Sandy soil liquefaction parameter rapid calibration method based on particle mesoscopic simulation
By combining the discrete element and lattice spring models with deep learning technology, the problems of low efficiency and poor adaptability of sand liquefaction parameter calibration are solved, and fast and accurate liquefaction parameter calibration is achieved, which is suitable for complex geological conditions.
Patent Information
- Application Number
- CN202510809183.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-09-26
AI Technical Summary
Existing technologies for studying the sand liquefaction mechanism require a large number of indoor experiments and have difficulty effectively correlating microscopic and macroscopic liquefaction behaviors with mechanical responses. Traditional methods are inefficient and have large parameter conversion errors, making them difficult to adapt to complex geological conditions.
Through collaborative simulation of discrete element models and discrete lattice spring models, combined with deep learning technology, the relationship between particle-scale liquefaction mechanism and macroscopic mechanical response is established, and the propagation neural network and professional enhanced large language model are used to achieve rapid calibration of liquefaction parameters.
It achieves rapid calibration of sand liquefaction parameters, breaks through the efficiency limitations and parameter conversion errors of traditional methods, adapts to complex geological conditions, and improves the model prediction accuracy and applicability.
Smart Images
Figure CN120706248A_ABST
Abstract
Description
Technical Field
[0001] The present invention is a method for quickly calibrating sand liquefaction parameters based on particle microscopic simulation, and relates to the technical fields related to sand liquefaction, discrete element simulation and artificial intelligence. Background Art
[0002] Sand liquefaction is a geotechnical instability phenomenon in which saturated sand, under dynamic loads such as earthquakes, experiences a dramatic increase in pore water pressure, leading to a disappearance of effective stress and a loss of shear strength, resulting in a liquid-like flow. This phenomenon is a major cause of engineering disasters such as building foundation instability and underground tunnel damage. Studying the microscopic mechanism of saturated sand liquefaction plays a crucial role in explaining various macroscopic phenomena during the liquefaction process. This not only scientifically elucidates the causal mechanisms of these macroscopic phenomena but also provides guidance for site liquefaction assessment and liquefaction prevention measures in engineering projects.
[0003] The discrete element method (DEM) is an effective tool for studying the microscopic mechanisms of sand liquefaction. Its unique advantage lies in its ability to analyze the influence of particle morphology on liquefaction behavior in real time and accurately capture the dynamic mechanical response of particle aggregates throughout the loading process. For studying the macroscopic response of liquefaction, the discrete lattice spring model (DLSM) can be used to analyze the response of these constitutive models at large scales, such as in experiments and engineering, by embedding them in various liquefaction constitutive models, thereby providing validation and support for theoretical research.
[0004] However, when using constitutive models in numerical simulation methods to study the mechanism of sand liquefaction, in-situ or indoor soil dynamics tests are usually required to obtain the liquefaction constitutive parameters. On the one hand, conducting a large number of indoor tests violates the starting point of efficient and rapid numerical simulation. On the other hand, some parameters calibrated through indoor tests cannot be directly applied to numerical simulation analysis at the engineering scale and often need to be converted based on experience. In traditional research, the mechanism simulation of micro- and macroscopic liquefaction behavior and mechanical response is often carried out separately, independently and lacking correlation. There is an urgent need to establish a connection between the micro-mechanical behavior of sand liquefaction in discrete element simulation and the selection of constitutive parameters in continuous media, thereby providing support for the reasonable selection of macroscopic liquefaction parameters. Summary of the Invention
[0005] In view of the above technical problems, the present invention discloses a rapid calibration method for sand liquefaction parameters based on particle microscopic simulation. The particle-scale liquefaction mechanism and macroscopic mechanical response are revealed through discrete unit model simulation. A macroscopic liquefaction behavior inversion verification mechanism is established in combination with a discrete lattice spring model. A quantitative relationship between basic physical parameters and liquefaction parameters is established. The above parameters are input as variables into a backpropagation neural network. A nonlinear mapping between parameters is established through training of a large number of data sets. Finally, intelligent analysis of geological survey data is achieved through the integration of a professional enhanced large language model. The neural network model is called to give recommended liquefaction parameter values, forming a liquefaction parameter calibration closed loop with self-optimization capabilities.
[0006] The technical solution of the present invention:
[0007] A method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation includes the following steps:
[0008] S1. Select a certain sand material, obtain basic physical parameters, use discrete element method to carry out triaxial shear numerical simulation, and calibrate the microscopic particle parameters of the sample;
[0009] S2. Conduct microscopic liquefaction mechanism research under dynamic loading conditions to explore the changes in sand-soil particle contact and mechanical response during liquefaction;
[0010] S3. Based on the discrete lattice numerical method (DLSM), a macroscopic liquefaction constitutive model is introduced, and the optimized macroscopic liquefaction constitutive model parameters are obtained through macroscopic inversion.
[0011] S4. The basic physical parameters of the selected sand material are used as independent variables, and the optimized macroscopic liquefaction parameters obtained by inversion are used as dependent variables. The parameters are input into the back-propagation neural network model built based on PyTorch to establish a nonlinear mapping between the physical parameters and the liquefaction parameters.
[0012] S5. Deploy a professional enhanced large language model for the field of geotechnical engineering. Based on the geological exploration information and natural language instructions provided by the user, call the BPNN model to give recommended optimal liquefaction parameter values and complete the rapid calibration of sand liquefaction parameters.
[0013] Preferably, the establishment of triaxial shear numerical simulation in step S1 includes the following steps:
[0014] S11. Select representative sand gradation distribution in the study area as a reference to determine the particle gradation distribution in the numerical simulation test; calculate the soil uniformity coefficient and continuity degree through the sand particle size gradation curve;
[0015] S12. According to the above-mentioned particle gradation, a spherical particle model is generated using a random arrangement method. To prevent the particles from flying out of the wall due to excessive overlap during particle generation, the initial coordinates of all particles are multiplied by a reduction coefficient and then dispersed. After a certain time step, the velocity is reset to zero using the clam command to achieve a balanced state.
[0016] S13. Based on the constructed spherical particle model, a sample with ellipsoidal particles as the basic unit is established. First, an ellipsoidal particle template with an aspect ratio of 2:1:1 is established in the external software. Each ellipsoidal unit is filled with three spherical particles with a particle size ratio of 1:1.4:1. The same initial parameters as those for the spherical particles are used to ensure consistent unit mass distribution.
[0017] S14. Establish a servo control system; by applying speed to the wall, make the internal particle porosity and stress parameters tend to be uniform. During the cycle, continuously adjust the servo speed of the wall, and at the same time count the number of contacts between the wall and the particles at each moment, and calculate the servo parameter G of the next time step, re-obtain the wall speed of the next time step, and continuously perform cyclic servo. When the average contact stress of all walls reaches the set value, the servo is completed.
[0018] Preferably, the calibration of the microscopic particle parameters of the sample in step S1 includes the following steps: according to the on-site measured values in the research area, 3D The ellipsoid model is used in the experiment, and the particle stiffness ratios k are different. n / k s , particle elastic modulus E C , the friction coefficient between particles f c The triaxial shear test was carried out under the condition of 1 / 40°, so as to obtain the discrete element microscopic parameters suitable for the macroscopic analysis of the research area; in the study of the friction coefficient f of the particles c In the experimental groups on the influence of the internal friction angle of the material, each group of experiments was carried out under confining pressures of 100kPa, 200kPa, and 300kPa, respectively, and the experiments in other groups were all carried out under a confining pressure of 100kPa; the internal friction angle was calculated based on the strength envelope obtained from the triaxial test.
[0019] Preferably, the triaxial test in step S1 includes the following steps: using a servo system to perform servo control on the sample separately so that the sample reaches an initial confining pressure state; maintaining the confining pressure constant by controlling the movement of the four walls on the side of the model, and moving the upper and lower walls in relative directions, controlling the wall speed to 0.2 mm / s, to ensure that the sample is in quasi-static equilibrium in each time step; after the strength curve of the sample basically no longer changes, stopping loading and exporting data.
[0020] Preferably, the study of the microscopic liquefaction mechanism under dynamic load conditions in step S2 includes the following steps:
[0021] S21. Use the constant volume method to achieve undrained loading in the dynamic triaxial numerical test, so that the contact between sand particles disappears and the sample liquefies.
[0022] S22. Use the transient limit equilibrium theory to describe the liquefaction phenomenon of sand and soil and determine whether the sample has liquefied;
[0023] S23. The cyclic loading of the sample adopts the control method of equal stress amplitude. After the servo control is completed, the servo control switch is turned off and the wall is given an initial speed. When the stress amplitude of the wall reaches the critical value, the wall changes its movement direction and moves back and forth multiple times until the sample reaches the initial liquefaction state.
[0024] Preferably, the optimized macroscopic liquefaction constitutive model parameters obtained by inversion in step S3 include the following steps:
[0025] S31. In the discrete medium model, the model with the same particle gradation characteristics as the indoor test is used to determine the maximum and minimum porosity ratios under the gradation. By continuously adjusting the particle elastic modulus E C , until the maximum and minimum porosity are close to the indoor test data; PFC 3D Elastic modulus E of the medium particle C The value affects the overlap volume between particles, and thus affects the maximum and minimum porosity ratios;
[0026] S32, the porosity of the triaxial specimens at relative densities of 40% and 50% is calculated by the maximum and minimum porosity of the specimens under the particle grading, using the macroscopic mechanical parameters of Harbin sand and PFC 3D Triaxial test, calibration of the micromechanical parameters of the numerical specimen;
[0027] S33. Calibrate the liquefaction parameter C1 in the continuum model using the dynamic triaxial test through the parameter automatic inversion program.
[0028] Preferably, the calibration method of the liquefaction parameters is:
[0029] S331, establish a macroscopic discrete lattice model, set the initial value of parameter C1, make the sample very easy to liquefy, and then gradually reduce the value of C1 in the subsequent program;
[0030] S332. Perform a first cyclic loading test on the discrete lattice model until the model liquefies, and obtain a pore water pressure variation curve for this time;
[0031] S333: Import PFC based on the curve in step S332 3D The program automatically compares the time it takes for the two curves to reach initial liquefaction.
[0032] S334. If the difference between the two curves is greater than 5%, re-assign the parameter C1 by subtracting 0.05 from the original value and repeat step S333. Exit the loop until the difference is less than 5%, and derive the C1 value of the last step and the corresponding pore pressure change curve. At this time, C1 is the optimized macroscopic liquefaction parameter.
[0033] Preferably, the specific method of the nonlinear mapping of physical parameters and liquefaction parameters in step S4 is: using the basic physical parameters of the selected sand material as independent variables, and the optimized macroscopic liquefaction parameters obtained by inversion as dependent variables, collecting example data sets, performing data preprocessing through Python, dividing the processed data sets into training sets and validation sets, accounting for 70% and 30% respectively, and then inputting the back propagation neural network model built based on PyTorch, the training set is used for model fitting training and feature extraction, and the validation set is used for parameter adjustment and calibration, using deep learning technology for fitting training, self-learning data features, and establishing a nonlinear mapping of physical parameters and liquefaction parameters.
[0034] Preferably, data preprocessing is performed by Python, specifically data normalization.
[0035] Preferably, rapid calibration of liquefaction parameters is achieved based on a back-propagation neural network, and the implementation steps are as follows: obtain basic physical property parameters such as elastic modulus, Poisson's ratio, and confining pressure of sand materials as independent variables, and use the optimized macroscopic liquefaction parameters obtained by inversion as dependent variables. Collect instance data sets, perform data preprocessing through Python, and divide the processed data sets into training sets and validation sets, accounting for 70% and 30% respectively. Then, input them into a back-propagation neural network model built based on PyTorch. The training set is used for model fitting training and feature extraction, and the validation set is used for parameter adjustment and calibration. Deep learning technology is used for fitting training, self-learning data features, and establishing a nonlinear mapping between physical parameters and liquefaction parameters.
[0036] Preferably, the specific steps of step S5 are: obtaining a professional enhanced large language model suitable for the field of geotechnical engineering through model distillation training, parsing the geological exploration data input by the user, obtaining the required basic physical parameter information therefrom, and then performing semantic parsing according to the user's instructions, externally calling the BPNN model through Function Calling, and giving recommended optimal liquefaction parameter values based on the input information, which is applicable to the macro liquefaction model established by the user.
[0037] Preferably, data preprocessing is performed by Python, specifically data normalization processing
[0038] The beneficial effects of the present invention are:
[0039] With the rapid development of computer technology, artificial intelligence (AI) methods, represented by deep learning algorithms, have achieved breakthroughs in complex pattern recognition tasks such as computer vision, natural language processing, and spatiotemporal data analysis. Deep learning, with its unique model architecture, possesses exceptional processing power in multidimensional function mapping and pattern classification. This paper leverages deep learning's ability to solve complex nonlinear problems and its natural language processing capabilities to effectively connect numerical simulation studies of liquefaction mechanisms at different scales, enabling rapid calibration of sand liquefaction parameters.
[0040] (1) Through the collaborative simulation of discrete element model (DEM) and discrete lattice spring model (DLSM), the cross-scale correlation between particle-scale liquefaction mechanism and macroscopic mechanical response is realized, breaking through the limitations of traditional micro-macro separation and forming a complete liquefaction mechanism verification mechanism from particle contact behavior to engineering scale.
[0041] (2) Based on the nonlinear mapping model of deep learning, the calibration process of physical parameters and liquefaction parameters is converted into a neural network training task. Compared with the traditional experimental calibration method, the efficiency is greatly improved and the error of empirical parameter conversion is avoided, providing a standardized technical path for the rapid inversion of liquefaction parameters under complex geological conditions.
[0042] (3) The back-propagation neural network and the professional enhanced large language model are integrated to realize the closed-loop feedback of geological survey data-simulation data-engineering application. The model prediction accuracy is continuously optimized as the data accumulates, and it can dynamically adapt to the differences in regional geological characteristics, breaking through the generalization bottleneck of the traditional static constitutive model.
[0043] (4) By constructing a liquefaction parameter database covering multiple sets of cross-operating samples, the model can automatically parse the physical parameter information in the survey report and output recommended values of liquefaction parameters for multiple constitutive models, thus solving the parameter adaptation problem of traditional methods in special geological scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 Provide a technical roadmap for a rapid calibration method of sand liquefaction parameters based on particle mesoscopic simulation;
[0045] Figure 2 This is a schematic diagram of the back-propagation neural network architecture;
[0046] Figure 3 The flowchart of the natural language conversational parameter selection method based on professional enhanced LLM. DETAILED DESCRIPTION
[0047] The preferred embodiments of the present invention are described in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0048] Example 1
[0049] This example discloses a rapid calibration method for sand liquefaction parameters based on particle microscopic simulation for the implementation of a case study on the liquefaction assessment of a flood control dam foundation in Xiongan New Area. Figures 1 to 3 , including the following steps:
[0050] S1. The basic physical parameters of the sand at the project site were obtained through indoor tests. The discrete element method was used to simulate triaxial shear tests and calibrate the microscopic particle parameters of the numerical model specimens. According to the foundation enclosure project of a flood control dam in Xiongan New Area, the measured parameters of the stratum and enclosure are as follows: the natural porosity ratio of the sand is e0 = 0.73 (porosity n = 0.42), the gradation parameter is C u =1.5, C=1.1, specific gravity G s =2.65.
[0051] The establishment of the triaxial shear discrete numerical model in step S1 includes the following steps:
[0052] S11. Take the representative sand gradation distribution in the Xiongan New Area project site as a reference to determine the particle gradation distribution C in the numerical simulation test. u The soil non-uniformity coefficient and continuity degree are calculated by the particle size gradation curve of this type of sand.
[0053] S12. Select the reference gradation curve of the indoor test and the gradation curve of the numerical model, where C in the numerical model is U =1.38,C C =1.16, which is close to the value measured experimentally. The porosity was taken as the measured value of 0.35. Following the aforementioned particle grading, spherical particles with a particle size distribution of 1 to 3 mm were generated using a random permutation method. To prevent excessive overlap during particle generation, resulting in excessive initial velocity and thus flying out of the wall, the three initial coordinates of all particles were first multiplied by a coefficient less than 1, so that the particle model was generated at a distance of 0.2 mm from the wall. The particles were then dispersed. Every 50 time steps, the velocity was reset to zero using the "clam" command until equilibrium was achieved.
[0054] S13. Based on the established spherical particle model, a specimen with ellipsoidal particles as the basic unit was established to study the effects of different particle morphologies on the mechanical behavior of sand liquefaction. First, an ellipsoidal particle template with an aspect ratio of 2:1:1 was established in the external software. Each ellipsoidal unit was filled with three spherical particles with a particle size ratio of 1:1.4:1. The ellipsoidal particles and spherical particles used the same physical and mechanical parameters and initial porosity, and the ellipsoidal size distribution adopted the same particle size distribution as the sphere. After multiple adjustments to the particle size, the mass distribution of the ellipsoidal particle unit was finally highly similar to that of the spherical particle unit. The ellipsoidal unit specimen model was established according to the principles of equal volume and equal mass.
[0055] S14. When the confining pressure at the model boundary fluctuates greatly during simulation, a servo control system should be used to allow the particles in the material to continuously generate contact motion to reach the initial state required for the simulation. Since the stress of the wall must be less than the absolute value of the difference between the current stress and the target stress during the cycle, a safety factor α = 0.8 is specified. By applying speed to the wall during the cycle and continuously making dynamic adjustments, parameters such as the porosity and stress of the internal particles tend to be uniform. During the cycle, the servo speed of the wall is continuously adjusted, and the number of contacts between the wall and the particles at each moment is counted. The servo parameter G for the next time step is calculated, the wall speed for the next time step is re-obtained, and the servo is continuously performed in a cycle. When the average contact stress of all walls reaches the set value, the servo is completed.
[0056] The calibration of the microscopic particle parameters of the sample in step S1 includes the following steps: based on the field survey measured data, 3D The ellipsoid model is used in the experiment, and the particle stiffness ratios k are different. n / k s , particle elastic modulus E C , the friction coefficient between particles f c The triaxial shear test was carried out under the condition of 1 / 40°, so as to obtain the discrete element microscopic parameters suitable for the macroscopic analysis of the research area; in the study of the friction coefficient f of the particles c In the experimental groups investigating the effect of the material's internal friction angle, each group was conducted at confining pressures of 100 kPa, 200 kPa, and 300 kPa, respectively. All other groups were conducted at a confining pressure of 100 kPa. By analyzing the stress-strain curves of the specimens under different operating conditions, plotting the Mohr-Coulomb strength envelope, and calculating the equivalent internal friction angle, the mesoscopic particle parameters of the discrete element model applicable to the study area were ultimately determined: the stiffness ratio kn / ks = 2.5, the elastic modulus Ec = 80 MPa, and the friction coefficient fc = 0.35. The equivalent internal friction angle φ' was verified to be 31.5° (with an error of <1% compared to the measured 31.2°).
[0057] The triaxial shear test in step S1 includes the following steps:
[0058] Servo control is performed on each sample to make the sample reach the initial confining pressure state;
[0059] The confining pressure is kept constant by controlling the movement of the four walls on the side of the model. The upper and lower walls move in relative directions and the wall speed is controlled at 0.2 mm / s, fully ensuring that the specimen is in quasi-static equilibrium in each time step.
[0060] When the strength curve of the sample basically stops changing, stop loading and export the data.
[0061] S2. Conduct microscopic liquefaction mechanism research under dynamic loading conditions to explore the changes in sand particle contact and mechanical response during the liquefaction process.
[0062] The basic earthquake intensity of the study area is VII. The soil in the project area is medium soft soil and the site category is Class III. After correction, the peak acceleration of the seismic motion in the project area is 0.125g. The loading waveform uses the 1994 Northridge earthquake wave in California, USA. This seismic wave is suitable for Class III site soil and the peak acceleration of the seismic wave is 0.13g. The model is loaded using velocity time history. Before loading, the seismic wave is baseline corrected and filtered, that is, the high-frequency part of the waveform is filtered and the displacement time history is baseline corrected.
[0063] The numerical simulation experiment of the mesoscopic liquefaction mechanism under dynamic load conditions in step S2 includes the following steps:
[0064] S21. Using the constant volume method, the movement of the six walls is controlled during the loading process to keep the sample volume floating within a small range (volume change less than 0.5%). This achieves undrained loading in the dynamic triaxial numerical test, causing the contact between the sand particles to disappear and the sample to liquefy. The sand liquefaction process is reflected by monitoring the stress on the walls.
[0065] S22. The liquefaction judgment method of the transient limit equilibrium theory is adopted. By recording the variables such as effective stress, shear stress, and pore water pressure during the experiment, the sample is considered to have reached initial liquefaction when the effective stress drops to 0 and the pore water pressure reaches the initial confining pressure (100 kPa).
[0066] S23. The cyclic loading of the sample adopts the control method of equal stress amplitude. After the servo control is completed, the servo control switch is turned off and the wall is given an initial speed. When the stress amplitude of the wall reaches the critical value, the wall changes its direction of movement and moves back and forth multiple times until the sample reaches the initial liquefaction state. The function of recording is to calculate the variables based on the stress on the wall and the movement of the wall. The liquefaction state is reached after 12.8 seconds (6.4 cycles) of cyclic loading. The shear strain γ at the initial liquefaction is 3.2%.
[0067] S3. Based on DLSM, a discrete lattice model of macro-engineering scale is established, and the DP macro-liquefaction constitutive model is introduced. The optimized macro-liquefaction constitutive model parameters are obtained through macro-simulation inversion.
[0068] The optimized macroscopic liquefaction constitutive model parameters obtained by inversion in step S3 include the following steps:
[0069] S31. In the discrete medium model, the model with the same particle gradation characteristics as the indoor test is used to determine the maximum and minimum porosity ratios under the gradation. By continuously adjusting the particle elastic modulus E C Until the maximum and minimum porosity are close to the indoor test data (PFC 3D Elastic modulus E of the medium particle C The value affects the overlap volume between particles, and thus affects the maximum and minimum porosity ratios). To obtain the calibrated porosity: adjust Ec so that e max =0.82 / e min =0.53 (measured value 0.83 / 0.52).
[0070] S32. The porosity of the triaxial specimens at relative densities of 40% and 50% is calculated by using the maximum and minimum porosity of the specimens under the particle grading. The porosity e=0.68 is obtained at a relative density of 45%. The macroscopic mechanical parameters of Harbin sand are used, and the same PFC3D triaxial shear test as in is used to calibrate the microscopic mechanical parameters of the numerical specimens.
[0071] S33. Calibrate the liquefaction parameter C1 in the continuum model using dynamic triaxial tests through an automatic parameter inversion procedure.
[0072] The calibration method of macroscopic liquefaction parameters is:
[0073] S331. Create a stratum model of the existing lock chamber section in DLSM. The model is 136m long, 21.6m wide, and 14.5m high. Set the initial value of parameter C1. In this example, the initial value is set to 1. At this time, the sample is very likely to liquefy, with a liquefaction time of 8.2s and an error of -35.9%. The C1 value will be gradually reduced in subsequent procedures.
[0074] S332. Perform the first cyclic loading test on the discrete lattice model, and the sample liquefies. Obtain the pore water pressure change curve of the model simulation experiment.
[0075] S333. Based on the dynamic pore pressure rise curve, the particle microscopic simulation model curve is compared and analyzed. The program automatically compares the time it takes for the two curves to reach initial liquefaction. The target liquefaction time is 12.8s.
[0076] If the difference between the two curves is significant, reassign parameter C1 by subtracting 0.05 from the original value, and repeat step S33 until the difference is less than 5%. The loop then exits, and the final C1 value and the corresponding pore pressure curve are derived. At this point, C1 is the optimized macroscopic liquefaction parameter. After four iterations: C1 = 0.95 → 0.85 → 0.79 → 0.76, the final liquefaction time is 12.6 seconds (error < 2%).
[0077] The liquefaction parameters in the continuum model were calibrated using dynamic triaxial tests through the parameter automatic inversion program, and C1=0.76 was finally determined to be the optimal parameter.
[0078] S4. The basic physical parameters of the selected sand material are used as independent variables, and the optimized macro-liquefaction parameters obtained by inversion are used as dependent variables. An example data set is collected and preprocessed using Python. The processed data set is divided into a training set and a validation set, accounting for 70% and 30% respectively. The training set is then input into a backpropagation neural network model built based on PyTorch. The training set is used for model fitting training and feature extraction, and the validation set is used for parameter adjustment and calibration. Deep learning technology is used for fitting training and self-learning data features. A high-precision nonlinear mapping between geotechnical physical parameters and liquefaction parameters is established through supervised learning training.
[0079] The field data of a flood control dam project in Xiongan New Area was normalized using Python.
[0080] The results of the backpropagation network model used in this embodiment are as follows:
[0081] Input layer: Ec (MPa) = 80, v = 0.3, σ3 (kPa) = 100, e0 = 0.73, D r (%) = 45, C u =1.5, G s =2.65.
[0082] Output layer: liquefaction parameter C1
[0083] Network architecture: 7-128-64-1 (ReLU activation)
[0084] Dataset: 800 samples (including 180 samples from the Xiongan New Area project site)
[0085] Prediction result: C1 = 0.76, converted to CSR pre d = 0.22
[0086] Verification error: 4.8% (measured )
[0087] S5. Through model distillation training, a professional enhanced large language model suitable for the field of geotechnical engineering is obtained. The input geological survey data of a flood control dam in Xiongan New Area is parsed to obtain the required basic physical parameters of the sand in the area. Then, semantic parsing is performed according to user instructions, and the BPNN model is externally called through Function Calling to provide a recommended optimal liquefaction parameter value CSR based on the input information. cal =0.20, completing the rapid calibration of sand liquefaction parameters.
[0088] In this example, parameter calibration takes approximately 3 minutes, compared to 8 weeks with traditional methods. This approach is approximately 1,000 times more efficient, with model error less than 5%, compared to 15-20% with traditional methods, a 300% improvement. Furthermore, this approach implements multi-scale fusion cross-validation decision-making, significantly improving its applicability to engineering geology.
[0089] Although some preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.
[0090] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications of the present invention fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.
Claims
1. A rapid calibration method for sand liquefaction parameters based on particle microscopic simulation, characterized in that: The following steps are involved: S1. Select a certain sand material, obtain basic physical parameters, use discrete element method to carry out triaxial shear numerical simulation, and calibrate the microscopic particle parameters of the sample; S2. Conduct microscopic liquefaction mechanism research under dynamic loading conditions to explore the changes in sand-soil particle contact and mechanical response during liquefaction; S3. Based on the discrete lattice numerical method (DLSM), a macroscopic liquefaction constitutive model is introduced, and the optimized macroscopic liquefaction constitutive model parameters are obtained through macroscopic inversion. S4. The basic physical parameters of the selected sand material are used as independent variables, and the optimized macroscopic liquefaction parameters obtained by inversion are used as dependent variables. The parameters are input into the back-propagation neural network model built based on PyTorch to establish a nonlinear mapping between the physical parameters and the liquefaction parameters. S5. Deploy a professional enhanced large language model for the field of geotechnical engineering. Based on the geological exploration information and natural language instructions provided by the user, call the BPNN model to give recommended optimal liquefaction parameter values and complete the rapid calibration of sand liquefaction parameters.
2. The method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 1 is characterized in that: The establishment of triaxial shear numerical simulation in step S1 includes the following steps: S11. Select representative sand gradation distribution in the study area as a reference to determine the particle gradation distribution in the numerical simulation test; calculate the soil uniformity coefficient and continuity degree through the sand particle size gradation curve; S12. According to the above-mentioned particle gradation, a spherical particle model is generated using a random arrangement method. To prevent the particles from flying out of the wall due to excessive overlap during particle generation, the initial coordinates of all particles are multiplied by a reduction coefficient and then dispersed. After a certain time step, the velocity is reset to zero using the clam command to achieve a balanced state. S13. Based on the constructed spherical particle model, a sample with ellipsoidal particles as the basic unit is established. First, an ellipsoidal particle template with an aspect ratio of 2:1:1 is established in the external software. Each ellipsoidal unit is filled with three spherical particles with a particle size ratio of 1:1.4:
1. The same initial parameters as those for the spherical particles are used to ensure consistent unit mass distribution. S14. Establish a servo control system; by applying speed to the wall, make the internal particle porosity and stress parameters tend to be uniform. During the cycle, continuously adjust the servo speed of the wall, and at the same time count the number of contacts between the wall and the particles at each moment, and calculate the servo parameter G of the next time step, re-obtain the wall speed of the next time step, and continuously perform cyclic servo. When the average contact stress of all walls reaches the set value, the servo is completed.
3. The method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 1 is characterized in that ,The calibration of the microscopic particle parameters of the sample in step S1 includes the following steps: ,according to the field measured values in the study area, ,the PFC 3D The ellipsoid model is used in the experiment, and the particle stiffness ratios k are different. n / k s , particle elastic modulus E C , the friction coefficient between particles f c The triaxial shear test was carried out under the condition of 1 / 40°, so as to obtain the discrete element microscopic parameters suitable for the macroscopic analysis of the research area; in the study of the friction coefficient f of the particles c In the experimental groups on the influence of the internal friction angle of the material, each group of experiments was carried out under confining pressures of 100kPa, 200kPa, and 300kPa, respectively, and the experiments in other groups were all carried out under a confining pressure of 100kPa; the internal friction angle was calculated based on the strength envelope obtained from the triaxial test.
4. A method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 3, characterized in that The triaxial test in step S1 includes the following steps: using a servo system to perform servo control on the specimens respectively so that the specimens reach an initial confining pressure state; maintaining the confining pressure constant by controlling the movement of the four walls on the side of the model, and moving the upper and lower walls in relative directions, controlling the wall speed to 0.2 mm / s, to ensure that the specimen is in quasi-static equilibrium in each time step; after the strength curve of the specimen basically no longer changes, stop loading and export the data.
5. The method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 1 is characterized in that ,The study of the microscopic liquefaction mechanism under dynamic load conditions in step S2 includes the following steps: S21. Use the constant volume method to achieve undrained loading in the dynamic triaxial numerical test, so that the contact between sand particles disappears and the sample liquefies. S22. Use the transient limit equilibrium theory to describe the liquefaction phenomenon of sand and soil and determine whether the sample has liquefied; S23. The cyclic loading of the sample adopts the control method of equal stress amplitude. After the servo control is completed, the servo control switch is turned off and the wall is given an initial speed. When the stress amplitude of the wall reaches the critical value, the wall changes its movement direction and moves back and forth multiple times until the sample reaches the initial liquefaction state.
6. The method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 1 is characterized in that The optimized macroscopic liquefaction constitutive model parameters obtained by inversion in step S3 include the following steps: S31. In the discrete medium model, the model with the same particle gradation characteristics as the indoor test is used to determine the maximum and minimum porosity ratios under the gradation. By continuously adjusting the particle elastic modulus E C , until the maximum and minimum porosity are close to the indoor test data; PFC 3D Elastic modulus E of the medium particle C The value affects the overlap volume between particles, and thus affects the maximum and minimum porosity ratios; S32, the porosity of the triaxial specimens at relative densities of 40% and 50% is calculated by the maximum and minimum porosity of the specimens under the particle grading, using the macroscopic mechanical parameters of Harbin sand and PFC 3D Triaxial test, calibration of the micromechanical parameters of the numerical specimen; S33. Calibrate the liquefaction parameter C1 in the continuum model using the dynamic triaxial test through the parameter automatic inversion program.
7. A method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 6, characterized in that ,The calibration method of liquefaction parameter C1 is: S331, establish a macroscopic discrete lattice model, set the initial value of parameter C1, make the sample very easy to liquefy, and then gradually reduce the value of C1 in the subsequent program; S332. Perform a first cyclic loading test on the discrete lattice model until the model liquefies, and obtain a pore water pressure variation curve for this time; S333: Import PFC based on the curve in step S332 3D The program automatically compares the time it takes for the two curves to reach initial liquefaction. S334. If the difference between the two curves is greater than 5%, re-assign the parameter C1 by subtracting 0.05 from the original value and repeat step S333. Exit the loop until the difference is less than 5%, and derive the C1 value of the last step and the corresponding pore pressure change curve. At this time, C1 is the optimized macroscopic liquefaction parameter.
8. The method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 1 is characterized in that The specific method of the nonlinear mapping between physical parameters and liquefaction parameters in step S4 is as follows: the basic physical parameters of the selected sand material are used as independent variables, and the optimized macroscopic liquefaction parameters obtained by inversion are used as dependent variables. An example data set is collected, and data preprocessing is performed through Python. The processed data set is divided into a training set and a validation set, accounting for 70% and 30% respectively. Then, the training set is input into a backpropagation neural network model built based on PyTorch. The training set is used for model fitting training and feature extraction, and the validation set is used for parameter adjustment and calibration. Deep learning technology is used for fitting training, self-learning data features, and a nonlinear mapping between physical parameters and liquefaction parameters is established.
9. A method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 8, characterized in that ,Data preprocessing through Python is specifically data normalization.
10. The method for rapid calibration of sand liquefaction parameters based on particle microscopic simulation according to claim 1 is characterized in that ,The specific steps of step S5 are as follows: a professional enhanced large language model suitable for the ,field of geotechnical engineering is obtained through model distillation training, ,the geological survey data input by the user is parsed to obtain the required ,basic physical parameter information from it, and then semantic parsing is performed according to the user ,instructions, and the BPNN model is externally called through Function Calling, and the ,recommended optimal liquefaction parameter values are given according to the input ,information, and the macro liquefaction model established by the user.