Cement-based material multi-scale intelligent design method based on machine learning

Abstract

The invention relates to the technical field of cement-based material design, and discloses a machine learning-based cement-based material multi-scale intelligent design method, which comprises the following steps of: constructing a system element database of a cement-based material under nano-scale, micro-scale, meso-scale and macro-scale; based on a machine learning algorithm, training the system element database to construct a machine learning model so as to establish mapping and association among different scales; designing parameters under any scale are input into the machine learning model, forward prediction is executed, and performance and structural features of the target object under the scale are output and obtained; and inputting the performance or structural characteristics of the target object under any scale into the machine learning model, executing reverse derivation, and outputting design elements required for realizing the performance or the structure. The method has the characteristics of intelligence, automation and high efficiency, the labor cost is remarkably reduced, the limitation that a traditional fitting method is long in period and low in efficiency is overcome, and the technical blank of multi-scale information transmission and performance mapping is filled.

Description

Technical Field

[0001] This invention relates to the field of cement-based material design technology, and in particular to a multi-scale intelligent design method for cement-based materials based on machine learning. Background Technology

[0002] As the most widely used building materials, the performance of cement-based materials directly determines the safety, durability, and long-term service capability of engineering structures.

[0003] Currently, the design of this material still heavily relies on engineering experience and the traditional "trial mixing method." This method requires a large number of repetitive experiments and has inherent drawbacks such as long R&D cycles, high economic costs, and high resource consumption. Furthermore, it is difficult to systematically reveal the intrinsic relationship between material composition, structure, and performance, which has become a bottleneck restricting the development of high-performance cement-based materials.

[0004] In recent years, artificial intelligence technology has brought new opportunities to materials design. Data-driven algorithms such as machine learning and deep learning have been used to predict and optimize material properties, showing great potential. However, most existing studies are limited to modeling and analysis at a single scale (such as only macroscopic or only microscopic), failing to effectively connect the complete scale chain from nanoscale, microscale, mesoscale to macroscale, resulting in a significant gap in the research on information transmission and performance coupling mechanisms between different scales.

[0005] Therefore, there is an urgent need for a multi-scale intelligent design method for cement-based materials that integrates multi-scale databases and intelligent algorithms to achieve collaborative correlation and performance prediction between different scales, and promote the intelligent, cross-scale and efficient design of cement-based materials. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-scale intelligent design method for cement-based materials based on machine learning. By utilizing high-throughput data processing, it achieves cross-scale, intelligent, and efficient design with extremely low labor costs, greatly compensating for the shortcomings of traditional trial mixing methods and filling the gap in multi-scale information transmission.

[0007] To achieve the above objectives, the following technical solution is adopted: A multi-scale intelligent design method for cement-based materials based on machine learning includes the following steps: Construct a database of system elements for cement-based materials at the nano, micro, meso, and macro scales; Based on machine learning algorithms, the system element database is trained to construct a machine learning model, thereby establishing mappings and associations between different scales; Input design parameters at any scale into the machine learning model, perform positive prediction, and output the performance and structural characteristics of the target object at that scale; Input the performance or structural characteristics of the target object at any scale into the machine learning model, perform reverse derivation, and output the design elements required to achieve the performance or structure.

[0008] Furthermore, the scale ranges of the nanoscale, microscale, mesoscale, and macroscale are respectively: nanoscale 1~100nm, microscale 0.1~100μm, mesoscale 0.1~100mm, and macroscale >0.1m; The system elements include target objects at various scales, environmental conditions, and materials that interact with the target objects; The target objects include: Nanoscale: Hydration products represented by calcium silicate hydrate gel (CSH); Microscale: heterogeneous multiphase cement paste composed of cement particles, hydration products, water, water-reducing agents, and pores; Detailed scale: Concrete is composed of cement paste, fine aggregate, coarse aggregate, auxiliary cementitious materials and other admixtures; Macro scale: Cast-in-place samples or components, primarily made of concrete.

[0009] Furthermore, the data used to construct the system element database originates from at least one of experimental testing and characterization, literature retrieval, theoretical calculation, or numerical simulation; the data types include numerical data and image data.

[0010] Furthermore, the system element database includes element data at any scale, and the element data at any scale includes design parameters at the corresponding scale and the performance and structural characteristics of the target object at the corresponding scale.

[0011] Furthermore, the design parameters include the phase composition, content, size, shape, spatial distribution, intrinsic properties of each phase, interactions between phases, intrinsic properties of related materials, interactions between the target object and related materials, and environmental conditions of the target object. The performance and structural characteristics include: the physical and chemical properties and structural information exhibited by the target object at a certain scale as a whole under set environmental conditions.

[0012] Furthermore, the performance and structural characteristics of the target object at a lower scale constitute part of the design parameters of the new target object at a higher scale.

[0013] Furthermore, the machine learning algorithm is any one of random forest, support vector machine, gradient boosting decision tree, or artificial neural network.

[0014] Furthermore, the mapping and association model includes one or more core models; in the core model, the design parameters at any scale are used as independent variables, and the performance and structural characteristics of the target object at the same scale are used as dependent variables.

[0015] Furthermore, the mapping and association model also includes one or more auxiliary models; in the auxiliary models, the image data in the design parameters at any scale is used as the independent variable, and the corresponding numerical data is used as the dependent variable.

[0016] Furthermore, when performing forward prediction, the input design parameters are within the critical range preset by the model; the output performance and structural characteristics are one or more preset features, or all features at the corresponding scale.

[0017] Furthermore, during the reverse derivation, the input performance or structural features are within the critical range preset by the model, and are one or more of the set features, or all features at the corresponding scale; the output design elements are the design parameters necessary to achieve the input.

[0018] The beneficial effects of this invention are reflected in: (1) This invention connects cross-scale element data of cement-based materials through high-throughput machine learning. After inputting relevant parameters of any scale, it can quickly output the corresponding property characteristics or design elements at that scale, realizing intelligent and efficient material design. Moreover, the output time is fast and the labor cost is low.

[0019] (2) The hyperparameters of the machine learning model structure obtained in this invention reflect the information transmission characteristics and weights between various elements. The model can be visualized and analyzed in a targeted manner to analyze the importance of different elements and provide theoretical guidance for further research on cross-scale performance transmission of cement-based materials.

[0020] (3) The database established in this invention has a wide range and a large scope, and is universal. It can be used to conduct relevant experiments based on different materials and research contents, and provide a design platform for the development of new ultra-high performance cement-based materials. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a multi-scale intelligent design method for cement-based materials based on machine learning, provided as an embodiment of the present invention; Figure 2 Example diagram of the system element database of cement-based materials at various scales provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the input / output data of the machine learning model provided in an embodiment of the present invention; Figure 4 This is a technical concept diagram of a multi-scale intelligent design method based on machine learning provided in an embodiment of the present invention. Detailed Implementation

[0022] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0023] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0024] This invention provides a multi-scale intelligent design method for cement-based materials based on machine learning, such as... Figure 1 The diagram shows a flowchart of the machine learning-based multi-scale intelligent design method for cement-based materials. This machine learning-based multi-scale intelligent design method for cement-based materials specifically includes the following steps S10-S40.

[0025] S10: Construct a database of system elements for cement-based materials at the nano, micro, meso, and macro scales.

[0026] Step S10 forms the data foundation of the entire method. It involves systematically collecting, filtering, and structuring various elemental data of cement-based materials at different scales (nanoscale, microscale, mesoscale, and macroscale). For each scale, the target object is clearly defined, such as nanoscale hydration products, microscale cement paste, mesoscale concrete, and macroscale cast-in-place components, and all data related to that object are compiled.

[0027] By implementing step S10, a standardized database covering the entire scale chain was constructed. This database not only contains traditional formulation and macroscopic performance data, but also deeply integrates structural and performance information at various scales, providing high-quality, high-throughput data raw materials for subsequent machine learning training, overcoming the shortcomings of traditional methods such as single data and fragmented scales.

[0028] In some embodiments, the scale ranges of the nanoscale, microscale, mesoscale, and macroscale are respectively: nanoscale 1~100nm, microscale 0.1~100μm, mesoscale 0.1~100mm, and macroscale >0.1m.

[0029] Specifically, for the nanoscale (1–100 nm), this range primarily defines the basic structural units of atoms, molecular clusters, and primary hydration products in cement-based materials. The core research objects are the core products of cement hydration: colloidal particles of calcium silicate hydrate (CSH) gel, crystal nuclei of calcium hydroxide crystals, and ionic clusters in the pore solution. At this scale, the focus is on the chemical bonding, intermolecular forces, and nanoscale pore structure of the material, which are the physicochemical basis determining the material's microscopic mechanical properties and durability.

[0030] For the microscale (0.1–100 μm), this range primarily defines multiphase composite systems composed of nanoscale units. The core research object is the microstructure of cement paste containing heterogeneous components such as unhydrated cement particles, various hydration products, capillaries, gel pores, and microcracks. At this scale, the focus is on the composition, distribution, morphology, and interfacial structures and interactions of each phase (CSH gel phase, CH crystalline phase, and porous phase). The structural characteristics at this scale directly determine the overall mechanical behavior and transport properties of the cement paste.

[0031] For the mesoscopic scale of 0.1–100 mm, this scale primarily defines heterogeneous materials composed of microscopic cement paste as the matrix phase, combined with reinforcing phases such as aggregates and fibers. The core research object is concrete, whose structural elements include cement mortar, coarse / fine aggregates, interfacial transition zones, and macroscopic defects (such as cracks and air bubbles). At this scale, the focus is on the spatial distribution and interactions of the components (especially the particle size, gradation, and shape of aggregates, as well as the properties of the interfacial transition zones). The structure at this scale is crucial in determining the macroscopic mechanical properties and failure modes of concrete.

[0032] For macroscopic scales > 0.1 m, this scale primarily defines specimens or components that can be tested and applied as engineering structural units. The core research objects are standard-cured concrete test blocks in the laboratory or cast-in-place components in actual engineering projects. At this scale, the focus is on the overall, average engineering properties exhibited by the material, such as compressive / flexural strength, modulus of elasticity, permeability, shrinkage, and creep; these properties are the direct basis for engineering design.

[0033] In some embodiments, the system elements include target objects at various scales, environmental conditions, and materials that interact with the target objects.

[0034] It should be noted that the target object, also known as the research object, specifically refers to the main material or representative volume element studied and analyzed at a specific scale. For example, at the nanoscale, the target object is hydration products such as calcium silicate hydrate (CSH) gel; at the microscale, the target object is cement paste containing cement particles, hydration products, water, and pores; at the mesoscale, the target object is concrete composed of cement paste, aggregates, and interfacial transition zones; and at the macroscale, the target object is concrete specimens or components.

[0035] Environmental conditions refer to the external physical and chemical fields in which the target object is located and is affected. These conditions may change over time and directly affect the performance evolution of the target object. Environmental conditions include, but are not limited to: temperature (e.g., curing temperature, ambient temperature), humidity (e.g., relative humidity, water curing conditions), external forces (e.g., loading method, stress level), and chemical environment (e.g., pH value of the solution, concentration of corrosive ions such as Cl). - SO4 2- (etc.) and time (such as curing age, loading duration). For example, the carbonation depth of concrete on a macroscopic scale depends on the CO2 concentration and humidity in the environment.

[0036] Interacting materials refer to other material phases or components that, at a specific scale, directly contact the target object and undergo physical, chemical, or mechanical interactions. These interactions are the direct cause of the evolution of material properties. Examples include, at the nanoscale, the interaction between water molecules and the surface of CSH gel, and the adsorption of water-reducing agent molecules on the surface of cement particles. At the microscale, the hydration reaction between unhydrated cement particles and water, and the interfacial bonding between fibers and the cement paste matrix. At the mesoscale, the interfacial transition zone formed between cement paste and coarse / fine aggregates, and the bridging and crack-resistant effects of fibers or polymer particles in concrete. At the macroscale, the bond-slip behavior between concrete and embedded reinforcing steel, and the compatibility between concrete and external protective coatings.

[0037] In some embodiments, the target object includes: nanoscale: hydration products represented by calcium silicate hydrate gel (CSH); microscale: heterogeneous multiphase cement paste composed of cement particles, hydration products, water, water-reducing agent, and pores; mesoscale: concrete composed of cement paste, fine aggregate, coarse aggregate, auxiliary cementitious materials, and other admixtures; macroscale: cast samples or components mainly composed of concrete.

[0038] In some embodiments, the data used to construct the system element database originates from at least one of experimental testing and characterization, literature retrieval, theoretical calculation, or numerical simulation; the data types include numerical data and image data.

[0039] In some embodiments, the system element database includes element data at any scale, which includes design parameters at the corresponding scale and the performance and structural characteristics of the target object at the corresponding scale. The design parameters include the phase composition, content, size, shape, spatial distribution, intrinsic properties of each phase, interactions between phases, intrinsic properties of related materials, interactions between the target object and related materials, and environmental conditions of the target object. The performance and structural characteristics include the physical, chemical properties and structural information exhibited by the target object as a whole at a certain scale under specified environmental conditions.

[0040] In some embodiments, the specific content and hierarchical division of the system element database can be referred to... Figure 2 The example shown. Figure 2 It clearly demonstrates the core elements contained in the database at four scales: nanoscale, microscale, granular scale, and macroscale.

[0041] For the nanoscale (1-100 nm), the research object is defined as hydration products, a typical example being calcium silicate hydrate (CSH) gel. Numerical data include the types, numbers, and arrangements of particles contained in the hydration products; the size and shape of the hydration product structure; and the property characteristics of the hydration products, such as nanoscale tensile strength and Young's modulus. Image data mainly consists of molecular models reflecting the atomic / molecular arrangement structure.

[0042] For the microscale (0.1–100 μm), the research object is defined as a heterogeneous multiphase system composed of cement particles, hydration products, water, pores, and admixtures, such as cement paste. Numerical data include the composition, content, size, shape, spatial distribution, intrinsic properties, and interfacial strength of each phase of the cement paste at different times; as well as the overall properties of the cement paste, such as microhardness and permeability. Image data includes two-dimensional and three-dimensional images of the microstructure obtained by scanning electron microscopy, as well as microstructure models reconstructed based on these images.

[0043] For the microscale (0.1–100 mm), the research object is defined as a composite material composed of cement paste, aggregates, and interfacial transition zones, such as concrete. Numerical data includes the composition, content, size, shape, spatial distribution, intrinsic properties, and interfacial strength of each phase of concrete at different times; as well as the overall properties of concrete, such as fracture energy and load transfer behavior at the microscale. Image data includes two-dimensional and three-dimensional images of the microstructure obtained through X-ray tomography and other methods, as well as microstructure models built based on these images.

[0044] For macroscopic scales (>0.1 m), the research object is defined as a cast sample or component prepared in a laboratory or engineering site. Numerical data includes the environmental conditions of the sample, such as temperature, humidity, and curing time; the composition, content, size, shape, spatial distribution, intrinsic properties, and interfacial strength of each phase of the sample; and the macroscopic properties of the sample, including fresh mixture properties (e.g., flowability), mechanical properties (e.g., compressive strength), durability properties (e.g., resistance to chloride ion penetration), and long-term properties (e.g., lifespan prediction). Image data includes two-dimensional images of the sample surface and three-dimensional structural models reflecting the overall internal structure.

[0045] S20: Based on machine learning algorithms, train the system element database to build a machine learning model, thereby establishing mapping and correlation between different scales.

[0046] In this embodiment, the database constructed in step S10 is used to select a suitable machine learning algorithm for model training. This process involves the construction of two types of models: Core Model: The model is trained by taking design parameters at any scale as input variables and performance and structural characteristics at the same scale as output variables. For example, design parameters at any scale can be nanoscale CSH gel structure information or macroscale water-gel ratio and curing regime; performance and structural characteristics at the same scale can be nanoscale gel strength or macroscale compressive strength.

[0047] Auxiliary Model: To process multimodal data, a model is specifically trained to connect image data and numerical data, enabling mutual conversion and feature extraction between different data formats.

[0048] The aforementioned core model and auxiliary model together constitute a mapping and correlation model that can understand and predict complex nonlinear relationships across scales.

[0049] The trained model can accurately capture the intrinsic connections and transmission laws from nanoscale structure to macroscopic performance. The structural hyperparameters (neural network weights) within the model objectively quantify the contribution of each design parameter to the performance, providing a data-driven insight tool for revealing the cross-scale evolution mechanism of material properties.

[0050] In some embodiments, the performance and structural characteristics of the target object at a low-level scale constitute part of the design parameters of the new target object at a higher-level scale. For example... Figure 3The diagram illustrates the input / output data of a machine learning model. For low-level scales, the design parameters are first used as the basic input conditions, along with corresponding numerical and image data. These low-level design parameters, auxiliary numerical and image data are then input into a suitable machine learning model. After training or inference, the model outputs low-level scale properties and features. In the high-level scale design process, these low-level scale properties and features become part of the high-level scale design parameters, meaning they encompass the performance and structural characteristics at the low-level scale. Similarly, high-level numerical and image data are collected, and these parameters, along with the auxiliary numerical and image data, are input into the corresponding machine learning model, ultimately outputting high-level scale properties and features. This data flow logic enables the transfer of low-level scale features to high-level scale design parameters, providing data and model-level support for multi-scale intelligent design.

[0051] In some embodiments, the machine learning algorithm is any one of random forest, support vector machine, gradient boosting decision tree, or artificial neural network.

[0052] In some embodiments, the mapping and association model includes one or more core models; in the core model, design parameters at any scale are used as independent variables, and the performance and structural characteristics of the target object at the same scale are used as dependent variables.

[0053] For each specific scale, all data records for that scale are extracted from the system element database. Each record contains a complete set of design parameters and corresponding performance and structural characteristics obtained through experiments or simulations. These data are used to train a selected machine learning algorithm, and through iterative optimization, the model learns the complex, nonlinear mapping relationship from design parameters to performance characteristics. The trained model constitutes the core prediction model for that scale. When performance prediction for new materials or formulations is required, a set of known design parameters at a specific scale are input into the core model corresponding to that scale. For example, at the macroscopic scale, parameters such as water-cement ratio, cementitious material composition, and curing regime are input. The model will perform high-speed calculations based on the learned mapping relationship and output one or more predicted performance and structural characteristics, such as 28-day compressive strength and chloride ion diffusion coefficient. When there is a clear performance target, the core model can be used for back-deriving. The desired performance and structural characteristics are used as the model's target output, and an optimization algorithm searches the model's input space to find the design parameter combination that makes the model output closest to the target value. This process enables reverse solving from effect to cause, providing key formulations and process parameters for directional material design.

[0054] In some embodiments, the mapping and association model further includes one or more auxiliary models; in the auxiliary models, image data in the design parameters at any scale is used as independent variables, and the corresponding numerical data is used as dependent variables.

[0055] In the auxiliary model, image data from the design parameters at any scale is used as the independent variable, and the corresponding numerical data reflecting the same physical essence is used as the dependent variable for training. The core function of this auxiliary model is to achieve high-precision conversion and information extraction between different data modalities.

[0056] Specifically, the training data for the auxiliary model comes from the system element database. For example, at the microscopic scale, a large number of scanning electron microscope images of cement slurry can be collected as independent variables, and corresponding numerical data (such as porosity, phase content, average pore size, etc.) obtained through image analysis techniques or experimental measurements can be used as dependent variables. By training this type of image-numerical pairing data with deep learning algorithms (such as convolutional neural networks CNN), the model can learn to automatically identify key features from complex images and quantify the inherent laws of their physical and geometric parameters.

[0057] The trained auxiliary model can automatically and quickly convert newly input, unlabeled image data into reliable numerical data. For example, given a new 2D CT scan image of concrete, the model can automatically output numerical parameters such as the spatial distribution statistics of aggregates and porosity. The high-value numerical data extracted from the image by the auxiliary model can be used as more reliable and comprehensive design parameters, directly input into the corresponding core model for forward performance prediction or reverse design. This solves the problem of decreased prediction accuracy of the core model due to mismatch or incompleteness between image and numerical information in the original data.

[0058] The auxiliary model also provides a reverse verification path. The performance characteristics predicted by the core model can be used to infer the possible corresponding microstructural image features through the auxiliary model, realizing numerical-image cross-validation and deepening the understanding of the structural mechanism behind the material properties.

[0059] S30: Input design parameters at any scale into the machine learning model, perform positive prediction, and output the performance and structural characteristics of the target object at that scale.

[0060] Step S20 enables performance-oriented prediction of materials. After obtaining the mature model trained in step S20, users can input one or more sets of design parameters at any scale according to design requirements. The model will perform high-speed calculations and inferences based on the learned complex mapping relationships, and output one or more performance and structural characteristics corresponding to that scale. As an example, the input design parameters are as follows: at the macroscopic scale, the inputs are: water-cement ratio 0.3, fly ash content 20%, and specific curing conditions. The output performance and structural characteristics are: 28-day compressive strength of 85 MPa, and average porosity of 8%.

[0061] Step S30 transforms traditional time-consuming and costly physical experiments into instantaneous computer predictions, significantly improving R&D efficiency. Designers can quickly assess the potential performance of different formulations and processes before putting them into actual production, enabling virtual screening and optimization of materials.

[0062] S40: Input the performance or structural characteristics of the target object at any scale into the machine learning model, perform reverse derivation, and output the design elements required to achieve the performance or structure.

[0063] Step S40 is used to achieve customized design of material performance. Specifically, users can directly input desired performance or structural characteristics into the model constructed in step S20 according to specific engineering requirements. The model will run a backpropagation algorithm to search for a combination of design parameters that can meet or approximate the target performance from a vast parameter space. As an example, the desired performance or structural characteristics input are: 7-day compressive strength ≥ 50 MPa and chloride ion permeability coefficient ≤ 2.0 × 10⁻⁶. -12 m 2 / s; The model output includes: recommended water-cement ratio range, admixture type and dosage, curing plan, etc.

[0064] Step S40 enables the reverse solution from performance targets to design solutions, providing a precise and efficient technical approach for developing customized, ultra-high-performance cement-based materials, and significantly reducing the design threshold and R&D cycle of high-performance materials.

[0065] In some embodiments, when performing forward prediction, the input design parameters are within the critical range preset by the model; the output performance and structural features are one or more preset features, or all features at the corresponding scale.

[0066] When performing positive prediction, the user inputs a set of design parameters at any scale into the trained machine learning model. The input parameters must be within a pre-defined critical range. This critical range is determined during model training based on the distribution range (such as minimum and maximum values) of each feature value in the training dataset. Keeping the input parameters within this range ensures the model operates within a familiar knowledge domain, thus guaranteeing the reliability of the output results. The model's output represents the performance and structural characteristics of the target object at that scale. This output can be one or more features pre-defined by the user based on specific research objectives. For example, at a macroscopic scale, the user can choose to predict only the compressive strength and chloride ion diffusion coefficient—two key indicators. Simultaneously, the model also supports outputting all performance and structural characteristics defined in the database at that scale, providing the user with a comprehensive performance prediction report.

[0067] In some embodiments, when performing reverse derivation, the input performance or structural features are within the critical range preset by the model, and are one or more of the set features, or all features at the corresponding scale; the output design elements are the design parameters necessary to realize the input.

[0068] During backpropagation, the user inputs the desired performance or structural characteristics into the model. Similar to forward prediction, these input features must also be within the model's predefined critical range, and can be one or more features set by the user, or all features at that scale as a comprehensive objective. The model's output consists of the design parameters necessary to achieve the input objective. It's important to note that "necessary" here means that the model, through its internal backpropagation or optimization search algorithm, finds the key design variable combination that optimally satisfies the input performance objective from a vast array of parameter combinations. The output typically includes the core formulation and process parameters for achieving the objective.

[0069] In some embodiments, such as Figure 4The diagram illustrates the technical concept of a multi-scale intelligent design method based on machine learning. This method uses fundamental science as its underlying theoretical basis, directly determining the setting of nanoscale design parameters. These nanoscale design parameters correspond to nanoscale property characteristics. These nanoscale property characteristics serve as input elements, supporting the formation of microscale design parameters, from which microscale property characteristics are derived. Similarly, microscale property characteristics further form mesoscale design parameters, which output mesoscale property characteristics. These mesoscale property characteristics then serve as input to form macroscale design parameters, from which macroscale property characteristics are derived. Ultimately, these macroscale property characteristics can be directly applied to engineering scenarios. Simultaneously, macroscale design parameters are input to a machine learning module, which possesses both forward prediction and reverse design capabilities. Forward prediction can deduce property characteristics at the corresponding scale based on the input design parameters; reverse design can infer suitable design parameters based on preset target property characteristics, thereby achieving intelligent iteration and precise optimization of the multi-scale design process.

[0070] Based on the above technical concept and the methods provided in the above embodiments, this embodiment gives two specific examples, as shown in Example 1 and Example 2 below.

[0071] Example 1: This embodiment provides a method for obtaining the property characteristics of ultra-high performance concrete based on design parameters, including the following four steps: (1) In the established multi-scale cement-based material database, select the raw material composition (type and proportion of cementitious materials, water-cement ratio, curing conditions, and admixture dosage) corresponding to the target concrete system as input features; (2) Input the design parameters into the trained cross-scale machine learning model (a prediction model based on GBDT or ANN), and the model will automatically complete feature extraction and cross-scale mapping; (3) The model outputs the performance characteristics of the corresponding system, including but not limited to compressive strength, porosity, carbonization depth, permeability and microstructure indicators, to realize intelligent prediction from design parameters to performance characteristics; (4) Compare the prediction results with the experimental data, and correct the model weights through the error feedback mechanism to continuously improve the prediction accuracy.

[0072] Example 2: This embodiment provides a method for obtaining design parameters based on the property characteristics of high-performance concrete, including the following four steps: (1) Input target performance: Set the target performance parameters as target compressive strength, durability grade or microstructure density according to engineering or research needs; (2) Input the target performance into the trained cross-scale machine learning inversion model. The model automatically derives the combination of key design parameters required to achieve the performance based on the established multi-scale database and feature mapping relationship. (3) Output multi-dimensional design parameters including water-cement ratio, admixture ratio, curing regime, and additive dosage to provide guidance for material preparation and optimization; (4) Perform small-scale verification based on the inversion results, feed the test data back to the model, and continuously improve the accuracy of the cross-scale database and model.

[0073] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A multi-scale intelligent design method for cement-based materials based on machine learning, characterized in that, Includes the following steps: Construct a database of system elements for cement-based materials at the nano, micro, meso, and macro scales; Based on machine learning algorithms, the system element database is trained to construct a machine learning model, thereby establishing mappings and associations between different scales; Input design parameters at any scale into the machine learning model, perform positive prediction, and output the performance and structural characteristics of the target object at that scale; Input the performance or structural characteristics of the target object at any scale into the machine learning model, perform reverse derivation, and output the design elements required to achieve the performance or structure.

2. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 1, characterized in that, The scale ranges of the nanoscale, microscale, mesoscale, and macroscale are respectively: nanoscale 1~100nm, microscale 0.1~100μm, mesoscale 0.1~100mm, and macroscale >0.1m; The system elements include target objects at various scales, environmental conditions, and materials that interact with the target objects; The target objects include: Nanoscale: Hydration products represented by calcium silicate hydrate gel (CSH); Microscale: heterogeneous multiphase cement paste composed of cement particles, hydration products, water, water-reducing agents, and pores; Detailed scale: Concrete is composed of cement paste, fine aggregate, coarse aggregate, auxiliary cementitious materials and other admixtures; Macro scale: Cast-in-place samples or components, primarily made of concrete.

3. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 1, characterized in that, The data used to construct the database of system elements comes from at least one of experimental testing and characterization, literature retrieval, theoretical calculation, or numerical simulation; the data types include numerical data and image data.

4. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 1, characterized in that, The system element database includes element data at any scale, and the element data at any scale includes the design parameters at the corresponding scale and the performance and structural characteristics of the target object at the corresponding scale.

5. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 4, characterized in that, The design parameters include the phase composition, content, size, shape, spatial distribution, intrinsic properties of each phase, interactions between phases, intrinsic properties of related materials, interactions between the target object and related materials, and environmental conditions of the target object. The performance and structural characteristics include: the physical and chemical properties and structural information exhibited by the target object at a certain scale as a whole under set environmental conditions.

6. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 4, characterized in that, The performance and structural characteristics of a target object at a lower scale constitute part of the design parameters of a new target object at a higher scale.

7. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 1, characterized in that, The mapping and association model includes one or more core models; in the core model, the design parameters at any scale are used as independent variables, and the performance and structural characteristics of the target object at the same scale are used as dependent variables.

8. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 7, characterized in that, The mapping and association model also includes one or more auxiliary models; in the auxiliary models, the image data in the design parameters at any scale is used as the independent variable, and the corresponding numerical data is used as the dependent variable.

9. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 1, characterized in that, When performing forward prediction, the input design parameters are within the critical range preset by the model; the output performance and structural characteristics are one or more preset features, or all features at the corresponding scale.

10. The multi-scale intelligent design method for cement-based materials based on machine learning according to claim 1, characterized in that, When performing reverse derivation, the input performance or structural features are within the critical range preset by the model, and are one or more of the set features, or all features at the corresponding scale; the output design elements are the design parameters necessary to achieve the input.

Images