Material property and carbon emission co-optimization prediction method and system for guardrail design

Abstract

The application discloses a kind of material performance and carbon emission collaborative prediction method for guardrail design, belong to the field of transportation infrastructure construction.It includes the following steps: collecting relevant data, converting category type features in data into model embedding vectors, and combining physical heuristic combined features and numerical features, obtaining an enhanced feature matrix after adaptive multi-kernel transformation;The enhanced feature matrix obtains the hidden layer representation matrix through the shared feature extraction layer;The hidden layer representation matrix is input into the material performance preliminary prediction branch and the carbon emission preliminary prediction branch to obtain the preliminary prediction result;The collaborative fusion layer uses trainable collaborative matrix and bias vector to collaboratively adjust the preliminary prediction of the two branches to obtain the collaborative prediction result;According to the collaborative prediction result, it is determined whether the scheme is feasible.The application can fully exploit the nonlinear relationship between data, significantly improve the generalization ability and prediction accuracy of the model.

Description

Technical Field

[0001] This invention belongs to the field of transportation infrastructure construction, and specifically relates to a method and system for synergistic prediction of material properties and carbon emissions for guardrail design. Background Technology

[0002] As a crucial protective facility ensuring driving safety, guardrails' design directly impacts the lives and property of road users. With the continuous growth of car ownership and the diversification of vehicle types, the road traffic environment is becoming increasingly complex, placing higher demands on the protective performance of guardrail structures. Simultaneously, the global climate change problem is becoming increasingly severe, and all industries are actively seeking effective ways to reduce carbon emissions. Infrastructure construction, as a significant source of carbon emissions, generates substantial carbon emissions from material production and component manufacturing. Against this backdrop, guardrail design faces a dual challenge: ensuring the structure possesses sufficient mechanical strength to withstand various complex collision conditions while minimizing the environmental impact of material use, achieving a balance between safety and environmental performance.

[0003] Existing technologies have significant shortcomings when facing performance prediction and carbon emission prediction: In terms of feature processing, existing technologies typically directly concatenate the raw data such as model number, yield strength, and cross-sectional dimensions into the model. This fails to transform discrete models into continuous representations containing semantic information, nor does it introduce any artificial features based on physical formulas. As a result, the model struggles to understand the mechanical essence behind the data, limiting prediction accuracy when facing complex working conditions. Furthermore, when dealing with the nonlinear relationship between material parameters and structural response, existing technologies, such as linear regression or support vector machines with fixed kernel functions, cannot capture the sensitivity changes of features across different value ranges. For example, the influence of yield strength on impact force may differ significantly between low and high strength regions, and traditional models cannot adaptively learn such local nonlinear features. In terms of model construction, most existing technologies separate material performance prediction and carbon emission accounting into two independent tasks, establishing separate prediction models. This approach ignores the inherent physical coupling between material strength, structural stiffness, material usage, and carbon emissions. This leads to situations where high-performance solutions may exceed carbon emission limits, while low-carbon solutions may fail to meet mechanical requirements, making it impossible to achieve synergistic optimization of performance and environmental protection during the design phase. At the same time, the training objective of existing models usually only focuses on minimizing the error between the predicted value and the true label, such as simply using the mean squared error as the loss function. This approach makes the model completely dependent on the distribution of training data, which can easily produce prediction results that violate physical laws in sparse data regions, such as unreasonable situations where the energy absorption capacity is much greater than the collision input energy. The generalization ability and reliability of the model are difficult to guarantee. Summary of the Invention

[0004] To address the shortcomings of existing technologies, such as the lack of prior knowledge in data processing and the inability of models to capture sensitive changes in features during prediction, this invention provides a collaborative prediction method for material properties and carbon emissions in guardrail design through multi-dimensional mechanism innovation and process optimization. This method enhances the physical meaning of data and improves the model's feature perception capabilities, significantly increasing the generalization ability and prediction accuracy of the prediction model. Furthermore, this invention also provides a system for implementing this method and a computer-readable storage medium, broadening the application scope of the technical solution and providing a universal and efficient technical means for performance evaluation and carbon emission accounting in the guardrail design stage.

[0005] This invention constructs a multi-source feature enhancement mechanism integrating physical priors, a collaborative prediction model architecture with dual-task coupling, and a composite loss function system with multi-constraint fusion. It breaks through the limitations of existing technologies in three dimensions: feature processing, model construction, and training optimization. First, it explicitly integrates the physical laws of materials mechanics and carbon emission accounting into the entire feature processing process, achieving effective vectorization of categorical features and nonlinear adaptive representation of numerical features, thus enhancing the expressive power of feature information. Second, it builds a three-order model architecture of "shared feature extraction layer + dual prediction branches + collaborative fusion layer," realizing mutual constraints, deep coupling, and collaborative optimization between material performance and carbon emission prediction. Third, it designs a composite loss function that integrates multi-task weighted loss and physical consistency loss, achieving dynamic self-balancing of multi-task prediction weights and forcing prediction results to conform to the basic physical laws of materials mechanics and carbon emission accounting, significantly improving the model's generalization ability, robustness, and interpretability.

[0006] The specific solution adopted in this invention is as follows: This invention provides a method for synergistic prediction of material properties and carbon emissions for guardrail design, comprising the following steps: S1. Data Collection: Collect sample data related to guardrails, including categorical features and numerical features. Organize the collected data according to a unified format to form the original dataset. S2. Feature Enhancement: Convert categorical features into model embedding vectors, construct combined features by combining the physical laws of materials mechanics and carbon emission accounting, fuse the combined features, numerical features and vectorized categorical features to obtain enhanced feature vectors, perform nonlinear mapping on the enhanced features through adaptive multi-kernel transformation, and construct the enhanced feature matrix as the training dataset. S3. Model Construction: A shared feature extraction layer is constructed based on gated linear units. The hidden layer representation matrix is ​​obtained through residual connection and layer normalization. The hidden layer representation matrix is ​​input into the preliminary prediction branch of material properties and the preliminary prediction branch of carbon emissions, respectively, to obtain the preliminary prediction matrix of material properties and the preliminary prediction matrix of carbon emissions. The preliminary prediction of the two branches is adjusted collaboratively using a trainable collaborative matrix and bias vector to obtain the collaborative prediction result. S4. Model Training: Construct a loss function that includes two parts: multi-task uncertainty weighted loss and physical consistency loss. Input the training dataset into the constructed model, use an optimization algorithm to iteratively update all trainable parameters in the model, gradually reduce the loss function value until the model converges, and save the converged model parameters. S5. Scheme Evaluation: Collect data on the new guardrail design scheme. After the data undergoes the feature enhancement steps described above, a feature enhancement matrix is ​​obtained. The feature enhancement matrix is ​​then input into the trained prediction model to obtain material performance prediction results and carbon emission prediction results. The performance and carbon emission accounting are evaluated using the prediction results of the new guardrail design scheme.

[0007] As a further technical solution, the standardization format organization in S1 specifically includes: Data was collected covering guardrail model information, material mechanical parameters, geometric dimension parameters, collision condition parameters, and carbon emission-related data. During processing, categorical features were assigned unique integer indices, while numerical features directly used their original values. Each sample was labeled with two categories of tags: material performance indicators and carbon emission indicators, resulting in the original dataset.

[0008] As a further technical solution, the entity embedding of categorical features in S2 specifically includes: Assign a unique integer index to each model and construct an embedding matrix where each row corresponds to an initial embedding vector for a model. These vectors are randomly initialized at the start of training. One sample, its model index This is used to query the corresponding row from the embedding matrix, thus obtaining the model embedding vector of the sample. .

[0009] As a further technical solution, the adaptive multi-kernel transformation in S2 specifically includes: For each feature dimension, a kernel function is used to model local sensitivity. The center and width of the kernel function are determined by data, adaptively learning the local sensitivity of the feature in different value ranges, specifically expressed as follows: In the formula, Indicates the first The feature is processed by the first The mapping values ​​of each kernel function, Indicates feature index, Indicates the kernel function index. Indicates the first The first feature The weight of each core, Represents the Gaussian radial basis function, which maps the input to... interval, Represents enhanced feature vectors The One portion, Indicates the first The first feature The center of each kernel function, Indicates the first The first feature The width of each kernel function; All kernel function mapping values ​​obtained by adaptive multi-kernel transformation for each sample are concatenated into a vector. The vectors of all samples are stacked row by row to obtain the enhanced feature matrix.

[0010] As a further technical solution, model construction in S3 specifically includes: Construct a shared feature extraction layer: A shared feature extraction layer based on gated linear units is constructed. This layer enhances training stability through residual connections and layer normalization, thereby obtaining the hidden layer representation, as follows: In the formula, The hidden layer representation matrix is ​​represented. Represents the enhanced feature matrix, This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer. This represents the linear transformation weight matrix in the residual connection. Presentation layer normalization operation, This represents the activation function of the linear rectifier unit; Preliminary prediction of material properties branch: The shared feature matrix is ​​mapped to the material property prediction space through a linear layer. The output is guaranteed to be positive by a smooth activation function. At the same time, the physical heuristic combined feature matrix is ​​mapped to a modulation factor through a trainable physical feature modulation matrix. After activation by a hyperbolic tangent activation function, it is added to an all-1 vector. Finally, it is multiplied element-wise with the preliminary prediction to obtain the preliminary prediction matrix of material properties. Preliminary carbon emission forecasting branch: The shared feature matrix is ​​activated by the linear rectifier unit after the linear layer to obtain a non-negative basic carbon emission prediction value. The physical correction part uses the carbon emission factor as a physical prior and generates correction terms through a gating mechanism. Finally, the two parts are fused to obtain the preliminary carbon emission prediction matrix. Collaborative Integration Layer: The initial prediction matrices of the two branches are concatenated column-wise to obtain the concatenated matrix, which is then subjected to a linear transformation and mapped to the hyperbolic tangent activation function. The interval is used to obtain the modulation factor, and this modulation factor is then compared with the all-1 vector. Adding them together gives The scaling factor of the range is then multiplied element-wise with the concatenated preliminary prediction to obtain the co-prediction result.

[0011] As a further technical solution, model training in S4 specifically includes: Loss function: Construct a loss function that includes two parts: multi-task uncertainty weighted loss and physical consistency loss, and obtain the difference between the predicted value and the true value; Forward propagation: The input enhanced feature matrix is ​​passed through a shared feature extraction layer to obtain the hidden layer representation matrix; then, the preliminary prediction matrix is ​​obtained through the preliminary prediction branch of material properties and the preliminary prediction branch of carbon emissions, respectively; finally, the collaborative prediction matrix is ​​obtained through a collaborative fusion layer. Backpropagation: After forward propagation is completed, the loss function value is calculated, and then the backpropagation algorithm is executed to calculate the gradient of the loss function with respect to each trainable parameter. The parameter values ​​are then updated using an optimization algorithm.

[0012] As a further technical solution, the evaluation of the solution in S5 specifically includes: Collect the input features of the new design scheme and construct an enhanced feature matrix; The enhanced feature matrix is ​​input into the trained model to obtain the collaborative prediction result; Based on the prediction results, determine whether the guardrail can withstand the impact load under the design conditions, whether it meets the deformation limit requirements, and assess the environmental impact of the design scheme.

[0013] This invention also provides a system for the coordinated prediction of material properties and carbon emissions for guardrail design, specifically including: Data acquisition module: used to collect information related to guardrails and organize the guardrail information into a unified standardized format to obtain the raw dataset; Feature enhancement module: By introducing entity embedding, physical relationships in materials mechanics and carbon emission accounting, and adaptive multi-kernel transformation, the original dataset is converted into an enhanced feature matrix; Model building module: used to build a collaborative prediction model of material properties and carbon emissions for guardrail design that incorporates physically heuristic combined features and carbon emission factors; Model training module: Iteratively updates the updatable parameters through optimization algorithms to obtain the optimal parameters of the prediction model; Solution Evaluation Module: Evaluate the performance of the guardrail design in the new solution based on the predicted results, and whether the carbon emissions meet the standards.

[0014] The present invention also provides a computer-readable storage medium storing a computer program that can be executed by a processor to implement a method for synergistic prediction of material properties and carbon emissions for guardrail design as described above.

[0015] Beneficial effects of this invention: This invention proposes a multi-source feature enhancement mechanism that integrates physical priors. It constructs embedded vectors by embedding discrete guardrail models into entities, and simultaneously constructs combined features with clear physical meaning based on materials mechanics and carbon emission calculation formulas. An adaptive multi-kernel transformation is used for mapping, effectively handling complex nonlinear mapping relationships while explicitly integrating domain prior knowledge into the model input. This solves the problem that purely data-driven models struggle to learn physical laws, significantly improving the model's prediction accuracy. A three-order model architecture of "shared feature extraction layer + dual prediction branches + collaborative fusion layer" is constructed. In the carbon emission branch, a gating mechanism is used to treat the carbon emission factor as a strong prior. This paper integrates empirical and data-driven prediction, and finally captures the intrinsic coupling relationship between material properties and carbon emissions through a trainable collaborative matrix, realizing the mutual constraint and collaborative optimization of the two tasks, and effectively improving the overall prediction accuracy. It proposes a composite loss function that integrates multi-task uncertainty weighting and physical consistency. On the one hand, it introduces learnable noise parameters to automatically balance the loss weights of the two tasks of material property prediction and carbon emission prediction, avoiding manual parameter tuning. On the other hand, it forces the predicted energy absorption capacity to approximate the theoretical collision energy factor and the carbon emission per unit length to approximate the carbon emission factor, ensuring that the prediction results conform to the basic physical laws and significantly enhancing the model's generalization ability and interpretability. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 A bar chart comparing the accuracy of the method of this invention with three conventional techniques in predicting material properties; Figure 3 This is a comparative analysis of the influence of yield strength variation on the peak impact force between the method of this invention and three conventional methods. Detailed Implementation

[0018] 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.

[0019] Example 1 In this embodiment, the present invention provides a method for synergistic prediction of material properties and carbon emissions for guardrail design, the specific steps of which include: S1. Data Acquisition and Construction of Original Dataset Data collection primarily comes from historical guardrail crash test reports, material performance databases, engineering design specifications, and relevant literature on carbon emission accounting. The specific data collected includes guardrail model information, material mechanical parameters, geometric dimensional parameters, crash test parameters, and carbon emission data. Guardrail model information records the specific model names of different guardrail products, serving as a crucial categorical characteristic for distinguishing different types of guardrails. Material mechanical parameters include the yield strength, elastic modulus, and density of the steel; these parameters directly determine the mechanical response behavior of the guardrail. Geometric dimensional parameters mainly collect the moment of inertia and cross-sectional area of ​​the guardrail section; these two indicators reflect the cross-sectional shape characteristics of the guardrail and have a significant impact on structural stiffness and material usage. Crash test parameters include loading angle, collision velocity, and collision height; these parameters describe the collision conditions the guardrail may encounter in actual use. Carbon emission data includes the carbon content of the materials, the carbon emission coefficient of the steel, and production process information; these data allow for the calculation of the guardrail's carbon emission level throughout its entire life cycle.

[0020] After data collection, each collected sample is organized according to a standardized format to form the original dataset. Each record in the original dataset corresponds to a crash test or a guardrail design scheme, containing all the characteristic information of that sample. For categorical features such as guardrail model, a unique integer index needs to be assigned, numbered consecutively starting from 1, with a total of M different guardrail models. Material mechanical parameters and geometric dimensional parameters are used directly as their original values ​​as numerical features. Crash condition parameters are also retained as numerical features, retaining their original measured values.

[0021] This invention requires labeling each sample with two types of tags: material performance indicators and carbon emission indicators. Material performance indicators include three specific parameters: peak impact force, energy absorption capacity, and deformation. Peak impact force refers to the maximum impact force experienced by the guardrail during a collision, measured in kilonewtons (kN); energy absorption capacity refers to the total energy absorbed by the guardrail from the moment it contacts the impacting object until it stops moving, measured in kilojoules (kJ); and deformation refers to the maximum permanent deformation displacement of the guardrail after the collision, measured in millimeters (mm). These three indicators are obtained through actual collision tests or, when testing is not possible, through high-precision finite element simulation calculations. Carbon emission indicators include two specific parameters: carbon emissions per unit length and total carbon emissions. Carbon emissions per unit length refer to the carbon dioxide equivalent emissions generated per meter of guardrail during the manufacturing process, measured in kilograms of carbon dioxide per meter; total carbon emissions refer to the total carbon dioxide equivalent emissions generated throughout the entire process of the guardrail component, from raw material extraction to completion of manufacturing, measured in kilograms of carbon dioxide.

[0022] The carbon emission index is calculated based on the amount of material used and the carbon emission coefficient. The specific method is as follows: calculate the material volume based on the cross-sectional area and length of the guardrail, multiply it by the material density to obtain the mass, and finally multiply it by the corresponding carbon emission coefficient to obtain the carbon emission amount.

[0023] Through the above data collection and annotation process, a final system containing... The original dataset contains 8 input features for each sample, including guardrail model index, yield strength, elastic modulus, density, cross section moment of inertia, loading angle, collision velocity, and collision height, as well as 5 output labels, including peak collision force, energy absorption capacity, deformation, carbon emissions per unit length, and total carbon emissions.

[0024] S2, Training Dataset Construction The guardrail model, material mechanical parameters, and carbon emission-related data include categorical and numerical features. The scale differences and nonlinear relationships of different features also limit the model's expressive power. This invention converts model numbers into continuous dense vectors through entity embedding, constructs combined features based on physical formulas, and employs adaptive multi-kernel transformation to perform nonlinear mapping on the features, thereby enhancing the information representation capability of the features. The specific steps are as follows: 1) Entity embedding of categorical features Entity embedding is performed on the categorical features in the original dataset, mapping each model to a trainable dense vector, thereby converting the categorical information into a continuous representation; definition Indicates the first The model embedding vector of each sample, with dimension [ ]. , by the embedding matrix according to the first Model index of each sample The query yielded that the dimension of the embedding matrix is... , are trainable parameters, and each row corresponds to an embedding vector of a model; This represents the total number of model categories, that is, the number of different guardrail models in the dataset. This represents the dimension of the embedding vector, with a value of 16. This represents the sample index, with a value range of [value range missing]. arrive ,in The total number of samples, Indicates the first The model index of each sample, with a value range of [value range missing]. .

[0025] In practice, entity embedding is implemented through a trainable embedding layer, assigning a unique integer index (from 1 to M) to each model, constructing an embedding layer of shape M. The embedding matrix is ​​a set of vectors, where each row corresponds to an initial embedding vector for a model. These vectors are randomly initialized at the start of training. For the _th ... One sample, its model index This is used to query the corresponding row from the embedding matrix, thus obtaining the model embedding vector of the sample. .

[0026] 2) Construction of Physics-Inspired Combinatorial Features By introducing the physical relationship between mechanics of materials and carbon emission accounting, five physical heuristic combination features are constructed: material strength ratio, cross-sectional stiffness index, carbon emission factor, collision energy factor, and angle influence coefficient. These features enhance the ability to express physical laws and are expressed as follows: In the formula, This represents a physically heuristic combined eigenvector, derived from the material strength ratio. Section stiffness index Carbon emission factors Collision energy factor and angle influence coefficient constitute; Indicates the material strength ratio. ,in Yield strength (unit: MPa). The reference yield strength is 235 MPa, and the material strength ratio reflects the strength of the material relative to the reference. Indicates the cross-sectional stiffness index. ,in It represents the elastic modulus (in GPa). Moment of inertia of cross section (unit) ), For reference span, the value is 1m. The section stiffness index measures the section's ability to resist bending deformation. Carbon emission factor (unit) ), ,in Material density (units) ), Cross-sectional area (unit) ), Carbon emission coefficient of steel (unit) The carbon emission factor is calculated based on the carbon content of the material corresponding to the model, and it represents the carbon emission per unit length of guardrail. Represents the collision energy factor. ,in The equivalent impact mass (in kg) is derived from the operating parameters. The collision velocity is expressed in m / s, and the collision energy factor characterizes the amount of energy during the collision process. Indicates the influence coefficient of angle. ,in The loading angle (which needs to be converted to radians) is used to characterize the influence of directional effects on the mechanical response.

[0027] 3) Construct enhanced feature vectors Combining the physical heuristic feature vector with the original numerical feature (yield strength) Elastic modulus ,density Moment of inertia of cross section Loading angle Collision speed Collision height The model embedding vector and the model number embedding vector are concatenated to form an enhanced feature vector. It represents the comprehensive information of each sample after feature enhancement, integrating the original numerical values, physical prior knowledge, and categorical embedding representation, with a dimension of [missing information]. ,Right now .

[0028] 4) Adaptive multi-kernel transformation An adaptive multi-kernel transformation is employed to map the enhanced feature vectors, eliminating feature scale differences and enhancing nonlinear expressive power. For each feature dimension, the adaptive multi-kernel transformation utilizes a kernel function for local sensitivity modeling. The center and width of the kernel function are determined by data-driven factors, enabling adaptive learning of the local sensitivity of features across different value ranges. This better handles the nonlinear distributions in material performance and carbon emission data, as expressed below: In the formula, Indicates the first The feature is processed by the first The mapping values ​​of each kernel function, through trainable center and width adaptive learning of features' local sensitivity in different value ranges, output values ​​in... An interval represents the similarity between the eigenvalue and the center of the kernel. This represents the feature index, with a value range of 100. arrive , This represents the kernel function index, with a value range of 1. arrive , This indicates the number of kernel functions, with a value of 16. Indicates the first The first feature The weights of each kernel are trainable parameters used to control the contribution of different kernel functions to the output; Represents the Gaussian radial basis function, which maps the input to... The interval is used to measure the similarity between the input and the kernel center; Represents enhanced feature vectors The Each component, i.e., the enhanced feature vector The One element; Indicates the first The first feature At the heart of each kernel function are the trainable parameters. Indicates the first The first feature The width of each kernel function is a trainable parameter.

[0029] 5) Construct the enhanced feature matrix All kernel function mapping values ​​obtained by adaptive multi-kernel transformation for each sample (total) (The vectors of each sample are concatenated into a single vector, and the vectors of all samples are stacked row-wise to obtain the enhanced feature matrix.) , dimension ,in The total number of samples, Each row corresponds to a vector formed by concatenating all kernel function mapping values ​​obtained from adaptive multi-kernel transformation for a sample.

[0030] S3. Construct a module for synergistic prediction of material properties and carbon emissions. This module includes a shared feature extraction layer, a preliminary material performance prediction branch, a preliminary carbon emission prediction branch, and a collaborative fusion layer. It fully utilizes physical heuristics to combine features, enabling the prediction results of the two tasks to mutually constrain and collaboratively optimize each other, thus better aligning with the actual physical laws of guardrail design. The specific steps are as follows: 1) Shared feature extraction layer To fully exploit the higher-order nonlinear relationships in the enhanced feature matrix, a shared feature extraction layer based on gated linear units is constructed. This layer enhances training stability through residual connections and layer normalization, thereby obtaining the hidden layer representation, as follows: In the formula, The hidden layer representation matrix represents the hidden layer representation output by the shared feature extraction layer, and has a dimension of . Each row corresponds to the hidden features of a sample, representing the potential patterns related to material properties and carbon emissions abstracted from the enhanced features. The hidden features of each sample integrate deep information from the original data, physical priors, and class embeddings. This represents the weight matrix of the first fully connected layer, with dimension 1. , are trainable parameters used to map input features to the first hidden layer space. This represents the bias vector of the first fully connected layer, with dimension . , are trainable parameters; This represents the weight matrix of the second fully connected layer, with dimension 1. , are trainable parameters used to map the first hidden layer to the second hidden layer. This represents the bias vector of the second fully connected layer, with dimension . , are trainable parameters; This represents the linear transformation weight matrix in the residual connection, with dimension 1. , are trainable parameters used to directly map input features to the output dimension to add residuals; The representation layer normalization operation normalizes the feature dimensions of each sample to stabilize training. This represents the activation function of a linear rectifier unit, introducing nonlinearity; This represents the output dimension of the first fully connected layer (i.e., the number of neurons in the first hidden layer), used for the high-order combination of initial feature extraction, and has a value of 128. This represents the output dimension of the second fully connected layer (i.e., the final hidden layer representation dimension), which serves as the input feature dimension for subsequent task branches, and its value is 64.

[0031] In practice, the shared feature extraction layer consists of two fully connected layers, with residual connections and layer normalization introduced, and the input enhanced feature matrix is... First, it undergoes the first level of linear transformation. , obtain dimensions The intermediate representation is obtained by introducing nonlinearity through the ReLU activation function, followed by a second linear transformation. , obtain dimensions Features; Input Enhanced Feature Matrix Through linear transformation in residual connection Directly mapped to the same dimension Add the two parts together, and then perform layer normalization to obtain the hidden layer representation matrix. .

[0032] 2) Preliminary prediction of material properties The hidden layer representation matrix is ​​input into the preliminary prediction branch of material properties, outputting multiple key indicators (peak impact force, energy absorption capacity, and deformation). Simultaneously, physical priors are introduced, and physical heuristic combined features are used to modulate the preliminary prediction, making the results more consistent with the laws of material mechanics. This is represented as follows: In the formula, This represents the preliminary prediction matrix for material properties, with dimensions of [missing information]. Each row corresponds to one sample. Preliminary predicted values ​​for several material performance indicators. The number of material performance indicators, with values ​​ranging from [value 1] to [value 2]. ; The weight matrix representing the material property branch has dimensions of . , are trainable parameters used to linearly map shared features to a material property prediction space. The bias vector representing the material property branch has dimensions of . , are trainable parameters; This represents a smoothing activation function that ensures the predicted values ​​of material performance indicators (such as impact force, energy absorption capacity, and deformation) are always positive. This represents element-wise multiplication. The dimension is A vector of all 1s is added to the modulation factor as a baseline value. This ensures that when the physical modulation is 0, the modulation term is 1, which does not affect the initial prediction. When the physical modulation is non-zero, it allows for a relative adjustment of the predicted value. This represents the hyperbolic tangent activation function, which maps the input to... interval; This represents the physical heuristic combined feature matrix, with dimension 1. Each row corresponds to a physical heuristic combined feature vector of a sample, including material strength ratio, cross-sectional stiffness index, carbon emission factor, collision energy factor and angle influence coefficient, which characterizes the mechanical and carbon emission characteristics of the guardrail from different perspectives. Through a trainable matrix, it is mapped to a modulation factor that matches the output dimension, so that the model can adaptively use this physical information to correct the prediction results. Represents the physical feature modulation matrix, with dimension . , are trainable parameters used to map physical features to modulation factors of the same dimension as the output.

[0033] In practice, the preliminary prediction branch of material properties uses a hidden layer to represent the matrix. As input, a preliminary prediction is first obtained through a linear layer, followed by a smoothing activation function to ensure a positive output. Simultaneously, a physics-based heuristic is used to combine the feature matrices. Through trainable physical feature modulation matrix The mapping is used as a modulation factor, which is then activated by the hyperbolic tangent activation function and added to an all-1 vector. Finally, it is multiplied element-wise with the preliminary prediction to obtain the preliminary prediction matrix of material properties. .

[0034] 3) Preliminary carbon emission forecasting branch The hidden layer representation matrix is ​​input into the preliminary carbon emission prediction branch, which outputs several key indicators (carbon emission per unit length, total carbon emission). This branch directly utilizes the carbon emission factor in the physical features for enhancement, making the predicted values ​​closely related to the amount of material used, as shown below: In the formula, This represents the preliminary carbon emission prediction matrix, with dimensions of [missing information]. Each row corresponds to one sample. Preliminary forecasts for each carbon emission indicator, The number of carbon emission credits, with a value of [value to be filled in]. ; The weight matrix representing the carbon emission branch has dimensions of . These are trainable parameters that map shared hidden layer features to the carbon emission prediction space. The bias vector representing the carbon emission branch has dimensions of . , are trainable parameters; The balancing contribution coefficient is a trainable scalar parameter used to balance the contributions of the underlying predictions and the physics-driven corrections. The carbon emission factor is a scalar value for each sample, which is replicated and expanded in the calculation. Column vectors, via broadcast mechanism Element-wise operations are performed on the dimensional vector to provide a physical reference for each output dimension; This represents element-wise multiplication. This represents the Sigmoid activation function, which maps the input to... interval; The weight matrix representing the physics-driven correction term has dimensions of . , are trainable parameters used to generate gating values ​​for each sample in each output dimension from the hidden layer, controlling the contribution strength of the physical factors. The bias vector representing the physics-driven correction term has dimensions of . , are trainable parameters.

[0035] In practice, the preliminary carbon emission prediction branch consists of two parts: a basic prediction part and a physical correction part. The basic prediction part obtains a non-negative basic carbon emission prediction value through the activation function of the linear rectifier unit after the linear layer, while the physical correction part utilizes the carbon emission factor. The scalar of each sample is used as a physical prior. A correction term is generated through a gating mechanism. Finally, the two parts are fused to obtain a preliminary carbon emission prediction matrix. This approach retains the flexibility of data-driven approaches while providing strong prior constraints through physical factors.

[0036] 4) Collaborative Integration Layer High-strength materials often lead to higher carbon emissions. To model the intrinsic coupling relationship between material properties and carbon emissions, a trainable co-adjustment matrix and bias vector are used to co-adjust the initial predictions of the two branches, resulting in the final material property prediction, expressed as: In the formula, This represents the final prediction matrix for material properties, with dimensions of [missing information]. Each row corresponds to one sample. The final predicted values ​​of each material performance index (such as peak impact force, energy absorption capacity, and deformation); This represents the final carbon emission prediction matrix, with dimensions of [missing information]. Each row corresponds to one sample. The final predicted values ​​of individual carbon emission indicators (such as carbon emissions per unit length, total carbon emissions); Indicates will and The matrix obtained by concatenating columns has dimensions of . Each row corresponds to a preliminary predicted concatenation vector for a sample; This represents element-wise multiplication. The dimension is A vector of all ones, used to add to the modulation factor. Represents the collaboration matrix, with dimension . These are trainable parameters used to capture the linear interaction between the outputs of different tasks. Represents the collaborative bias vector, with dimension . , are trainable parameters.

[0037] In practice, the collaborative fusion layer uses a trainable collaborative matrix. and cooperative bias vector To model the intrinsic coupling relationship between material properties and carbon emissions, specifically, the preliminary prediction matrices of the two branches... and The dimension is obtained by concatenating columns. The matrix is ​​then linearly transformed and mapped to the hyperbolic tangent activation function. The interval is used to obtain the modulation factor, and this modulation factor is then compared with the all-1 vector. Adding them together gives The scaling factor of the range is then multiplied element-wise with the initial prediction after splicing to obtain the final prediction results of material properties and carbon emissions. This allows the predictions of the two tasks to influence each other and adjust in a coordinated manner, which is more in line with actual physical laws.

[0038] S4. Training of the Collaborative Prediction Model for Material Properties and Carbon Emissions After completing the construction of the material properties and carbon emission synergistic prediction model, the loss function is first constructed, and then the model is trained using the labeled training dataset so that the model can learn the mapping relationship between input features and output labels, while satisfying the physical consistency constraint.

[0039] To drive the model to accurately predict both material properties and carbon emissions, and to ensure that the prediction results conform to physical laws, the loss function consists of two parts: multi-task uncertainty weighted loss and physical consistency loss. During training, the loss function dynamically balances the weights of each task and forces the predicted values ​​to approximate the theoretical values ​​derived from physical features, thereby improving the model's generalization ability and interpretability.

[0040] 1) Multi-task uncertainty-weighted loss Drawing on the weighting strategy based on homoscedastic uncertainty in multi-task learning, a learnable noise parameter is introduced to dynamically balance the contributions of material performance prediction and carbon emission prediction, avoiding manual parameter tuning. This is expressed as: In the formula, This represents the multi-task uncertainty-weighted loss, a scalar. By introducing a learnable noise parameter, it dynamically balances the loss contributions of the two tasks, material performance prediction and carbon emission prediction, avoiding manual weight adjustment. Indicates the first The final predicted vector of material properties for each sample has dimensions of . It is the final prediction matrix of material properties. The i-th row vector; Indicates the first The material property true label vector for each sample has a dimension of . , Indicates the first The final predicted carbon emission vector for each sample has dimensions of . It is the final carbon emission prediction matrix. The i-th row vector, Indicates the first The true carbon emission label vector for each sample, with dimension [missing information]. , This represents the L2 norm, which is equivalent to the calculation method of Euclidean distance; The first noise parameter is a trainable positive scalar (constrained to be positive), representing the uncertainty (i.e., noise standard deviation) of the material property prediction task. It serves as a trainable parameter in the loss function, and its weights are automatically adjusted through optimization. The second noise parameter is a trainable positive scalar (constrained positive value) that characterizes the uncertainty of the carbon emission prediction task (i.e., noise standard deviation). It is used as a trainable parameter in the loss function and the weights of each task are automatically adjusted through optimization.

[0041] 2) Loss of physical consistency To ensure that the prediction results conform to the fundamental physical relationship between mechanics of materials and carbon emission accounting, constraint terms are constructed using physical heuristic combined features. These physical heuristic combined features refer to five features constructed based on the physical relationship between mechanics of materials and carbon emission accounting, including the material strength ratio. Section stiffness index Carbon emission factors Collision energy factor and angle influence coefficient They depict the mechanical and carbon emission characteristics of the guardrail from different perspectives, embodying physical laws.

[0042] Specifically, the predicted energy absorption capacity of the material properties should approximate the collision energy factor, and the predicted carbon emissions should approximate the theoretical carbon emission factor. A learnable scaling factor should be introduced to accommodate deviations in the actual data, expressed as: In the formula, This represents the loss of physical consistency, which forces the model's predicted values ​​to meet basic physical relationships. By penalizing the deviation between the predicted energy absorption capacity and the collision energy factor, and between the predicted carbon emissions per unit length and the carbon emission factor, the model incorporates physical priors on the basis of data-driven analysis, thereby improving the physical rationality and generalization ability of the prediction results. This represents the balance coefficient, used to adjust the relative importance of two physical constraint terms, and has a value of 0.4. Indicates the first The energy absorption capacity predicted for each sample is a component of the final predicted vector of material properties. This represents the first scaling parameter, a scalar used to adjust the ratio between energy absorption capacity and collision energy factor. Indicates the first Collision energy factor of each sample; Indicates the first The carbon emissions per unit length predicted for each sample are a component of the final carbon emission prediction vector. This represents the second scaling parameter, a scalar used to adjust the ratio between carbon emissions per unit length and the carbon emission factor. Indicates the first Carbon emission factors of each sample.

[0043] At the start of training, all trainable parameters are initialized. The embedding matrix is ​​randomly initialized, and the weight matrix is ​​initialized using Xavier. During training, a batch of samples is randomly selected from the training dataset each time, with the batch size set appropriately based on computational resources. For each batch of data, forward propagation is first performed, processing the input features in the order of the model's modules. Finally, the data passes through a collaborative fusion layer to obtain the final material property prediction matrix and carbon emission prediction matrix.

[0044] After forward propagation is completed, the loss function value is calculated. After obtaining the loss value, the backpropagation algorithm is executed to calculate the gradient of the loss function with respect to each trainable parameter. Then, the parameter values ​​are updated using the mini-batch stochastic gradient descent algorithm.

[0045] The termination of model training iterations needs to be determined based on preset judgment conditions. This invention adopts a comprehensive judgment strategy, simultaneously monitoring multiple indicators to decide whether to terminate training. First, during the training process, the model performance is periodically evaluated on a validation dataset, which is an independent subset of the original training data and does not participate in parameter updates. When the loss function value on the validation set no longer decreases or begins to increase for several consecutive training epochs, it indicates that the model may be overfitting, and training is stopped at this point. Simultaneously, the loss function value on the training set is monitored. When the loss value decreases to a point where it tends to stabilize and the change amplitude is less than a preset threshold, it indicates that the model has basically converged. In addition, a maximum number of training epochs is set as a hard limit to prevent training from continuing indefinitely. Combining these conditions, when the validation set loss does not improve for ten consecutive epochs, or the training loss tends to stabilize, or the maximum number of training epochs is reached, the model training process is terminated, and the model parameters at the point where the validation set performance is optimal are saved as the final trained material performance and carbon emission co-prediction model.

[0046] In one embodiment, the accuracy of this method in predicting material properties is compared with that of three conventional techniques. Experimental data were obtained from historical guardrail collision test reports and finite element simulation results, totaling 150 valid samples. Each sample contained eight input features, including yield strength and elastic modulus, as well as the true values ​​of three material property indicators: peak impact force, energy absorption capacity, and deformation. The comparison methods included the empirical formula method, random forest, and backpropagation (BP) neural network. The empirical formula method, based on classical material mechanics theory, directly uses linear regression to fit the relationship between input and output without introducing any data-driven optimization. The random forest used default parameters, with 100 trees and a maximum depth of 10. The BP neural network contained two hidden layers with 64 and 32 neurons per layer, respectively, and used root mean square error as the loss function. The experimental results were evaluated using root mean square error as the evaluation metric. Figure 2 The results clearly show that, in terms of peak impact force, the error value of the method described in this invention is 2.1 kN, the error value of the empirical formula method is 29 kN, the error value of the random forest method is 4 kN, and the error value of the BP neural network method is 6 kN; in terms of energy absorption capacity, the error value of the method described in this invention is 3.2 kJ, the error value of the empirical formula method is 8.1 kJ, the error value of the random forest method is 5.5 kJ, and the error value of the BP neural network method is 5.6 kJ; in terms of deformation, the error value of the method described in this invention is 1 mm, the error value of the empirical formula method is 15.2 mm, the error value of the random forest method is 2.1 mm, and the error value of the BP neural network method is 4.2 mm. From these three indicators, the method described in this invention has the lowest error, indicating that this method can more accurately capture the nonlinear laws of material properties and has high prediction accuracy.

[0047] S5. Synergistic Prediction of Material Properties and Carbon Emissions in New Scheme For a new guardrail design scheme, the first step is to collect its input features in the same way as the original data, and then organize the collected data into an input vector with a consistent format as the original dataset. The prepared original dataset is input into the training dataset building module to obtain the enhanced feature matrix; The enhanced feature matrix is ​​input into the trained model to obtain the final predicted values ​​of material properties and carbon emissions; These predicted values ​​can be directly used for performance evaluation and carbon emission calculation of guardrail design schemes. Designers can determine whether the guardrail can withstand the impact load under design conditions based on the predicted peak impact force, assess the energy dissipation performance of the guardrail based on its energy absorption capacity, and check whether it meets the deformation limit requirements based on the deformation amount. At the same time, the environmental impact of the design scheme can be assessed based on the carbon emission per unit length and the total carbon emission, providing a quantitative basis for low-carbon material selection. Through this collaborative prediction method, both the mechanical properties of materials and the carbon emission level can be considered at the design stage, achieving synergistic optimization of guardrail structure performance and environmental protection.

[0048] In one embodiment, the influence of yield strength variation on peak impact force is analyzed, and the degree of agreement between the predicted curves of different methods and the actual test points is compared to verify the sensitivity capture capability of this technology for key material parameters and the accuracy of nonlinear mapping. Figure 3 The horizontal axis represents yield strength in megapascals (MPa), and the vertical axis represents peak impact force in kilonewtons (kN). The black scatter points represent baseline values ​​obtained from real-world collision tests or high-precision simulations; these points exhibit some fluctuation, reflecting the noise characteristics of actual data. The prediction curve of this technique (solid green line) closely follows the distribution of real points and retains local fluctuations in the data without excessive smoothing; the finite element simulation curve (dashed blue line) shows a similar overall trend but deviates somewhat in the peak region; the single-task learning curve (dotted orange line) fluctuates significantly and deviates from the true value in some intervals; the empirical formula curve (dotted red line) exhibits a simple linear trend and cannot capture the complex nonlinear relationship between yield strength and impact force. Comparative results show that by integrating physically heuristic features and adaptive multi-kernel transformation, the method of this invention can more flexibly learn the local sensitivity of features in different value ranges and maintain high prediction fidelity when parameters change.

[0049] Example 2 A system for synergistic prediction of material properties and carbon emissions for guardrail design, specifically comprising: Data acquisition module: used to collect information related to guardrails and organize the guardrail information into a unified standardized format to obtain the raw dataset; Feature enhancement module: By introducing entity embedding, physical relationships in materials mechanics and carbon emission accounting, and adaptive multi-kernel transformation, the original dataset is converted into an enhanced feature matrix; Model building module: used to build a collaborative prediction model of material properties and carbon emissions for guardrail design that incorporates physically heuristic combined features and carbon emission factors; Model training module: Iteratively updates the updatable parameters through optimization algorithms to obtain the optimal parameters of the prediction model; Solution Evaluation Module: Evaluate the performance of the guardrail design in the new solution based on the predicted results, and whether the carbon emissions meet the standards.

[0050] Example 3 A computer-readable storage medium storing a computer program that executes a method for co-predicting material properties and carbon emissions for guardrail design, the computer program having the following characteristics: 1. The program code adopts a modular design, corresponding to the five core steps of data acquisition, feature enhancement, model building, model training, and scheme evaluation. The code of each module is independent and callable. 2. The program code predefines the optimal initial values ​​for all hyperparameters, including the embedding vector dimension, the number of kernel functions, the batch size, the initial learning rate, and the loss balance coefficient. These do not require manual modification and can be directly called by the processor during execution. 3. The program code includes logic for determining model convergence, which can automatically monitor the training process and realize parameter iteration and optimal model saving; 4. The program code supports CPU / GPU hybrid acceleration. Time-consuming steps such as matrix operations and feature mapping are accelerated by the GPU, while data processing and evaluation report generation are executed by the CPU, improving running efficiency.

[0051] The computer-readable storage medium can be a solid-state drive, a hard disk drive, a USB flash drive, an optical disc, etc. The program code can be written based on the Python or PyTorch / TensorFlow deep learning framework and is compatible with mainstream operating systems (Windows / Linux).

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for synergistic prediction of material properties and carbon emissions for guardrail design, characterized in that, Includes the following steps: S1. Data Collection: Collect sample data related to guardrails, including categorical features and numerical features. Organize the collected data according to a unified format to form the original dataset. S2. Feature Enhancement: Convert categorical features into model embedding vectors, construct combined features by combining the physical laws of materials mechanics and carbon emission accounting, fuse the combined features, numerical features and vectorized categorical features to obtain enhanced feature vectors, perform nonlinear mapping on the enhanced features through adaptive multi-kernel transformation, and construct the enhanced feature matrix as the training dataset. S3. Model Construction: A shared feature extraction layer is constructed based on gated linear units. The hidden layer representation matrix is ​​obtained through residual connection and layer normalization. The hidden layer representation matrix is ​​input into the preliminary prediction branch of material properties and the preliminary prediction branch of carbon emissions, respectively, to obtain the preliminary prediction results. The preliminary predictions of the two branches are adjusted collaboratively using a trainable collaborative matrix and bias vector to obtain the collaborative prediction results. S4. Model Training: Construct a loss function that includes two parts: multi-task uncertainty weighted loss and physical consistency loss. Input the training dataset into the constructed model, iteratively update all trainable parameters in the model through optimization algorithms, gradually reduce the loss function value until the model converges, and save the converged model parameters. S5. Scheme Evaluation: Collect data on the new guardrail design scheme. After the data undergoes the feature enhancement steps described above, a feature enhancement matrix is ​​obtained. The feature enhancement matrix is ​​then input into the trained prediction model to obtain material performance prediction results and carbon emission prediction results. The performance and carbon emission accounting are evaluated using the prediction results of the new guardrail design scheme.

2. The method for synergistic prediction of material properties and carbon emissions for guardrail design according to claim 1, characterized in that, The standardization process in step S1 includes: Data was collected covering guardrail model information, material mechanical parameters, geometric dimension parameters, collision condition parameters, and carbon emission-related data. During processing, categorical features were assigned unique integer indices, while numerical features directly used their original values. Each sample was labeled with two categories of tags: material performance indicators and carbon emission indicators, resulting in the original dataset.

3. The method for synergistic prediction of material properties and carbon emissions for guardrail design according to claim 1, characterized in that, The entity embedding of categorical features in step S2 specifically includes: Assign a unique integer index to each model and construct an embedding matrix where each row corresponds to an initial embedding vector for a model. These vectors are randomly initialized at the start of training. One sample, its model index This is used to query the corresponding row from the embedding matrix, thus obtaining the model embedding vector of the sample. .

4. The method for synergistic prediction of material properties and carbon emissions for guardrail design according to claim 1, characterized in that, Step 2, the adaptive multi-kernel transformation, specifically includes: For each feature dimension, a kernel function is used to model local sensitivity. The center and width of the kernel function are determined by data, adaptively learning the local sensitivity of the feature in different value ranges, specifically expressed as follows: In the formula, Indicates the first The feature is processed by the first The mapping values ​​of each kernel function, Indicates feature index, Indicates the kernel function index. Indicates the first The first feature The weight of each core, Represents the Gaussian radial basis function, which maps the input to... interval, Represents enhanced feature vectors The One portion, Indicates the first The first feature The center of each kernel function, Indicates the first The first feature The width of each kernel function; All kernel function mapping values ​​obtained by adaptive multi-kernel transformation for each sample are concatenated into a vector. The vectors of all samples are stacked row by row to obtain the enhanced feature matrix.

5. The method for synergistic prediction of material properties and carbon emissions for guardrail design according to claim 1, characterized in that, The model construction in step S3 specifically includes: S31. Construct a shared feature extraction layer A shared feature extraction layer based on gated linear units is constructed. This layer enhances training stability through residual connections and layer normalization, thereby obtaining the hidden layer representation, as follows: In the formula, The hidden layer representation matrix is ​​represented. Represents the enhanced feature matrix, This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer. This represents the linear transformation weight matrix in the residual connection. Presentation layer normalization operation, This represents the activation function of the linear rectifier unit; S32. Preliminary Prediction of Material Properties (Branch) The shared feature matrix is ​​mapped to the material property prediction space through a linear layer. The output is guaranteed to be positive by a smooth activation function. At the same time, the physical heuristic combined feature matrix is ​​mapped to a modulation factor of 1 through a trainable physical feature modulation matrix. After activation by a hyperbolic tangent activation function, it is added to an all-1 vector. Finally, it is multiplied element-wise with the preliminary prediction to obtain the preliminary prediction matrix of material properties. S33. Preliminary Carbon Emission Forecasting Branch The shared feature matrix is ​​activated by the linear rectifier unit after the linear layer to obtain a non-negative basic carbon emission prediction value. The physical correction part uses the carbon emission factor as a physical prior and generates correction terms through a gating mechanism. Finally, the two parts are fused to obtain the preliminary carbon emission prediction matrix. S34. Collaborative Integration Layer The initial prediction matrices of the two branches are concatenated column-wise to obtain a concatenated matrix, which is then subjected to a linear transformation and mapped to a hyperbolic tangent activation function. The interval is used to obtain the modulation factor 2. This modulation factor 2 is then compared with the all-1 vector. Adding them together gives The scaling factor of the range is then multiplied element-wise with the concatenation matrix to obtain the co-prediction result.

6. The method for synergistic prediction of material properties and carbon emissions for guardrail design according to claim 1, characterized in that, Step S4, model training, specifically includes: S41. Loss Function: Construct a loss function that includes two parts: multi-task uncertainty weighted loss and physical consistency loss, to obtain the difference between the predicted value and the true value; S42. Forward Propagation: Input the enhanced feature matrix, obtain the hidden layer representation matrix through the shared feature extraction layer; then obtain the preliminary prediction matrix through the preliminary prediction branch of material properties and the preliminary prediction branch of carbon emissions respectively; finally, obtain the collaborative prediction result through the collaborative fusion layer. S43. Backpropagation: After completing the forward propagation, calculate the loss function value, then execute the backpropagation algorithm to calculate the gradient of the loss function with respect to each trainable parameter, and use the optimization algorithm to update the parameter values.

7. The method for synergistic prediction of material properties and carbon emissions for guardrail design according to claim 1, characterized in that, Step S5, the scheme evaluation specifically includes: S51. Collect the input features of the new design scheme and construct an enhanced feature matrix; S52. Input the enhanced feature matrix into the trained model to obtain the collaborative prediction results; S53. Based on the prediction results, determine whether the guardrail can withstand the impact load under the design conditions, whether it meets the deformation limit requirements, and assess the environmental impact of the design scheme.

8. A system for co-predicting material properties and carbon emissions for guardrail design, comprising executing the method for co-predicting material properties and carbon emissions for guardrail design as described in any one of claims 1-7, characterized in that, include: Data acquisition module: used to collect information related to guardrails and organize the guardrail information into a unified standardized format to obtain the raw dataset; Feature enhancement module: By introducing entity embedding, physical relationships in materials mechanics and carbon emission accounting, and adaptive multi-kernel transformation, the original dataset is converted into an enhanced feature matrix; Model building module: used to build a collaborative prediction model of material properties and carbon emissions for guardrail design that incorporates physically heuristic combined features and carbon emission factors; Model training module: Iteratively updates the updatable parameters through optimization algorithms to obtain the optimal parameters of the prediction model; Solution Evaluation Module: Evaluate the performance of the guardrail design in the new solution based on the predicted results, and whether the carbon emissions meet the standards.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the method for co-predicting material properties and carbon emissions for guardrail design as described in any one of claims 1-7.

Images