≥ 30% as the full score benchmark, with 3 points deducted for every 50 kg / m³ increase in carbon emissions and 4 points deducted for every 5% decrease in admixture replacement rate. A comprehensive score ≥ 80 points indicates that the low-carbon compatibility is qualified.

[0075] Error correction addresses prediction errors under different scenarios and low-carbon levels by constructing a scenario- and low-carbon dual adaptive error correction model. Through SHAP analysis of error sources, error correction functions for different scenarios and low-carbon levels are established to dynamically correct the model's prediction results. At the same time, an error feedback mechanism is set up to use the corrected results back to fine-tune the model, further improving the prediction accuracy of the model under different scenarios and low-carbon levels, and meeting the actual needs of engineering.

[0076] Among them, the engineering adaptation phase designs an engineering adaptation module that integrates interpretability, engineering, and low carbonization. The interpretability module adopts the SHAP and LIME dual algorithms to perform interpretability analysis on the model's prediction results, calculate the contribution value and influence weight of each input feature (especially GO micro features and low carbon features) on the predicted bonding performance and low carbon optimization potential, clarify the influence law and weight ranking of key factors such as GO content, water-cement ratio, interface roughness, and admixture substitution rate, and generate a visual interpretation report that includes feature contribution charts and low carbon optimization suggestion charts.

[0077] The visualization report comprises three core modules: First, a feature weight ranking chart (bar chart format), showing the contribution percentage of the top 10 key features to each prediction indicator; second, a factor influence trend chart (line chart format), presenting the dynamic influence of core features such as GO content and water-cement ratio on bonding performance and low-carbon potential within reasonable value ranges; and third, a low-carbon optimization scheme comparison table, listing three sets of parameter combinations with different priorities—high strength priority, low carbon priority, and balanced priority—and their corresponding prediction results, facilitating engineers to select the appropriate option as needed.

[0078] The engineering interface encapsulates the trained and optimized hybrid machine learning model into a lightweight engineering module. It features dual interfaces for PCs and mobile devices, allowing engineers to input actual engineering parameters such as GO micro-parameters, UHPC mix proportions, curing conditions, and low-carbon goals. The module outputs predicted results including bond strength, slippage, interface damage level, bond failure warnings, and low-carbon optimization potential. The response time is ≤0.8s. This response time is based on the following hardware environment: the PC requires an Intel Core i5 or higher processor and 8GB or more of RAM; the mobile device requires a Snapdragon 855 or higher processor and 6GB or more of RAM. Model quantization and compression further ensure response speed under different hardware environments during deployment.

[0079] It also supports batch data input, result export and printing functions. The parameter recommendation function is based on the model's prediction results, feature contribution analysis and low-carbon goals. For different engineering needs, such as high-strength bonding, high-durability bonding, green low-carbon bonding and extreme environment bonding, it automatically recommends the optimal GO content, UHPC mix ratio, interface treatment scheme and low-carbon optimization scheme. For example, it recommends the mix ratio with the lowest carbon emissions while ensuring bonding strength.

[0080] The field calibration module allows engineers to input actual field measurement data to calibrate the model's prediction results in real time, correcting errors caused by differences between the field environment and the laboratory environment, further improving the model's accuracy in field applications, and also supporting the automatic uploading of calibration data to the dataset.

[0081] The on-site calibration adopts the deviation compensation logic between measured and predicted values: when the deviation between the measured on-site bonding strength and the model prediction value is ≤5%, the model parameters are not adjusted; when the deviation is between 5% and 10%, the input weights of the corresponding environmental parameters (temperature and humidity) are linearly corrected according to the deviation ratio; when the deviation is >10%, local parameter fine-tuning is triggered, and the weight ratio of the environmental feature layer in the dual-branch attention mechanism is adjusted based on the deviation between the measured and predicted values. At the same time, the calibration data is marked as high priority for the next round of model iteration.

[0082] The update and iteration phase establishes a dynamic update and iteration mechanism for data updates, model iterations, version management, and low-carbon adaptation. The data update interface adopts a three-in-one design of automatic data collection, manual input, and federated learning collaborative updates, which can access new GO-UHPC bonding performance test data, engineering application data, low-carbon data, and extreme environment test data in real time. At the same time, it can realize data collaborative updates of multiple laboratories and multiple engineering units through federated learning, expanding the coverage, sample size, and low-carbon scenarios of the dataset.

[0083] The model iteration is set to a regular iteration cycle of 6 months and a triggered iteration mechanism. When the amount of new data reaches a threshold, the model prediction error exceeds a threshold, or a new GO material, a new UHPC mix ratio, or a new engineering scenario appears, the iteration is automatically triggered.

[0084] The specific trigger threshold is set as follows: the amount of new data is ≥ 30% of the total amount of the original dataset; the average relative error (RE) of the model in 5 consecutive engineering cases is > 8%; a new GO material (the difference between the thickness of the sheet and the content of oxygen-containing functional groups and the existing data is ≥ 20%), a new UHPC mix ratio (the water-cement ratio fluctuates by ±0.1 or the steel fiber content fluctuates by ±2%), or a new engineering scenario (such as a high-altitude strong ultraviolet environment) appears. If any of these conditions are met, the iteration will be triggered.

[0085] The uniformity of GO dispersion was quantified by scanning electron microscopy image analysis: Three different observation areas (each area ≥ 10 μm × 10 μm) were selected, and the coefficient of variation of the distribution density of GO sheets was statistically analyzed. A coefficient of variation ≤ 0.2 was considered uniform dispersion, with a quantification value of 0.8-1.0; a coefficient of variation 0.21-0.4 was considered relatively uniform dispersion, with a quantification value of 0.4-0.79; and a coefficient of variation > 0.4 was considered non-uniform dispersion, with a quantification value of 0.0-0.39. This achieved a quantitative characterization of the dispersion state.

[0086] The existing model is retrained and its hyperparameters are optimized using new data and transfer learning. The model structure, feature system and hyperparameters are adjusted in combination with changes in new materials, new engineering scenarios and low-carbon goals to continuously improve the model’s prediction accuracy, generalization ability and low-carbon adaptability. A model version management mechanism and scenario and low-carbon tag system are established to record the model parameters, performance indicators, applicable scenarios, low-carbon adaptability level and improvement content of each iteration.

[0087] Version management uses a major version number, minor version number, and revision number numbering rule, such as V1.0.0. The major version number corresponds to a major adjustment to the model structure, the minor version number corresponds to the optimization of the feature system, and the revision number corresponds to the fine-tuning of hyperparameters. The recorded content is stored in a structured document format, including version number, iteration time, training dataset information, core performance indicators (R², MAE, etc.), a list of applicable scenarios, low-carbon adaptation level (three levels: A / B / C), and improvement descriptions. It supports multi-dimensional querying and retrieval by version number, applicable scenario, and iteration time.

[0088] The technology integration and iteration regularly integrates the latest technological achievements in the fields of new materials (such as novel graphene oxide derivatives), artificial intelligence (such as novel deep learning algorithms and attention mechanisms), and low-carbon building materials to upgrade and optimize the model structure, feature system, and engineering adaptation modules.

[0089] This invention also provides a machine learning-based GO-UHPC adhesion prediction system, which, based on the above method, includes:

[0090] The data acquisition and preprocessing module includes a parameter acquisition unit and a data processing unit. It acquires the microscopic characteristic parameters of GO through scanning electron microscope and X-ray diffraction equipment, and acquires the macroscopic mixing ratio, interface treatment, environmental curing parameters and measured values ​​of bonding performance of UHPC through experiments. It completes outlier removal, missing value filling and dimension elimination by using the Laida criterion, K-nearest neighbor interpolation method and Z-score standardization method, and outputs a multi-source heterogeneous dataset.

[0091] The feature processing module adopts a hierarchical screening-coupling construction strategy. It uses mutual information method, grey relational analysis and variance analysis to complete the initial feature screening, coupled feature construction and final feature determination, and outputs a feature set suitable for model training.

[0092] The model building and training module has a built-in CNN-LSTM-Attention hybrid machine learning model. It completes the model structure initialization and hierarchical training, and uses Bayesian optimization and 10-fold cross-validation to achieve dynamic optimization of hyperparameters. It suppresses overfitting through Dropout layer, L2 regularization and early stopping mechanism, and outputs the best prediction model after training.

[0093] The validation and correction module divides the dataset into training, testing and validation sets according to a set ratio. It validates the model performance through multiple metrics such as R², MAE, and RMSE and multiple engineering scenarios. It also has a built-in scene adaptive error correction function to dynamically correct the prediction results.

[0094] The engineering adaptation module includes an interpretability unit, an interface unit, and a parameter recommendation unit. It generates a visual explanation report through the SHAP algorithm, provides input and output interfaces, and can automatically recommend the optimal parameter scheme.

[0095] The dynamic update module is designed with an automatic data update interface and a periodic iteration mechanism. It can integrate new data to complete model retraining and optimization, and establish a version management unit.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based GO-UHPC adhesion prediction method, characterized in that, include: In the data construction phase, a multi-source heterogeneous dataset was constructed, including GO micro-characteristics, UHPC macro-parameters, interface processing parameters, environmental parameters, measured values ​​of bonding performance, and low-carbon parameters. Federated learning was used to achieve collaborative data sharing, parameter determination and data preprocessing were completed, and a dataset labeling system was established. In the feature optimization stage, a comprehensive strategy of hierarchical screening, association construction and dynamic verification is adopted to hierarchically divide the features of the dataset and construct core association features, determine the optimal model input features, and form a feature system adapted to GO-UHPC bonding prediction. In the model building phase, a hybrid machine learning prediction model combining CNN, LSTM, Attention and residual connections is constructed by combining the diverse characteristics of GO-UHPC bonding performance. A dual-branch attention mechanism is designed to optimize feature extraction and weight allocation. Initial parameters of the model are set, and a multi-output structure is adopted to achieve simultaneous prediction of multiple indicators and simplify redundant parameters of the model. During the model training phase, a collaborative training strategy combining hierarchical training, dynamic hyperparameter optimization, and transfer learning is adopted to complete hierarchical training and dynamic hyperparameter optimization of the model, thereby suppressing model overfitting. In the verification and correction phase, a model verification system is constructed to complete the model performance verification under different engineering scenarios and extreme scenarios. A dual adaptive error correction strategy based on scenario and low carbon is adopted to improve the model prediction accuracy. During the engineering adaptation phase, an engineering adaptation module integrating interpretability, engineering feasibility, and low carbon footprint is designed to achieve interpretability analysis and visualization of the model, and to complete the deployment of engineering interfaces, parameter recommendation, and on-site calibration functions. During the update and iteration phase, a dynamic update and iteration mechanism is established for data updates, model iterations, version management, and low-carbon adaptation to achieve dataset expansion and model iteration optimization.

2. The GO-UHPC adhesion prediction method based on machine learning according to claim 1, characterized in that, In the data construction phase, the microscopic characteristic parameters of GO in the multi-source heterogeneous dataset include GO sheet thickness, number of layers, type and content of oxygen-containing functional groups, dispersion uniformity, and quantitative characterization parameters of microscopic defects; UHPC macroscopic parameters include mix proportion parameters and low-carbon index interface treatment parameters including interface roughness parameters, aging performance parameters of interface agents, diameter and surface condition of reinforcing bars, type and dosage of interface agents; environmental parameters include environmental curing parameters and extreme environment simulation parameters; and the measured bond performance values ​​are quantitative indicators obtained by combining standard pull-out tests with microscopic tests. Data preprocessing includes using dual outlier removal and joint interpolation to process the data, eliminating the influence of dimensions through standardization, and completing the preprocessing by combining data normalization and outlier backtracking verification; the federated learning framework is used to realize multi-agent data sharing and collaborative expansion, and the dataset labeling system labels the scene type, low-carbon level and GO micro-defect level of each set of data.

3. The GO-UHPC adhesion prediction method based on machine learning according to claim 2, characterized in that, In the feature optimization stage, the five-dimensional correlation feature system is divided into the GO micro-feature layer, the UHPC macro-feature layer, the interface parameter layer, the environmental feature layer, and the low-carbon feature layer; the core correlation features include the correlation features between GO micro-properties and UHPC macro-parameters, GO micro-properties and interface parameters, environmental parameters and interface parameters, and low-carbon parameters and bonding performance. The optimal model input features were determined through a triple screening method of mutual information, grey relational analysis and variance analysis. Feature interaction verification was combined to ensure that the feature system was suitable for GO-UHPC bonding prediction.

4. The GO-UHPC adhesion prediction method based on machine learning according to claim 3, characterized in that, In the model building phase, the hybrid machine learning prediction model optimizes gradient propagation through residual connections, and uses a dual-branch attention mechanism to optimize the feature extraction effect and weight allocation of CNN and LSTM, focusing on the weight adaptation of key features and low-carbon features. The initial parameters of the model are set using a parameter adaptive initialization algorithm. The output layer is a multi-output structure, which simultaneously predicts five indicators: bond strength, slip, interface damage degree, bond failure warning threshold, and low-carbon optimization potential. Quantization compression technology is used to simplify redundant parameters and reduce computational complexity.

5. The GO-UHPC adhesion prediction method based on machine learning according to claim 4, characterized in that, In the model training phase, the hierarchical training is divided into three levels: feature extraction layer, attention layer, and prediction layer. After each layer is trained in sequence, joint training is performed. Hyperparameter optimization adopts a combination of Bayesian optimization and reinforcement learning, and combines multi-fold cross-validation and leave-one-out validation to determine the optimal hyperparameter combination. Overfitting is suppressed by adaptive Dropout layer, L2 regularization, early stopping mechanism, and model ensemble strategy. Transfer learning trains a base model using the general UHPC performance dataset and then fine-tunes the model using the GO-UHPC small sample dataset to achieve knowledge transfer.

6. The GO-UHPC adhesion prediction method based on machine learning according to claim 5, characterized in that, In the verification and correction phase, the dataset is divided into a training set, a test set, and a verification set according to a preset ratio, which are used for model training, preliminary verification, and stability and generalization ability verification, respectively. An extreme scenario-specific dataset is added to verify the model's adaptability to extreme environments. Model performance is evaluated using four core indicators: coefficient of determination, mean absolute error, root mean square error, and relative error, as well as low-carbon adaptability indicators. Multi-scenario validation covers five typical engineering scenarios. Error correction is achieved by analyzing error sources through SHAP, constructing a dual adaptive error correction model, and combining it with an error feedback mechanism to fine-tune the model in reverse.

7. The GO-UHPC adhesion prediction method based on machine learning according to claim 6, characterized in that, During the engineering adaptation phase, the interpretability module uses the SHAP and LIME dual algorithms to analyze the contribution value and influence weight of each input feature and generate a visual explanation report. The engineering interface is encapsulated into a lightweight module, supporting multi-terminal access, parameter input and output, batch processing, and result export and printing. The parameter recommendation function automatically recommends the optimal parameters and low-carbon optimization schemes based on model prediction results, feature contribution analysis, and low-carbon goals. The on-site calibration module supports input of actual on-site measurement data, corrects errors caused by environmental differences, and automatically uploads calibration data.

8. The GO-UHPC adhesion prediction method based on machine learning according to claim 7, characterized in that, During the update and iteration phase, data updates are conducted through a combination of automatic collection, manual input, and federated learning to expand the dataset coverage and sample size. Model iteration combines periodic iteration with triggered iteration, optimizing the model in light of new data, new materials, new scenarios, and changes in low-carbon goals. The version management unit records the model parameters, performance indicators and improvements for each iteration, and regularly integrates the latest technological achievements in related fields to upgrade and optimize the model and engineering adaptation modules; The technology integration and iteration regularly incorporates the latest technological achievements in the fields of new materials, artificial intelligence, and low-carbon building materials to upgrade and optimize the model structure, feature system, and engineering adaptation modules.

9. A machine learning-based GO-UHPC adhesion prediction system, based on the method of claim 8, characterized in that, include: The data acquisition and preprocessing module is used to collect various parameters, including GO micro-characteristics and UHPC macro-parameters, and outputs multi-source heterogeneous datasets through outlier removal, missing value imputation, and standardization. The feature processing module adopts a hierarchical screening-coupling construction strategy. It completes feature screening, coupling construction and final feature determination through multiple screening methods, and outputs a feature set that is suitable for model training. The model building and training module has a built-in CNN-LSTM-Attention hybrid machine learning model to complete model initialization, hierarchical training, hyperparameter optimization and overfitting suppression, and output the optimal prediction model. The validation and correction module divides the dataset and validates the model performance through multiple indicators and scenarios, and uses a dual adaptive error correction model to dynamically correct the prediction results. The engineering adaptation module includes interpretability, interface, and parameter recommendation units, enabling visual explanation, engineering interface deployment, and optimal parameter recommendation. The dynamic update module features automatic data updates, regular model iterations, and version management, and supports the integration of new data to optimize the model.

Images