Geotechnical material data driving construction modeling method and device based on knowledge migration strategy

Abstract

The invention relates to the technical field of geotechnical material local modeling, in particular to a geotechnical material data driving local modeling method and device based on a knowledge migration strategy, and the method comprises the steps: building a first mapping data pair based on the stress-strain physical quantity of a test curve sampling point, and building a high-fidelity data set of a geotechnical material; constructing a second mapping data pair based on the stress-strain state of the simulation curve sampling point and the intermediate physical quantity label, and establishing a low-fidelity data set of the geotechnical material; pre-training the target data-driven constitutive model by using the low-fidelity data set to generate an initial geotechnical material data-driven constitutive model; and driving the model to be fine-tuned by using the high-fidelity data set for the initial geotechnical material data, and generating final geotechnical material data to drive the model. Therefore, the problem of insufficient generalization ability and prediction precision of a data-driven model due to limited training data and unpredictability of a stress path in the prior art is solved.

Description

Technical Field

[0001] This application relates to the field of geotechnical material constitutive modeling technology, and in particular to a data-driven constitutive modeling method and apparatus for geotechnical materials based on a knowledge transfer strategy. Background Technology

[0002] The mechanical response of geotechnical materials is influenced by multiple coupled factors, including particle morphology, particle size distribution, particle contact state, and particle fragmentation degree. This results in significant nonlinearity, dilatation, and anisotropy in their mechanical properties. Due to these complex mechanical properties, the development of traditional constitutive models has encountered bottlenecks: on the one hand, simple constitutive models such as the Mohr-Coulomb model and the modified Cambridge model are insufficient to accurately describe the mechanical response of materials under multi-factor coupling and multi-path loading conditions; on the other hand, while advanced constitutive models such as the boundary surface model, the Unified Hardening (UH) model, and the subplastic model can better characterize the complex mechanical properties of materials, their complex mathematical expressions and the large number of material parameters that need to be determined make their application in practical engineering difficult. Therefore, the development of constitutive theory for geotechnical materials urgently needs breakthroughs.

[0003] In related technologies, artificial intelligence is driving a profound transformation in scientific research paradigms, and data-driven methods are being widely applied to the construction of material constitutive models. The core logic of these methods is to use machine learning techniques such as artificial neural networks to capture the complex mapping relationship between stress and strain, and then construct alternative constitutive models. Once trained, the models can not only predict the mechanical behavior of materials under different loading conditions, but can also be directly embedded into the numerical solution framework of boundary value problems (BVPs), providing new pathways for engineering simulation.

[0004] However, when applied to real-world complex engineering problems, the aforementioned methods still face two major challenges: One issue is the insufficient robustness of the model. In real-world engineering, BVPs often involve a large number of unpredictable stress paths, while purely data-driven alternative models are highly dependent on the training data. Once faced with data distribution shifts (such as stress paths exceeding the training range), the model's predictive stability will significantly decrease.

[0005] Secondly, there is the challenge of small-sample modeling. While physical experiments can accurately obtain data on the mechanical properties of materials, the experimental process is time-consuming and labor-intensive, making it extremely difficult to acquire large-scale data. The limited amount of data directly restricts the training effect of complex alternative models, and researchers often have to rely on synthetic data with lower accuracy for model training. These two challenges overlap, ultimately leading to an irreconcilable contradiction between the scarcity of actual engineering data and the need to build robust constitutive models.

[0006] Therefore, how to effectively improve the generalization ability and prediction accuracy of data-driven constitutive models in scenarios with scarce data has become a technical problem that the industry urgently needs to solve. Summary of the Invention

[0007] This application provides a data-driven constitutive modeling method and apparatus for geotechnical materials based on a knowledge transfer strategy, in order to solve the problems of insufficient generalization ability and prediction accuracy of data-driven models due to limited training data and the unpredictability of stress paths in related technologies.

[0008] The first aspect of this application provides a data-driven constitutive modeling method for geotechnical materials based on a knowledge transfer strategy, comprising the following steps: constructing a first mapping data pair of current state, strain increment, and stress increment based on the stress-strain physical quantities of sampling points of experimental curves, to establish a high-fidelity dataset of geotechnical materials based on the first mapping data pair; constructing a second mapping data pair based on the stress-strain state and intermediate physical quantity labels of sampling points of simulated curves, to establish a low-fidelity dataset of geotechnical materials based on the second mapping data pair; pre-training the target data-driven constitutive model using the low-fidelity dataset until a first preset iteration stopping condition is met, generating an initial geotechnical material data-driven constitutive model; and fine-tuning the initial geotechnical material data-driven constitutive model using the high-fidelity dataset until a second preset iteration stopping condition is met, generating a final geotechnical material data-driven constitutive model.

[0009] Through the above-mentioned technical means, the embodiments of this application can construct a first mapping data pair using experimental curve sampling points to form a high-fidelity dataset, and construct a second mapping data pair using simulated curve sampling points and intermediate physical quantity labels to form a low-fidelity dataset. First, the constitutive model is driven by the target data pre-trained using the low-fidelity dataset to obtain an initial model, and then the initial model is fine-tuned using the high-fidelity dataset until the stopping condition is met to generate the final model. This not only efficiently utilizes simulated data to complete the construction of the model's basic capabilities and reduces the dependence on high-fidelity data and training costs, but also fine-tunes the model deviation through experimental data, thereby significantly improving the accuracy of the final model in characterizing and predicting the real stress-strain behavior of geotechnical materials, and providing reliable model support for geotechnical engineering mechanics analysis and design.

[0010] Optionally, in one embodiment of this application, establishing a high-fidelity dataset for geotechnical materials includes: collecting triaxial shear test data of geotechnical materials and generating test curves based on the triaxial shear test data; performing smoothing and noise reduction processing on the test curves to generate smooth curves that reflect the overall trend of the data points; sampling the smooth curves to calculate the average principal stress, generalized shear stress, Lode angle, volumetric strain, shear strain, cumulative absolute volumetric strain increment, and cumulative absolute shear strain increment of the target sampling point, and calculating the volumetric strain increment, shear strain increment, average principal stress increment, and generalized shear stress increment from the target sampling point to the next sampling point, so as to construct a first mapping data pair of the current state, strain increment, and stress increment of the smooth curve sampling points, and establish the high-fidelity dataset.

[0011] Through the above-mentioned technical means, the embodiments of this application can collect triaxial shear test data of geotechnical materials and generate test curves. After smoothing and noise reduction processing to eliminate data interference and retain the overall trend, multi-dimensional stress and strain parameters and increments are calculated through targeted sampling. Finally, a mapping data pair between the current state and strain increments and stress increments is constructed to form a high-fidelity dataset. This provides high-quality and high-reliability data support for geotechnical material-related research or applications, and improves the accuracy and effectiveness of subsequent analysis or modeling based on this dataset.

[0012] Optionally, in one embodiment of this application, establishing the low-fidelity dataset of geotechnical materials includes: generating random curves based on a Gaussian process, and linearly transforming the random curves to generate random stress curves within a target range; using the random stress curves as loads for numerical simulation to obtain the stress-strain state of the unit at each time step, generating simulation curves based on the stress-strain states of the unit at each time step, and calculating the stress-strain state of the sampling points of the simulation curves; using the stress-strain states of the sampling points for loading and unloading stress detection of the unit, generating the total strain and residual strain of the unit, and evaluating the elastic parameters, plastic flow direction, and plastic modulus of the unit under the current state based on the difference between the total strain and the residual strain, generating the intermediate physical quantity labels; constructing the second mapping data pair according to the stress-strain states of the sampling points of the simulation curves and the intermediate physical quantity labels, and establishing a low-fidelity numerical simulation dataset.

[0013] Through the aforementioned technical means, this application embodiment generates random curves through a Gaussian process and obtains random stress curves within the target range through linear transformation. Using these curves as load simulations, the stress-strain state of the unit body is obtained and a simulation curve is generated. Combined with loading and unloading stress detection, the total strain and residual strain are obtained, and intermediate physical quantity labels such as elastic parameters are evaluated. Finally, a mapping data pair between stress-strain state and intermediate physical quantity labels is constructed, forming a low-fidelity numerical simulation dataset. This dataset can supplement geotechnical material research with rich simulation data that conforms to physical laws, broadening the data sources while ensuring the data's adaptability and usability for subsequent analysis or modeling.

[0014] Optionally, in one embodiment of this application, the step of fine-tuning the initial geotechnical material data-driven constitutive model using the high-fidelity dataset includes: partitioning the high-fidelity dataset to construct a training set for model training; loading the network weight parameters of the initial geotechnical material data-driven constitutive model and freezing the network weights; unfreezing the network layer weights of the initial geotechnical material data-driven constitutive model layer by layer to retrain single network layer weights using the training set, generating data-driven constitutive models with different network layer fine-tuning; and using the data-driven constitutive models with different network layer fine-tuning... The system continuously predicts triaxial shear tests of unit cells, generates data-driven model prediction results, and uses the initial geotechnical material data-driven constitutive model as a benchmark model to predict triaxial shear tests of unit cells, generating physical test results. It also calculates the error evaluation index between the data-driven model prediction results and the physical test results, and identifies network layers that meet preset conditions based on the error evaluation index. The system then reloads the initial geotechnical material constitutive model and unfreezes the weights of the network layers that meet the preset conditions. Finally, it uses the training set to fine-tune the initial geotechnical material data-driven constitutive model, generating the final data-driven constitutive model.

[0015] Through the above-mentioned technical means, the embodiments of this application can generate a training set by dividing a high-fidelity dataset, load the initial geotechnical material data to drive the constitutive model weights and freeze and unfreeze them layer by layer to retrain a single network layer to obtain different fine-tuned models. Then, these fine-tuned models are used to carry out continuous prediction of triaxial shear tests of unit cells. The network layer that meets the preset conditions is calculated by combining the error evaluation index with the physical test results of the initial model. Finally, the initial model is formally fine-tuned based on the network layer to generate the final model. This can efficiently utilize high-fidelity data, avoid blind fine-tuning, improve the prediction accuracy of the model for triaxial shear tests of unit cells, ensure the applicability and reliability of the final data-driven constitutive model, and provide effective model support for the analysis of the mechanical properties of geotechnical materials.

[0016] Optionally, in one embodiment of this application, the intermediate physical quantity label is generated by evaluating the elastic parameters, plastic flow direction, and plastic modulus of the unit body in its current state, wherein the calculation formula for the intermediate physical quantity label is as follows:

[0017]

[0018]

[0019] Among them, superscript This indicates that the variable is a data label. Indicates bulk modulus. Indicates shear modulus, Indicates the average stress increment. Indicates the equivalent stress increment. Represents the volumetric elastic strain increment. This represents the equivalent elastic strain increment. n g，p and n g，q Representing the two components of the plastic flow direction in the pq stress space, and These represent the volumetric plastic strain increment and the equivalent plastic strain increment, respectively.

[0020] Through the above-mentioned technical means, the embodiments of this application can evaluate the elastic parameters, plastic flow direction and plastic modulus under the current state in the constitutive model of geotechnical materials. This can greatly improve the model's prediction accuracy of elastic deformation, plastic deformation path and strength hardening / softening response of geotechnical materials, break through the limitations of traditional fixed parameter models to adapt to different stress levels and complex working conditions (such as foundation pit excavation and tunnel construction), and can also dynamically feed back the real-time mechanical state of geotechnical materials. This provides key basis for differentiated engineering design, support scheme optimization and risk management (such as slope instability early warning), and effectively connects constitutive model theory with engineering practice.

[0021] A second aspect of this application provides a data-driven constitutive modeling device for geotechnical materials based on a knowledge transfer strategy, comprising: a first construction module, configured to construct a first mapping data pair of current state, strain increment, and stress increment based on stress-strain physical quantities at sampling points of an experimental curve, to establish a high-fidelity dataset of geotechnical materials based on the first mapping data pair; a second construction module, configured to construct a second mapping data pair based on stress-strain state and intermediate physical quantity labels at sampling points of a simulated curve, to establish a low-fidelity dataset of geotechnical materials based on the second mapping data pair; a generation module, configured to pre-train a target data-driven constitutive model using the low-fidelity dataset until a first preset iteration stopping condition is met, generating an initial geotechnical material data-driven constitutive model; and a modeling module, configured to fine-tune the initial geotechnical material data-driven constitutive model using the high-fidelity dataset until a second preset iteration stopping condition is met, generating a final geotechnical material data-driven constitutive model.

[0022] Through the above-mentioned technical means, the embodiments of this application can construct a first mapping data pair using experimental curve sampling points to form a high-fidelity dataset, and construct a second mapping data pair using simulated curve sampling points and intermediate physical quantity labels to form a low-fidelity dataset. First, the constitutive model is driven by the target data pre-trained using the low-fidelity dataset to obtain an initial model, and then the initial model is fine-tuned using the high-fidelity dataset until the stopping condition is met to generate the final model. This not only efficiently utilizes simulated data to complete the construction of the model's basic capabilities and reduces the dependence on high-fidelity data and training costs, but also fine-tunes the model deviation through experimental data, thereby significantly improving the accuracy of the final model in characterizing and predicting the real stress-strain behavior of geotechnical materials, and providing reliable model support for geotechnical engineering mechanics analysis and design.

[0023] Optionally, in one embodiment of this application, the first construction module includes: a first generation unit, used to collect triaxial shear test data of soil and rock materials and generate test curves based on the triaxial shear test data; a second generation unit, used to perform smoothing and noise reduction processing on the test curves to generate smooth curves that reflect the overall trend of the data points; and a construction unit, used to sample the smooth curves to calculate the average principal stress, generalized shear stress, Lode angle, volumetric strain, shear strain, cumulative absolute volumetric strain increment, and cumulative absolute shear strain increment of the target sampling point, and to calculate the volumetric strain increment, shear strain increment, average principal stress increment, and generalized shear stress increment from the target sampling point to the next sampling point, so as to construct a first mapping data pair of the current state, strain increment, and stress increment of the smooth curve sampling point, and establish the high-fidelity dataset.

[0024] Through the above-mentioned technical means, the embodiments of this application can collect triaxial shear test data of geotechnical materials and generate test curves. After smoothing and noise reduction processing to eliminate data interference and retain the overall trend, multi-dimensional stress and strain parameters and increments are calculated through targeted sampling. Finally, a mapping data pair between the current state and strain increments and stress increments is constructed to form a high-fidelity dataset. This provides high-quality and high-reliability data support for geotechnical material-related research or applications, and improves the accuracy and effectiveness of subsequent analysis or modeling based on this dataset.

[0025] Optionally, in one embodiment of this application, the second construction module includes: a third generation unit, configured to generate a random curve based on a Gaussian process, and linearly transform the random curve to generate a random stress curve within a target range; a calculation unit, configured to use the random stress curve as a load for numerical simulation to obtain the stress-strain state of the unit body at each time moment, generate a simulation curve based on the stress-strain state of the unit body at each time moment, and calculate the stress-strain state of the sampling points of the simulation curve; a fourth generation unit, configured to use the stress-strain state of the sampling points for loading and unloading stress detection of the unit body, generate the total strain and residual strain of the unit body, and evaluate the elastic parameters, plastic flow direction, and plastic modulus of the unit body under the current state based on the difference between the total strain and the residual strain, and generate the intermediate physical quantity label; and an establishment unit, configured to construct the second mapping data pair according to the stress-strain state of the sampling points of the simulation curve and the intermediate physical quantity label, and establish a low-fidelity numerical simulation dataset.

[0026] Through the aforementioned technical means, this application embodiment generates random curves through a Gaussian process and obtains random stress curves within the target range through linear transformation. Using these curves as load simulations, the stress-strain state of the unit body is obtained and a simulation curve is generated. Combined with loading and unloading stress detection, the total strain and residual strain are obtained, and intermediate physical quantity labels such as elastic parameters are evaluated. Finally, a mapping data pair between stress-strain state and intermediate physical quantity labels is constructed, forming a low-fidelity numerical simulation dataset. This dataset can supplement geotechnical material research with rich simulation data that conforms to physical laws, broadening the data sources while ensuring the data's adaptability and usability for subsequent analysis or modeling.

[0027] Optionally, in one embodiment of this application, the modeling module includes: a fifth generation unit, used to partition the high-fidelity dataset and construct a training set for model training; a sixth generation unit, used to load the network weight parameters of the initial geotechnical material data-driven constitutive model, freeze the network weights, and unfreeze the network layer weights of the initial geotechnical material data-driven constitutive model layer by layer, so as to retrain the single network layer weights using the training set to generate a data-driven constitutive model with different network layer fine-tuning; and a search unit, used to perform triaxial shearing of the element volume using the data-driven constitutive model with different network layer fine-tuning. The system continuously predicts shear tests, generates data-driven model prediction results, and uses the initial geotechnical material data-driven constitutive model as a benchmark model to perform triaxial shear test predictions on unit cells, generating physical test results. It also calculates the error evaluation index between the data-driven model prediction results and the physical test results, and identifies network layers that meet preset conditions based on the error evaluation index. The seventh generation unit is used to reload the initial geotechnical material constitutive model and unfreeze the weights of the network layers that meet the preset conditions. It then uses the training set to fine-tune the initial geotechnical material data-driven constitutive model, generating the final data-driven constitutive model.

[0028] Through the above-mentioned technical means, the embodiments of this application can generate a training set by dividing a high-fidelity dataset, load the initial geotechnical material data to drive the constitutive model weights and freeze and unfreeze them layer by layer to retrain a single network layer to obtain different fine-tuned models. Then, these fine-tuned models are used to carry out continuous prediction of triaxial shear tests of unit cells. The network layer that meets the preset conditions is calculated by combining the error evaluation index with the physical test results of the initial model. Finally, the initial model is formally fine-tuned based on the network layer to generate the final model. This can efficiently utilize high-fidelity data, avoid blind fine-tuning, improve the prediction accuracy of the model for triaxial shear tests of unit cells, ensure the applicability and reliability of the final data-driven constitutive model, and provide effective model support for the analysis of the mechanical properties of geotechnical materials.

[0029] Optionally, in one embodiment of this application, the calculation formula for the intermediate physical quantity label is as follows:

[0030]

[0031]

[0032] Among them, superscript This indicates that the variable is a data label. Indicates bulk modulus. Indicates shear modulus, Indicates the average stress increment. Indicates the equivalent stress increment. Represents the volumetric elastic strain increment. This represents the equivalent elastic strain increment. n g，p and n g，q Representing the two components of the plastic flow direction in the pq stress space, and These represent the volumetric plastic strain increment and the equivalent plastic strain increment, respectively.

[0033] Through the above-mentioned technical means, the embodiments of this application can evaluate the elastic parameters, plastic flow direction and plastic modulus under the current state in the constitutive model of geotechnical materials. This can greatly improve the model's prediction accuracy of elastic deformation, plastic deformation path and strength hardening / softening response of geotechnical materials, break through the limitations of traditional fixed parameter models to adapt to different stress levels and complex working conditions (such as foundation pit excavation and tunnel construction), and can also dynamically feed back the real-time mechanical state of geotechnical materials. This provides key basis for differentiated engineering design, support scheme optimization and risk management (such as slope instability early warning), and effectively connects constitutive model theory with engineering practice.

[0034] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the geotechnical material data-driven modeling method based on a knowledge transfer strategy as described in the above embodiments.

[0035] A fourth aspect of this application provides a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described geotechnical material data-driven modeling method based on a knowledge transfer strategy.

[0036] A fifth aspect of this application provides a computer program product that stores a computer program that, when executed by a processor, implements the above-described geotechnical material data-driven modeling method based on a knowledge transfer strategy.

[0037] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0038] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a geotechnical material data-driven modeling method based on a knowledge transfer strategy, according to an embodiment of this application. Figure 2This is a schematic diagram of an apparatus for a physical test according to a specific embodiment of this application; Figure 3 This is a physical test curve before and after smoothing according to a specific embodiment of this application; Figure 4 This is a distribution diagram of a low-fidelity dataset according to a specific embodiment of this application; Figure 5 This is a schematic diagram of a data-driven constitutive model architecture for code generalized plasticity according to a specific embodiment of this application; Figure 6 This is a schematic diagram illustrating the training and validation results of a pre-trained model according to a specific embodiment of this application; Figure 7 This is a schematic diagram of the layer-by-layer sensitivity analysis results of a model according to a specific embodiment of this application; Figure 8 This is a schematic diagram showing the prediction effect of the model before and after fine-tuning under a triaxial drainage shear path according to a specific embodiment of this application. Figure 9 This is a schematic diagram showing the prediction effect of the model before and after fine-tuning according to a specific embodiment of this application under triaxial non-draining shear and isopic triaxial path. Figure 10 This is a schematic diagram of the structure of the geotechnical material data-driven modeling device based on a knowledge transfer strategy, according to an embodiment of this application. Figure 11 This is a schematic diagram of an electronic device structure according to an embodiment of this application. Detailed Implementation

[0039] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0040] Geotechnical materials are widely present in the natural environment and are also core materials in key engineering fields such as water conservancy, civil engineering, and transportation. The accurate characterization of their mechanical behavior is directly related to the quality of engineering projects throughout their entire life cycle, from safe design and efficient construction to long-term stable operation and maintenance. Therefore, they hold a vital position in engineering practice.

[0041] The following describes, with reference to the accompanying drawings, a data-driven constitutive modeling method and apparatus for geotechnical materials based on a knowledge transfer strategy, according to embodiments of this application. Addressing the issues mentioned in the background section regarding the insufficient generalization ability and prediction accuracy of data-driven models due to limited training data and the unpredictability of stress paths, this application provides a data-driven constitutive modeling method for geotechnical materials based on a knowledge transfer strategy. In this method, a high-fidelity dataset of geotechnical materials can be constructed through physical experiments; a low-fidelity dataset of geotechnical materials can be constructed through numerical simulations; and a data-driven constitutive model can be pre-trained based on the low-fidelity dataset; and the data-driven constitutive model can be fine-tuned based on the high-fidelity dataset. This method integrates low-fidelity numerical simulation data and high-fidelity physical experiment data through transfer learning technology. It utilizes the low-fidelity data to ensure the model's path generalization ability and utilizes the high-fidelity data to improve the model's prediction accuracy, ultimately providing a practical solution for developing robust data-driven constitutive models for geotechnical materials under conditions of data scarcity. This solves the problems of insufficient generalization ability and prediction accuracy of data-driven models due to limited training data and the unpredictability of stress paths in related technologies.

[0042] Specifically, Figure 1 This is a flowchart illustrating a data-driven modeling method for geotechnical materials based on a knowledge transfer strategy, as provided in an embodiment of this application.

[0043] like Figure 1 As shown, the geotechnical material data-driven modeling method based on a knowledge transfer strategy includes the following steps: In step S101, a first mapping data pair of current state, strain increment and stress increment is constructed based on the stress and strain physical quantities of the sampling points of the test curve, so as to establish a high-fidelity dataset of geotechnical materials based on the first mapping data pair.

[0044] The stress and strain physical quantities at the sampling point may include, but are not limited to: the average principal stress, generalized shear stress, Lode angle, volumetric strain, shear strain, cumulative absolute volumetric strain increment and cumulative absolute shear strain increment at the sampling point, as well as the volumetric strain increment, shear strain increment, average principal stress increment and generalized shear stress increment from the sampling point to the next sampling point. The specific values ​​can be set by those skilled in the art in combination with the actual application scenario, and no specific limitations are made here.

[0045] Furthermore, high-fidelity datasets for geotechnical materials refer to datasets that can accurately, comprehensively, and realistically reproduce the physical and mechanical properties and engineering responses of geotechnical materials, with "fidelity" being the core. They must possess characteristics such as accurate data (measured values ​​closely approximate real-world properties), complete information (covering key physical and mechanical parameters), scenario matching (experimental conditions closely match actual engineering projects), and data consistency (no logical contradictions). They can provide precise input for numerical simulations in geotechnical engineering and a reliable foundation for AI model applications, serving as a key support for promoting the precision development of geotechnical engineering.

[0046] In one embodiment of this application, establishing a high-fidelity dataset for geotechnical materials includes: collecting triaxial shear test data of geotechnical materials and generating test curves based on the triaxial shear test data; performing smoothing and noise reduction processing on the test curves to generate smooth curves that reflect the overall trend of the data points; sampling the smooth curves to calculate the average principal stress, generalized shear stress, Lode angle, volumetric strain, shear strain, cumulative absolute volumetric strain increment, and cumulative absolute shear strain increment of the target sampling point, and calculating the volumetric strain increment, shear strain increment, average principal stress increment, and generalized shear stress increment from the target sampling point to the next sampling point, so as to construct a mapping data pair of the current state, strain increment, and stress increment of the smooth curve sampling points, and establish a high-fidelity dataset.

[0047] The experimental data may include, but are not limited to, data on axial load, deformation, and drainage volume; the experimental curves may include, for example, deviatoric stress curves and volumetric strain curves.

[0048] As one possible approach, embodiments of this application can construct high-fidelity datasets of geotechnical materials through physical experiments. Specific steps can be set as follows: S101.1: The high-fidelity physical test dataset is constructed based on triaxial shear tests of soil and rock materials. The test can be conducted using a large triaxial apparatus for coarse-grained soil. The sample is consolidated by injecting water into the saturation tank, and the sample is loaded by displacement control of the upper loading plate. During the test, data such as axial load, deformation, and drainage volume are collected and organized into deviatoric stress curves and volumetric strain curves.

[0049] For example, this step can be performed as follows: Figure 2 The experimental setup shown was used to conduct physical tests on the rockfill material for a high-faced rockfill dam project; according to the design requirements for rockfill material filling, as per section 2.13 The density standard controls the compaction degree. During sample preparation, to prevent the separation of coarse and fine particles and the distribution of particles with different densities along the height, a five-layer layering method is used to prepare the test specimen. After each layer is filled, it is compacted with a sledgehammer, and about 5% water is added to reduce dust and friction between particles. The test specimen is surrounded by two layers of rubber membrane. The inner membrane is used to maintain the shape of the specimen and prevent sharp edges of the gravel inside the specimen from piercing the rubber membrane, while the outer membrane is used for waterproofing. The specimen is saturated using a combination of vacuum saturation and water head saturation. A vacuum pump is used to evacuate air from the top of the specimen until the negative pressure reaches 60~80. After stabilizing for a period of time, it reached 2 Water is slowly pumped into the sample from the bottom until water comes out of the drain pipe on the sample; then the vacuum is released and water is pumped into the saturation tank to saturate the sample using the water head.

[0050] Furthermore, respectively according to =0.2, 0.4, 0.8, 1.6, 2.5 The specimens were consolidated under the confining pressure level, with the pressure applied slowly during consolidation to ensure sufficient drainage. During shearing, strain-controlled loading was used, with the strain rate controlled at 0.5%. Throughout the test, the computer automatically collected data on axial load, deformation, and shear drainage volume. After preliminary processing, the deviatoric stress curve and volumetric strain curve of the test rockfill material during the conventional triaxial shear test were obtained.

[0051] S101.2: Smooth spline interpolation is used to smooth and reduce noise in the test curves. Specifically, it can be set as follows: For a given array of curve coordinates and Construct a cubic spline function Optimization methods are used to make The following optimization objectives must be met:

[0052] In the formula, It is the first on the curve coordinate points, spline function The second derivative of the first term is summed over all data points to penalize the deviation between the spline function and the data points; the integral of the second term is applied over the entire interval. The above is used to penalize large values ​​of the second derivative; It is a non-negative penalty parameter, when When only the first objective remains, the resulting curve will be forced to pass through each data point. In this case, the optimization objective simultaneously includes fitting the data points and restricting the second derivative, resulting in a smooth curve that reflects the overall trend of the data points.

[0053] Due to objective factors such as environmental mechanical vibration interference, loading system noise, and human operational errors, physical test data are inevitably affected by noise disturbances, resulting in fluctuations in the obtained test curves. These fluctuations can lead to inaccurate stress increment labels or even stress increments in the opposite direction when preparing incremental datasets for materials, thus significantly affecting the data-driven model's learning of the material's true mechanical properties. To mitigate the impact of unreasonable data fluctuations caused by these errors on model training, this application embodiment employs smooth spline interpolation to smooth, reduce noise, and refine the aforementioned test curves.

[0054] After preliminary calculations and adjustments, different penalty coefficients were selected for different experimental curves. For deviatoric stress curves with significant data fluctuations, a penalty coefficient is applied. To enhance the smoothing effect, a penalty coefficient is applied to volumetric strain curves where data fluctuations are not significant. To ensure that the smoothed results better match the original experimental data, the experimental curves before and after smoothing are as follows: Figure 3 As shown.

[0055] S101.3: Sample the smoothed test curve and calculate the sampling points. Average principal stress at Generalized shear stress Lode Corner Volumetric strain Shear strain Cumulative absolute volumetric strain increment and cumulative absolute shear strain increment Simultaneously calculate the volumetric strain increment from the sampling point to the next sampling point. Shear strain increment Mean principal stress increment and generalized shear stress increment The specific calculation formula is as follows: Mean principal stress: , Generalized shear stress: , Lode's Corner: , Volumetric strain: , Shear strain: , Cumulative absolute volumetric strain increment: , Cumulative absolute shear strain increment: , Volumetric strain increment: , Shear strain increment: , Mean principal stress increment: , Generalized shear stress increment: , in, and These are the sampling points The stress tensor and strain tensor at the point, It is a second-order unit tensor. and Sampling points The second and third order invariants of the deviatoric stress tensor at the location.

[0056] Second-order invariants of deviatoric stress: , Third-order invariants of deviatoric stress: , Using the above physical quantities, a mapping data pair of current state, strain increment, and stress increment can be constructed to establish a high-fidelity physical test dataset.

[0057] For example, this step can be configured to sample the smoothed test curve every 0.01% of axial strain and calculate the sampling points. Average principal stress at Generalized shear stress Lode Corner , Volumetric strain Shear strain Cumulative absolute volumetric strain increment and cumulative absolute shear strain increment Simultaneously calculate the volumetric strain increment from the sampling point to the next sampling point. Shear strain increment Mean principal stress increment and generalized shear stress increment We construct mapping data pairs of current state, strain increment, and stress increment to establish a high-fidelity physical test dataset.

[0058] Through the above-mentioned technical means, the embodiments of this application can collect triaxial shear test data of geotechnical materials and generate test curves. After smoothing and noise reduction processing to eliminate data interference and retain the overall trend, multi-dimensional stress and strain parameters and increments are calculated through targeted sampling. Finally, a mapping data pair between the current state and strain increments and stress increments is constructed to form a high-fidelity dataset. This provides high-quality and high-reliability data support for geotechnical material-related research or applications, and improves the accuracy and effectiveness of subsequent analysis or modeling based on this dataset.

[0059] In step S102, a second mapping data pair is constructed based on the stress-strain state and intermediate physical quantity labels of the sampling points of the simulated curve, so as to establish a low-fidelity dataset of geotechnical materials based on the second mapping data pair.

[0060] In this context, the stress-strain state at the sampling points of the simulated curve can be understood as the stress state and cumulative strain value at each sampling point. Low-fidelity datasets for geotechnical materials refer to data sets that, compared to high-fidelity datasets, have certain shortcomings in accuracy, completeness, and scene representativeness, but possess characteristics such as low cost and large quantity. They are typically generated through numerical simulations, empirical formulas, etc. Although the data deviation is relatively large and the accuracy is not high, they can provide certain references and supplements for geotechnical engineering research and analysis. Especially when high-fidelity data is difficult to obtain or too costly, low-fidelity datasets can be combined with high-fidelity datasets, using data fusion and other techniques to improve the model's prediction accuracy and generalization ability.

[0061] Optionally, in one embodiment of this application, establishing a low-fidelity dataset for geotechnical materials includes: generating random curves based on a Gaussian process and linearly transforming the random curves to generate random stress curves within a target range; using the random stress curves as loads for numerical simulation to obtain the stress-strain state of the element at each time step; generating simulation curves based on the stress-strain state of the element at each time step and calculating the stress-strain state at the sampling points of the simulation curves; using the stress-strain state at the sampling points for loading and unloading stress detection of the element to generate the total strain and residual strain of the element; evaluating the elastic parameters, plastic flow direction, and plastic modulus of the element under the current state based on the difference between the total strain and the residual strain, and generating intermediate physical quantity labels; constructing mapping data pairs based on the stress-strain state of the sampling points of the simulation curves and the intermediate physical quantity labels to establish a low-fidelity numerical simulation dataset.

[0062] As one possible approach, embodiments of this application can construct low-fidelity datasets of geotechnical materials through numerical simulation. The specific steps are as follows: S102.1: Generate three random curves using a Gaussian process. A Gaussian process is a method for generating random curves from a continuous domain. Indexed random variables The resulting stochastic process can be denoted as:

[0063] In the formula, and These are the mean function and the kernel function, respectively, which describe the mean of the Gaussian distribution at the sampling point and the covariance between different sampling points.

[0064] For simplicity, the mean function can be a constant zero, and the kernel function can be a radial basis function, denoted as .

[0065] in and These are different sampling points. and These are the parameters of the radial basis functions. (Symbol) express The Euclidean norm can describe the sampling points. and sampling points The distance between them.

[0066] For example, this step can use a Gaussian process to generate 1800 random curves, and divide them into groups of 3, forming 600 groups of random curves.

[0067] S102.2: By generating random parameters using a random function and performing a linear transformation on the random curve, a random stress curve within a specified stress range is obtained. Specifically:

[0068] in, and Random parameters generated by a random function.

[0069] For example, this step can use a random function to generate random parameters and perform a linear transformation on 600 sets of random curves to obtain an initial stress range of 100. up to 5 The random stress path.

[0070] S102.3: Three random stress curves are applied as loads to the three principal directions of the element, and numerical simulation is performed to record the stress-strain state of the element at each time point.

[0071] For example, this step can use three random stress curves from a set of random stress paths as loads, applied to the three principal directions of the element, and use the uniform hardening (UH) model considering particle breakage as the constitutive model of the dam rockfill to carry out numerical simulation, record the stress and strain state of the element at each time, obtain the strain response of the rockfill under random stress paths, and construct stress-strain data pairs. Here, the above process can be repeated 600 times using 600 sets of random stress paths to obtain 600 sets of low-fidelity stress-strain data pairs.

[0072] S102.4: Calculate the stress state and cumulative strain value at each sampling point according to the method in S101.3. Extract the results of the sampling point state for loading and unloading stress detection. Record the total strain and residual strain during the stress detection process. The total strain is used as the total strain increment during the detection process, and the residual strain is used as the plastic strain increment during the detection process. The difference between the total strain and the residual strain is used as the elastic strain increment during the detection process to evaluate the elastic parameters, plastic flow direction, and plastic modulus under the current state. The calculation formulas for these intermediate physical quantities can be set as follows:

[0073]

[0074]

[0075] Among them, superscript This indicates that the variable is a data label. Indicates bulk modulus. Indicates shear modulus, This represents the average stress increment (the average principal stress increment during the detection process). It represents the equivalent stress increment (the generalized shear stress increment during the detection process). It represents the volumetric elastic strain increment (elastic volumetric strain increment). It represents the equivalent elastic strain increment (shear strain increment). n g，p and n g，q Representing the two components of the plastic flow direction in the pq stress space, and These represent the volumetric plastic strain increment (plastic volumetric strain increment) and the equivalent plastic strain increment (plastic shear strain increment), respectively.

[0076] In the formula, , , and It can be calculated based on the total strain and residual strain during the detection process.

[0077] Furthermore, in this embodiment of the application, the stress-strain state of the sampling points and the aforementioned intermediate physical quantity labels can be used to construct mapping data pairs and establish a low-fidelity numerical simulation dataset.

[0078] Specifically, in this embodiment of the application, each random stress path can be sampled 50 times at equal intervals, and the sampling points can be calculated. Average principal stress at Generalized shear stress Lode Corner Volumetric strain Shear strain Cumulative absolute volumetric strain increment and cumulative absolute shear strain increment Simultaneously calculate the volumetric strain increment from the sampling point to the next sampling point. Shear strain increment Mean principal stress increment and generalized shear stress increment .

[0079] Furthermore, the results of each sampling point's state are extracted for loading and unloading stress detection. The total strain and residual strain are recorded during the stress detection process. The total strain is used as the total strain increment during the detection process, and the residual strain is used as the plastic strain increment. The difference between the total strain and the residual strain is used as the elastic strain increment during the detection process. This allows for the evaluation of the elastic parameters, plastic flow direction, and plastic modulus under the current state. Furthermore, mapping data pairs are constructed using the stress-strain state of the sampling points and intermediate physical quantity labels to establish a low-fidelity numerical simulation dataset. The low-fidelity dataset is distributed as follows: Figure 4 As shown.

[0080] Through the aforementioned technical means, this application embodiment generates random curves through a Gaussian process and obtains random stress curves within the target range through linear transformation. Using these curves as load simulations, the stress-strain state of the unit body is obtained and a simulation curve is generated. Combined with loading and unloading stress detection, the total strain and residual strain are obtained, and intermediate physical quantity labels such as elastic parameters are evaluated. Finally, a mapping data pair between stress-strain state and intermediate physical quantity labels is constructed, forming a low-fidelity numerical simulation dataset. This dataset can supplement geotechnical material research with rich simulation data that conforms to physical laws, broadening the data sources while ensuring the data's adaptability and usability for subsequent analysis or modeling.

[0081] In step S103, the target data-driven constitutive model is pre-trained using a low-fidelity dataset until the first preset iteration stopping condition is met, thereby generating an initial geotechnical material data-driven constitutive model.

[0082] The first preset iteration stop condition can be set by relevant technical personnel according to actual needs, and no specific limitation is made here.

[0083] Specifically, as one possible approach, embodiments of this application can pre-train a data-driven constitutive model based on a low-fidelity dataset. The specific steps can be set as follows: S103.1: Divide the low-fidelity numerical simulation dataset into training and validation sets.

[0084] For example, this step can divide the low-fidelity numerical simulation data into a training set and a validation set in a 7:3 ratio. The training set is used to train the data-driven constitutive model, and the validation set is used to prevent overfitting.

[0085] S103.2: Optimize the hyperparameters of the training model (i.e. the constructed data-driven constitutive model), use the optimized hyperparameters to perform and complete the pre-training of the data constitutive model on a low-fidelity data training set, and at the same time avoid model overfitting by monitoring the validation set loss.

[0086] For example, this step can employ a data-driven constitutive modeling framework that encodes generalized plasticity to construct a data-driven constitutive model, such as... Figure 5 As shown, the model consists of three sub-networks: the first predicts elastic parameters, the second predicts the plastic flow direction, and the third predicts the plasticity model. The stress increment is calculated by combining the predictions from these three sub-networks with the generalized plasticity framework. To fully learn the physical knowledge contained in the data, the data-driven constitutive model is pre-trained using a loss function constructed from intermediate physical quantities such as elastic parameters, plastic flow direction, and plastic modulus.

[0087] After hyperparameter optimization, the hyperparameters shown in Table 1 were finally selected for pre-training of the data-driven constitutive model: Table 1

[0088] The network structure in the table represents the number of neurons in each layer of each sub-network. The model is trained using the Adam optimizer with default parameters, and the initial learning rate is [value missing]. When the loss function fails to decrease, the ReduceLROnPlateau strategy is used to automatically reduce the learning rate. The validation set loss is monitored in real time during pre-training, and no overfitting occurs. Model weights are saved after training is complete.

[0089] S103.3: Using the trained data-driven constitutive model, continuous predictions are made for different stress paths. The prediction results are then compared with the numerical simulation results under the same stress path. An evaluation index is calculated between the data-driven constitutive model's prediction results and the original numerical simulation results to confirm whether the data-driven pre-trained model has grasped the mechanical properties of the numerical simulation data. The calculation formulas for each evaluation index are as follows: Coefficient of determination:

[0090] Mean square error:

[0091] Mean absolute error:

[0092] In the formula, This is a reference result. It is a prediction result. It is the average of the reference results. It refers to the number of samples.

[0093] For example, this step can use the trained data-driven constitutive model to continuously predict triaxial drained shear tests, triaxial undrained shear tests, and isopeconic triaxial shear tests, respectively. The results of triaxial drained shear tests, triaxial undrained shear tests, and isopeconic triaxial shear tests using the UH model considering particle breakage can be used as a control. The control results are as follows: Figure 6 As shown in Table 2, the evaluation index of the model prediction results is calculated. The R² of both the data-driven model prediction results and the UH model prediction results considering particle breakage are greater than 0.99, which verifies that the pre-trained model has mastered the mechanical properties of the UH model considering particle breakage under different stress paths.

[0094] Table 2

[0095] This application embodiment can divide the low-fidelity numerical simulation dataset into training and validation sets, optimize hyperparameters to complete the pre-training of the data-driven constitutive model, and avoid overfitting by using validation set loss monitoring. Then, by continuously predicting different stress paths and comparing them with numerical simulation results, and calculating evaluation indicators, the low-fidelity data can be used to achieve effective model pre-training, ensuring model reliability. At the same time, it can accurately confirm whether the model has mastered the mechanical properties of the numerical simulation data, thus laying the foundation for the subsequent application or further optimization of the data-driven constitutive model.

[0096] In step S104, the initial geotechnical material data is used to fine-tune the model until the second preset iteration stop condition is met, and the final geotechnical material data is generated to drive the model.

[0097] The second preset iteration stop condition can be set by relevant technical personnel according to actual needs, and no specific limitation is made here.

[0098] Optionally, in one embodiment of this application, fine-tuning the pre-trained model using a high-fidelity dataset includes: dividing the high-fidelity dataset to construct a training set for model training; loading the network weight parameters of the initial geotechnical material data-driven constitutive model and freezing the network weights, then unfreezing the network layer weights of the initial geotechnical material data-driven constitutive model layer by layer to retrain the weights of a single network layer using the training set, generating data-driven constitutive models with different network layer fine-tuning; continuously predicting triaxial shear tests of unit cells using data-driven constitutive models with different network layer fine-tuning, generating data-driven model prediction results, and using the initial geotechnical material data-driven constitutive model as a benchmark model to predict triaxial shear tests of unit cells, generating physical test results, and calculating the error evaluation index between the data-driven model prediction results and the physical test results, and identifying network layers that meet preset conditions based on the error evaluation index; reloading the initial geotechnical material data-driven constitutive model and unfreezing the network layer weights that meet the preset conditions, then formally fine-tuning the initial geotechnical material data-driven constitutive model using the training set, generating the final data-driven constitutive model.

[0099] Specifically, based on a high-fidelity dataset, the data-driven constitutive model is fine-tuned. The specific steps can be set as follows: S104.1: Divide the high-fidelity physics experiment dataset into a training set and a validation set. This step can divide the high-fidelity physics experiment data into a training set and a validation set in a 7:3 ratio. The training set is used for training the data-driven constitutive model, and the validation set is used to prevent overfitting during training.

[0100] S104.2: Load the network weight parameters of the pre-trained model (i.e., the initial constitutive model of soil and rock materials), freeze the network weights of the pre-trained model, and then unfreeze the network layer weights of the pre-trained model layer by layer. Retrain the weights of individual network layers using a high-fidelity physical experiment training set to obtain models with fine-tuned different network layers; this step allows for the use of a smaller learning rate during fine-tuning. The batch size (4) and the number of training rounds are uniformly set to 100 to obtain a model that fine-tunes different network layers.

[0101] S104.3: Continuous prediction of triaxial shear tests on unit cells is performed using various fine-tuned data-driven constitutive models. Using the pre-trained model as the baseline, error evaluation indices are calculated between the data-driven model prediction results and the physical test results. The network layers that significantly improve the model's prediction accuracy are identified. The results are as follows: Figure 7 As shown, the network layers that have the greatest impact on model improvement in this fine-tuning are KG_2 and dg_2 layers.

[0102] S104.4: Based on the effect of different network layers on improving the model's predictive ability, select the network layer weights that have a higher improvement in the model's prediction accuracy and unfreeze them. Use the training set of physical experimental data to perform formal fine-tuning of the model. The validation set is still used to prevent the model from overfitting. This step can be based on the results obtained in S104.3. Reload the pre-trained model, freeze all network weights, unfreeze the network layer weights of KG_2 and dg_2 layers, and use the training set of physical experimental data to perform formal fine-tuning of the model. The hyperparameter settings for formal fine-tuning are consistent with those in the network layer sensitivity analysis. The number of training rounds is increased to 500. During the fine-tuning process, the validation set loss is monitored, and no overfitting occurs.

[0103] S104.5: Continuously calculate the stress-strain curves of riprap under different stress paths using the models before and after fine-tuning, calculate the evaluation index in S3.3, and verify the effectiveness of the fine-tuned model. This step involves continuously predicting the stress-strain curves of riprap under triaxial drained shear tests, triaxial undrained shear tests, and isopeconic triaxial shear tests using the models before and after fine-tuning. The triaxial drained shear test is used to verify numerical accuracy, while the triaxial undrained shear and isopeconic triaxial shear tests are used to verify the model's generalization ability.

[0104] Figure 8 shows the results of the triaxial drained shear test before and after fine-tuning. The triaxial drained shear physical test is used as a control. The evaluation index before and after fine-tuning is calculated, and the results are shown in Table 3. Compared with the pre-trained model, the prediction accuracy is significantly improved after fine-tuning.

[0105] Table 3

[0106] Figure 9 The results of triaxial undrained shear and isoprene triaxial shear tests on the model before and after fine-tuning are presented. The stress-strain response law of the model after fine-tuning is consistent with that before fine-tuning, and it still has the path generalization ability. Moreover, the numerical variation law of the model after fine-tuning in triaxial undrained shear path and isoprene triaxial shear path is consistent with that in triaxial drained shear path: both can improve the shear strength of the pre-trained model and reduce the shear dilatation ability of the pre-trained model. This verifies that the data-driven constitutive model constructed by this method has both strong path generalization ability and good prediction accuracy.

[0107] Through the above-mentioned technical means, the embodiments of this application can generate a training set by dividing a high-fidelity dataset, load the initial geotechnical material data to drive the constitutive model weights and freeze and unfreeze them layer by layer to retrain a single network layer to obtain different fine-tuned models. Then, these fine-tuned models are used to carry out continuous prediction of triaxial shear tests of unit cells. The network layer that meets the preset conditions is calculated by combining the error evaluation index with the physical test results of the initial model. Finally, the initial model is formally fine-tuned based on the network layer to generate the final model. This can efficiently utilize high-fidelity data, avoid blind fine-tuning, improve the prediction accuracy of the model for triaxial shear tests of unit cells, ensure the applicability and reliability of the final data-driven constitutive model, and provide effective model support for the analysis of the mechanical properties of geotechnical materials.

[0108] In summary, the geotechnical material data-driven modeling method based on a knowledge transfer strategy proposed in this application can be briefly summarized as follows: The data-driven constitutive modeling method for geotechnical materials based on a knowledge transfer strategy proposed in this application can construct a high-fidelity dataset of geotechnical materials through physical experiments; construct a low-fidelity dataset of geotechnical materials through numerical simulations; further, pre-train the data-driven constitutive model based on the low-fidelity dataset; and fine-tune the data-driven constitutive model based on the high-fidelity dataset. This method integrates low-fidelity numerical simulation data and high-fidelity physical experiment data through transfer learning technology. It utilizes the low-fidelity data to ensure the model's path generalization ability and utilizes the high-fidelity data to improve the model's prediction accuracy, ultimately providing a practical solution for developing robust data-driven constitutive models for geotechnical materials under conditions of data scarcity. This solves the problems of insufficient generalization ability and prediction accuracy of data-driven models caused by limited training data and the unpredictability of stress paths in related technologies.

[0109] Next, refer to the appendix. Figure 10 This application describes a geotechnical material data-driven modeling device based on a knowledge transfer strategy, according to an embodiment of the present application.

[0110] Figure 10 This is a block diagram of the geotechnical material data-driven modeling device based on a knowledge transfer strategy, according to an embodiment of this application.

[0111] like Figure 10 As shown, the geotechnical material data-driven modeling device 10 based on knowledge transfer strategy includes: a first construction module 100, a second construction module 200, a generation module 300, and a modeling module 400.

[0112] The first construction module 100 is used to construct a first mapping data pair of current state, strain increment and stress increment based on the stress and strain physical quantities of the sampling points of the test curve, so as to establish a high-fidelity dataset of geotechnical materials based on the first mapping data pair.

[0113] The second construction module 200 is used to construct a second mapping data pair based on the stress-strain state and intermediate physical quantity labels of the sampling points of the simulated curve, so as to establish low-fidelity data of geotechnical materials based on the second mapping data pair.

[0114] The generation module 300 is used to pre-train the target data-driven constitutive model using a low-fidelity dataset until the first preset iteration stopping condition is met, thereby generating an initial soil and rock material data-driven constitutive model.

[0115] The modeling module 400 is used to fine-tune the initial geotechnical material data using a high-fidelity dataset until the second preset iteration stopping condition is met, and then generate the final geotechnical material data to drive the model.

[0116] Optionally, in one embodiment of this application, the first construction module 100 includes: a first generation unit, a second generation unit, and a construction unit; wherein, the first generation unit is used to collect triaxial shear test data of soil and rock materials and generate test curves based on the triaxial shear test data; the second generation unit is used to smooth and denoise the test curves to generate smooth curves that reflect the overall trend of the data points; the construction unit is used to sample the smooth curves to calculate the average principal stress, generalized shear stress, Lode angle, volumetric strain, shear strain, cumulative absolute volumetric strain increment, and cumulative absolute shear strain increment of the target sampling point, and to calculate the volumetric strain increment, shear strain increment, average principal stress increment, and generalized shear stress increment from the target sampling point to the next sampling point, so as to construct a first mapping data pair of the current state, strain increment, and stress increment of the smooth curve sampling point and establish a high-fidelity dataset.

[0117] Optionally, in one embodiment of this application, the second construction module 200 includes: a third generation unit, a calculation unit, a fourth generation unit, and an establishment unit; wherein, the third generation unit is used to generate random curves based on a Gaussian process, and linearly transform the random curves to generate random stress curves within a target range; the calculation unit is used to perform numerical simulation using the random stress curves as loads to obtain the stress-strain state of the unit body at each time moment, generate simulation curves based on the stress-strain state of the unit body at each time moment, and calculate the stress-strain state of the sampling points of the simulation curves; the fourth generation unit is used to use the stress-strain state of the sampling points for loading and unloading stress detection of the unit body, generate the total strain and residual strain of the unit body, and evaluate the elastic parameters, plastic flow direction, and plastic modulus of the unit body under the current state based on the difference between the total strain and the residual strain, and generate intermediate physical quantity labels; the establishment unit is used to construct a second mapping data pair based on the stress-strain state of the sampling points of the simulation curves and the intermediate physical quantity labels, and establish a low-fidelity numerical simulation dataset.

[0118] Optionally, in one embodiment of this application, the modeling module 400 includes: a fifth, a sixth, and a seventh generation unit and a search unit; wherein, the fifth generation unit is used to partition the high-fidelity dataset and construct a training set for model training; the sixth generation unit is used to load the network weight parameters of the initial soil and rock material data-driven constitutive model, freeze the network weights, and unfreeze the network layer weights of the initial soil and rock material data-driven constitutive model layer by layer, so as to retrain the weights of a single network layer using the training set to generate a data-driven constitutive model with fine-tuned different network layers; the search unit is used to utilize the data-driven constitutive model with fine-tuned different network layers. The dynamic constitutive model continuously predicts triaxial shear tests of unit cells, generating data-driven model prediction results. It then uses the initial soil and rock material data-driven constitutive model as a benchmark model to predict triaxial shear tests of unit cells, generating physical test results. The error evaluation index between the data-driven model prediction results and the physical test results is calculated, and network layers that meet preset conditions are identified based on the error evaluation index. The seventh generation unit is used to reload the initial soil and rock material constitutive model and unfreeze the weights of network layers that meet preset conditions. The initial soil and rock material data-driven constitutive model is then fine-tuned using the training set to generate the final data-driven constitutive model.

[0119] Optionally, in one embodiment of this application, the calculation formula for the intermediate physical quantity label is as follows:

[0120]

[0121]

[0122] Among them, superscript This indicates that the variable is a data label. Indicates bulk modulus. Indicates shear modulus, Indicates the average stress increment. Indicates the equivalent stress increment. Represents the volumetric elastic strain increment. This represents the equivalent elastic strain increment. n g，p and n g，q Representing the two components of the plastic flow direction in the pq stress space, and These represent the volumetric plastic strain increment and the equivalent plastic strain increment, respectively.

[0123] It should be noted that the foregoing explanation of the embodiment of the geotechnical material data-driven construction modeling method based on knowledge transfer strategy also applies to the geotechnical material data-driven construction modeling device based on knowledge transfer strategy in this embodiment, and will not be repeated here.

[0124] The geotechnical material data-driven constitutive modeling device based on a knowledge transfer strategy proposed in this application can construct a high-fidelity dataset of geotechnical materials through physical experiments; construct a low-fidelity dataset of geotechnical materials through numerical simulations; further pre-train the data-driven constitutive model based on the low-fidelity dataset; and fine-tune the data-driven constitutive model based on the high-fidelity dataset. This method integrates low-fidelity numerical simulation data and high-fidelity physical experiment data through transfer learning technology. It utilizes the low-fidelity data to ensure the model's path generalization ability and utilizes the high-fidelity data to improve the model's prediction accuracy, ultimately providing a practical solution for developing robust geotechnical material data-driven constitutive models under data-scarce conditions. This solves the problems of insufficient generalization ability and prediction accuracy of data-driven models caused by limited training data and the unpredictability of stress paths in related technologies.

[0125] Figure 11 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 1101, the processor 1102, and the computer program stored on the memory 1101 and executable on the processor 1102.

[0126] When the processor 1102 executes the program, it implements the geotechnical material data-driven modeling method based on the knowledge transfer strategy provided in the above embodiments.

[0127] Furthermore, electronic devices also include: Communication interface 1103 is used for communication between memory 1101 and processor 1102.

[0128] The memory 1101 is used to store computer programs that can run on the processor 1102.

[0129] The memory 1101 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0130] If the memory 1101, processor 1102, and communication interface 1103 are implemented independently, then the communication interface 1103, memory 1101, and processor 1102 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 11 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0131] Optionally, in a specific implementation, if the memory 1101, processor 1102, and communication interface 1103 are integrated on a single chip, then the memory 1101, processor 1102, and communication interface 1103 can communicate with each other through an internal interface.

[0132] The processor 1102 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0133] This application also provides a non-volatile computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described geotechnical material data-driven modeling method based on a knowledge transfer strategy.

[0134] This application also provides a computer program product storing a computer program that, when executed by a processor, implements the above-described geotechnical material data-driven modeling method based on a knowledge transfer strategy.

[0135] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0136] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0137] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0138] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0139] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0140] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0141] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0142] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A data-driven constitutive modeling method for geotechnical materials based on a knowledge transfer strategy, characterized in that, Includes the following steps: Based on the stress and strain physical quantities of the experimental curve sampling points, a first mapping data pair of current state, strain increment and stress increment is constructed, and a high-fidelity dataset of geotechnical materials is established based on the first mapping data pair; A second mapping data pair is constructed based on the stress-strain state and intermediate physical quantity labels of the sampling points of the simulated curve, and a low-fidelity dataset of geotechnical materials is established based on the second mapping data pair; The target data-driven constitutive model is pre-trained using the low-fidelity dataset until the first preset iteration stopping condition is met, thereby generating an initial geotechnical material data-driven constitutive model. The initial geotechnical material data-driven constitutive model is fine-tuned using the high-fidelity dataset until the second preset iteration stopping condition is met, generating the final geotechnical material data-driven constitutive model.

2. The method according to claim 1, characterized in that, The establishment of a high-fidelity dataset for geotechnical materials includes: Collect triaxial shear test data of soil and rock materials, and generate test curves based on the triaxial shear test data; The experimental curves are smoothed and denoised to generate smooth curves that reflect the overall trend of the data points. The smooth curve is sampled to calculate the average principal stress, generalized shear stress, Lode angle, volumetric strain, shear strain, cumulative absolute volumetric strain increment, and cumulative absolute shear strain increment at the target sampling point. The volumetric strain increment, shear strain increment, average principal stress increment, and generalized shear stress increment from the target sampling point to the next sampling point are also calculated to construct the first mapping data pair of the current state, strain increment, and stress increment of the smooth curve sampling point, thus establishing the high-fidelity dataset.

3. The method according to claim 1, characterized in that, The establishment of the low-fidelity dataset for geotechnical materials includes: Random curves are generated based on a Gaussian process, and then linearly transformed to generate random stress curves within the target range. The random stress curve is used as a load for numerical simulation to obtain the stress and strain state of the unit at each time moment. Based on the stress and strain state of the unit at each time moment, a simulation curve is generated and the stress and strain state of the sampling point of the simulation curve is calculated. The stress-strain state of the sampling points is used for loading and unloading stress detection of the unit body. The total strain and residual strain of the unit body are recorded. Based on the difference between the total strain and the residual strain, the elastic parameters, plastic flow direction and plastic modulus of the unit body under the current state are evaluated, and the intermediate physical quantity label is generated. The second mapping data pair is constructed based on the stress-strain state of the sampling points of the simulated curve and the intermediate physical quantity labels to establish a low-fidelity numerical simulation dataset.

4. The method according to claim 1, characterized in that, The step of using the high-fidelity dataset to drive fine-tuning of the constitutive model based on the initial geotechnical material data includes: The high-fidelity dataset is divided to construct a training set for model training; Load the network weight parameters of the initial geotechnical material data-driven constitutive model and freeze the network weights. Unfreeze the network layer weights of the initial geotechnical material data-driven constitutive model layer by layer to retrain the single network layer weights using the training set, and generate data-driven constitutive models with fine-tuned different network layers. The constitutive model driven by different network layers is fine-tuned to continuously predict the triaxial shear test of the unit cell, generating the prediction results of the data-driven model. The constitutive model driven by the initial soil and rock material data is used as the benchmark model to predict the triaxial shear test of the unit cell, generating the physical test results. The error evaluation index between the prediction results of the data-driven model and the physical test results is calculated. Based on the error evaluation index, the network layer that meets the preset conditions is identified. The initial soil and rock material constitutive model is reloaded and the network layer weights that meet the preset conditions are unfrozen. The initial soil and rock material data-driven constitutive model is then fine-tuned using the training set to generate the final data-driven constitutive model.

5. The method according to claim 3, characterized in that, The process involves evaluating the elastic parameters, plastic flow direction, and plastic modulus of the unit cell under its current state, and generating intermediate physical quantity labels. The calculation formula for the intermediate physical quantity labels is as follows: Among them, superscript This indicates that the variable is a data label. Indicates bulk modulus. Indicates shear modulus, Indicates the average stress increment. Indicates the equivalent stress increment. Represents the volumetric elastic strain increment. This represents the equivalent elastic strain increment. n g，p and n g，q Representing the two components of the plastic flow direction in the pq stress space, and These represent the volumetric plastic strain increment and the equivalent plastic strain increment, respectively.

6. A geotechnical material data-driven constitutive modeling device based on a knowledge transfer strategy, characterized in that, include: The first construction module is used to construct a first mapping data pair of current state, strain increment and stress increment based on the stress and strain physical quantities of the sampling points of the test curve, so as to establish a high-fidelity dataset of geotechnical materials based on the first mapping data pair. The second construction module is used to construct a second mapping data pair based on the stress-strain state and intermediate physical quantity labels of the sampling points of the simulated curve, so as to establish a low-fidelity dataset of geotechnical materials based on the second mapping data pair; The generation module is used to pre-train the target data-driven constitutive model using the low-fidelity dataset until the first preset iteration stopping condition is met, thereby generating the initial geotechnical material data-driven constitutive model. The modeling module is used to fine-tune the initial geotechnical material data using the high-fidelity dataset until the second preset iteration stopping condition is met, and to generate the final geotechnical material data to drive the model.

7. The apparatus according to claim 6, characterized in that, The first building module includes: The first generation unit is used to collect triaxial shear test data of soil and rock materials and generate test curves based on the triaxial shear test data; The second generation unit is used to smooth and reduce noise on the test curve to generate a smooth curve that reflects the overall trend of the data points. A construction unit is used to sample the smooth curve to calculate the average principal stress, generalized shear stress, Lode angle, volumetric strain, shear strain, cumulative absolute volumetric strain increment, and cumulative absolute shear strain increment of the target sampling point, and to calculate the volumetric strain increment, shear strain increment, average principal stress increment, and generalized shear stress increment from the target sampling point to the next sampling point, so as to construct the first mapping data pair of the current state, strain increment, and stress increment of the smooth curve sampling point, and establish the high-fidelity dataset.

8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the geotechnical material data-driven modeling method based on a knowledge transfer strategy as described in any one of claims 1-5.

9. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the geotechnical material data-driven modeling method based on the knowledge transfer strategy as described in any one of claims 1-5.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement the geotechnical material data-driven modeling method based on a knowledge transfer strategy as described in any one of claims 1-5.

Images