Variable permeability water permeable pipe design optimization method based on ai simulation technology

Abstract

This invention discloses a design optimization method for variable permeability permeable pipes based on AI simulation technology, belonging to the field of intelligent design technology for geotechnical engineering pipes. The method includes acquiring original geological survey data and historical hydrological monitoring data of the engineering site, cleaning and standardizing them to form a multi-source basic dataset, constructing a three-dimensional geological and hydrological simulation model of the site including soil stratification, groundwater level fluctuations, and surrounding loads, and generating initial design parameters for the permeable pipe. The virtual model of the permeable pipe is then embedded into the simulation model, dynamic groundwater level and infiltration boundary conditions are set, and coupled numerical simulations of the seepage field and stress field are conducted. Permeability and stress deformation data at different locations and times are extracted in real time, and compared with preset drainage performance targets and durability constraints to form an evaluation result. The design parameters are iteratively adjusted using a proxy model optimized through deep reinforcement learning. This method enables the design parameters of the permeable pipe to be accurately adapted to site conditions, thereby meeting its drainage and durability performance requirements.

Description

Technical Field

[0001] This invention belongs to the field of intelligent design technology for geotechnical engineering pipes, specifically a design optimization method for variable permeability permeable pipes based on AI simulation technology. Background Technology

[0002] Current design optimization of variable permeability permeable pipes often adopts conventional geological and hydrological analysis methods, directly using original geological survey data and historical hydrological monitoring data to carry out parameter design. However, the multi-source data is not cleaned and standardized for fusion, making it impossible to form a unified and complete site basic data system.

[0003] Conventional techniques often construct simplified two-dimensional geological and hydrological analysis models, without integrating soil stratification, groundwater level fluctuation characteristics, and surrounding load information. Simulation calculations only set fixed groundwater levels and static infiltration conditions, without conducting coupled numerical calculations of seepage field and stress field, and cannot obtain real-time data on the actual permeability and stress deformation of permeable pipes at different locations and times.

[0004] Traditional permeable pipe design parameters rely on manual experience to formulate and adjust through trial calculations. The parameter correction process lacks objective and quantitative evaluation criteria, fails to integrate deep reinforcement learning algorithms with the permeable pipe design optimization process, and lacks an optimization proxy model that can autonomously iterate parameters. Parameter adjustments cannot directly correspond to preset drainage performance targets and durability constraints. The design parameters have poor adaptability to the actual geological and hydrological conditions of the engineering site, making it difficult for variable permeability permeable pipes to achieve the expected drainage effect and durability.

[0005] Existing technologies lack a high-fidelity simulation environment that can accurately reflect the dynamic performance of permeable pipes under real and complex working conditions. Furthermore, the adjustment of design parameters relies on manual experience and cannot be automatically optimized. This makes it difficult to achieve a precise and automated optimal balance between drainage performance and structural durability in the design of variable permeability permeable pipes. Therefore, this invention aims to solve the core problem of intelligently and accurately adapting permeable pipe design parameters to the specific geological and hydrological conditions of engineering sites by constructing a closed-loop system that combines a high-fidelity simulation environment with an intelligent optimization agent. Summary of the Invention

[0006] This invention aims to solve one or more of the following technical problems: How can we build a high-fidelity simulation environment that can accurately reflect complex, dynamic, and real-world working conditions? Existing technologies use unprocessed raw data and simplified models, which cannot accurately simulate the performance of permeable pipes under real-world site conditions. How can we achieve direct and quantitative benchmarking and evaluation between design parameters and drainage performance and durability constraints? Traditional methods rely on manual experience, lack objective evaluation basis, and lead to blind design adjustments. How can we achieve adaptive and automated optimization of design parameters to accurately match specific site conditions? Existing methods struggle to find the optimal solution efficiently under multiple objectives and constraints, resulting in poor adaptability between design parameters and engineering requirements. To this end, the present invention proposes a design optimization method for variable permeability permeable pipes based on AI simulation technology, including: The original geological survey data and historical hydrological monitoring data of the engineering site are obtained, and the original geological survey data and historical hydrological monitoring data are cleaned, standardized and fused to generate a multi-source basic dataset of the site. Based on the site's multi-source basic dataset, a three-dimensional geological and hydrological simulation model of the site is constructed, which includes soil layering structure, groundwater level fluctuation characteristics, and surrounding load information. Based on the preset drainage performance targets and durability constraints, an initial design parameter set for the permeable pipe is generated; In the three-dimensional geological and hydrological simulation model of the site, the virtual model of the permeable pipe defined by the initial design parameter set of the permeable pipe is implanted, and the groundwater level boundary conditions and surface infiltration conditions that change with time are set. The numerical simulation calculation of coupled seepage and stress field is started. During the numerical simulation calculation, the actual permeability and stress deformation data of the virtual model of the permeable pipe at different locations and at different times are extracted in real time. The actual permeability and stress deformation data are compared with the preset drainage performance targets and durability constraints to generate the initial evaluation results of the design parameters. Based on the initial evaluation results, a deep reinforcement learning-based optimization agent model is driven to iteratively adjust the initial design parameter set of the permeable pipe, generating an optimized design parameter set for the permeable pipe.

[0007] Furthermore, the original geological survey data and historical hydrological monitoring data are cleaned, standardized, and fused to generate a multi-source basic dataset for the site, specifically including: Identify anomalous borehole coordinates, missing soil layer descriptions, and illogical physical and mechanical indicators in the original geological exploration data, and use geostatistical interpolation methods for repair and correction. Historical hydrological monitoring data are periodically decomposed to separate long-term trend terms, seasonal periodic terms, and irregular residual terms, and obvious outliers caused by equipment failures are removed. Establish a unified spatiotemporal coordinate framework to associate and fuse the corrected geological exploration data and the processed hydrological monitoring data on the three-dimensional spatial grid nodes; For the fused three-dimensional spatial grid data, the representative soil permeability coefficient, void ratio and compression modulus of each grid cell are calculated to form a discretized site physical and mechanical property field; Integrating spatiotemporal coordinate frameworks, three-dimensional spatial grid data, and site physical and mechanical property fields, it encapsulates them into a structured multi-source basic dataset for the site.

[0008] Furthermore, the construction of the site's three-dimensional geological and hydrological simulation model, which includes soil stratification, groundwater level fluctuation characteristics, and surrounding load information, specifically includes: Spatial distribution interface and thickness variation information of different soil layers are extracted from the multi-source basic dataset of the site. A non-uniform rational B-spline surface fitting method is used to generate a continuous and smooth soil layer structure surface model. Based on the long-term trend term and seasonal periodic term decomposed from historical hydrological monitoring data, a mathematical expression for the change of groundwater level over time is constructed, and this expression is used as a dynamic boundary condition to be assigned to the bottom boundary of the soil stratification surface model. Identify existing structures and planned loads around the engineering site, simplify the loads into equivalent static or dynamic pressures, and map them onto the corresponding surface areas of the soil layer structure surface model. The unsaturated seepage theory framework and the elastoplastic constitutive model are selected to describe the transport law of water in soil and the deformation behavior of soil under stress, respectively. The soil layered structure surface model, dynamic water level boundary conditions, surrounding load mapping results, unsaturated seepage theory framework and elastoplastic constitutive model are integrated and assembled into a three-dimensional geological and hydrological simulation model for the site that can be used for coupled calculations.

[0009] Furthermore, based on preset drainage performance targets and durability constraints, the initial design parameter set for the permeable pipe is generated, specifically including: The initial design parameters for the permeable pipe include pipe diameter, perforation pattern, perforation ratio, material composition, and wall thickness. Obtain engineering design requirements documents, extract the allowable surface water accumulation depth and groundwater level control elevation under specific return period rainfall, and quantify them as drainage performance targets; From the material library and specifications, the long-term strength decay curve, corrosion resistance index and minimum safe thickness requirements of the materials used in permeable pipes are extracted to form durability constraints. Using a parametric modeling method, the pipe diameter, perforation pattern shape, perforation distribution density, and equivalent perforation diameter of the permeable pipe are defined as variable design variables. For each variable design variable, within its engineering feasible range, a set of uniformly distributed design parameter samples are generated according to the Latin hypercube sampling method; Combining drainage performance objectives and durability constraints, the generated design parameter samples are screened for feasibility, samples that clearly do not meet the constraints are removed, and the remaining samples are combined into the initial design parameter set for the permeable pipe.

[0010] Furthermore, numerical simulation calculations of coupled seepage and stress fields are initiated. During the numerical simulation calculation process, the actual permeability and stress deformation data of the virtual model of the permeable pipe at different locations and times are extracted in real time, specifically including: In the three-dimensional geological and hydrological simulation model of the site, the virtual model of the implanted permeable pipe is discretized into a finite element mesh, and its nodes are coordinated and connected with those of the surrounding soil mesh. Set the total simulation duration and time step. Within each time step, solve the unsaturated seepage control equation and the soil stress balance equation in sequence, and consider the influence of the volume force caused by the seepage field on the stress field, as well as the influence of the stress field change on the soil permeability coefficient. After each time step of the solution is completed, the velocity vector data flowing through the pipe wall and the stress tensor data of the pipe wall elements are read from the finite element mesh elements of the permeable pipe virtual model. Based on the velocity vector data and the geometric information of the orifice pattern, the local actual permeability at each orifice location under the current head difference is dynamically calculated; Based on the stress tensor data, the equivalent stress and local deformation of the permeable pipe body are calculated and summarized into stress-deformation data.

[0011] Furthermore, the comparison of actual permeability and stress-deformation data with preset drainage performance targets and durability constraints to generate initial evaluation results of design parameters specifically includes: From the drainage performance targets, the groundwater level control elevation required at a specific time is obtained, and it is spatially compared with the groundwater level distribution at the corresponding time obtained from the simulation calculation to calculate the water level control compliance rate. Based on actual permeability data, the average drainage flow per unit length of the permeable pipe virtual model was statistically analyzed during the preset high infiltration intensity period. The average drainage flow rate is compared with the theoretically required drainage flow rate calculated based on the surface water depth requirement to determine the drainage capacity satisfaction. From the stress and deformation data, identify the maximum equivalent stress and the maximum local deformation of the pipe wall that the virtual model of the permeable pipe experiences throughout the entire simulation period. The maximum equivalent stress is compared with the long-term strength of the material, and the maximum local deformation is compared with the allowable deformation limit to determine whether the durability constraint has been violated. By integrating the compliance rate of water level control, the satisfaction of drainage capacity, and the violation of durability constraints, a quantitative, multi-dimensional initial evaluation result of design parameters is generated.

[0012] Furthermore, based on the initial evaluation results, a deep reinforcement learning-based optimization surrogate model is driven to iteratively adjust the initial design parameter set of the permeable pipe, specifically including: The initial training sample set is formed by combining each design parameter sample in the initial design parameter set of the permeable pipe and its corresponding initial evaluation result. A deep reinforcement learning network is constructed, whose state space is a combination of design parameters and the current hydrological state of the site, and whose action space is the fine-tuning amount of the design parameters. The reward function is constructed based on the water level control compliance rate and drainage capacity satisfaction in the initial evaluation results, and the violation of durability constraints is penalized. The deep reinforcement learning network is pre-trained using the initial training sample set to learn the preliminary mapping relationship between the design parameters and the evaluation results. Based on the pre-trained network, the deep reinforcement learning agent explores in the simulated 3D geological and hydrological model of the site, obtains new state-action-reward samples by interacting with the environment, and continuously updates the network weights. When the fluctuation range of the reward function value is less than the pre-set convergence threshold in a series of preset iterations, training stops, and the optimal action sequence given by the deep reinforcement learning agent is applied to the initial design parameter set of the permeable pipe to generate an optimized design parameter set for the permeable pipe.

[0013] Furthermore, the step of dynamically calculating the local actual permeability at each opening location under the current head difference based on the flow velocity vector data and the geometric information of the opening pattern specifically includes: For each perforated element in the virtual model of the permeable pipe, collect the head values ​​of the surrounding soil elements and the inner wall elements of the pipe at the current time step. Calculate the head difference corresponding to the opening unit, where the head difference is the difference between the water head in the soil and the water head inside the pipe; Extract the normal velocity component perpendicular to the opening surface from the velocity vector data corresponding to the opening unit; Based on the geometric information of the opening pattern, the actual opening area represented by the opening unit is determined; Multiply the normal velocity component by the actual orifice area to obtain the instantaneous flow rate through the corresponding orifice. Then divide the instantaneous flow rate by the head difference and the orifice characteristic area to calculate the local actual permeability of the corresponding orifice location at the current time step.

[0014] Furthermore, the exploration within the simulated three-dimensional geological and hydrological model of the site, through interaction with the environment to obtain new state-action-reward samples, specifically includes: The deep reinforcement learning agent observes the current state, which includes the permeable pipe design parameters of the current iteration step, as well as the key hydrological state variables of the previous simulation read from the site's three-dimensional geological and hydrological simulation model. The agent outputs an action based on its network policy, which is an adjustment vector for the current permeable pipe design parameters; Based on the adjustment vector, the design parameters of the permeable pipe virtual model are modified, and a new, full-duration coupled seepage and stress field numerical simulation calculation is initiated in the three-dimensional geological and hydrological simulation model of the site. From the new simulation results, the updated actual permeability and stress deformation data are extracted, and the water level control compliance rate and drainage capacity satisfaction are recalculated, and the durability constraints are checked. Based on the results of the recalculation and inspection, and combined with the predefined reward function calculation formula, calculate the immediate reward obtained by the current state-action pair; The current state, the action performed, the immediate reward obtained, and the new state entered after performing the action are stored together as a new state-action-reward sample for subsequent network weight updates.

[0015] Furthermore, it also includes: The optimized set of permeable pipe design parameters was substituted into the site's three-dimensional geological and hydrological simulation model for verification simulation calculations, specifically including: From the optimized set of permeable pipe design parameters, the design parameter combination with the highest reward function value is selected as the final candidate scheme; A new virtual model of a permeable pipe is completely reconstructed based on the pipe diameter, opening pattern, opening ratio, material composition, and wall thickness parameters defined in the final candidate scheme. The new permeable pipe virtual model is implanted into the original site 3D geological and hydrological simulation model, replacing the original permeable pipe virtual model. Set a longer simulation duration or more stringent hydrological boundary conditions than in the optimization phase, and initiate the final round of coupled seepage and stress field numerical simulation calculations. After the verification simulation calculation is completed, the dynamic change process of the groundwater level, the time series of drainage flow of the permeable pipe, and the stress and strain history data of the permeable pipe throughout its entire life cycle are extracted and used as the basis for the final design verification.

[0016] Compared with the prior art, the beneficial effects of the present invention are: The original geological survey data and historical hydrological monitoring data of the engineering site are cleaned, standardized, and fused to remove anomalies and unify the format and dimensions of multi-source data, forming a standardized multi-source basic dataset for the site. This provides stable data support for subsequent simulation modeling. Based on this dataset, a three-dimensional geological and hydrological simulation model of the site is constructed, incorporating soil stratification, groundwater level fluctuation characteristics, and surrounding load information. This model can fully reproduce the geological composition and hydrological variation characteristics of the engineering site. A virtual model of permeable pipes is embedded in this model, and groundwater level boundary conditions and surface infiltration conditions that vary with time are set. Numerical simulation calculations of coupled seepage and stress fields are carried out. During the simulation process, the actual permeability and stress deformation data of the virtual model of permeable pipes at different locations and times can be extracted in real time. The acquired performance data has both spatial distribution and temporal variation characteristics, truly reflecting the working state of permeable pipes under actual site conditions. The dimensionality and accuracy of the data acquisition are superior to conventional analysis methods.

[0017] By directly comparing the real-time extracted actual permeability and stress deformation data with the preset drainage performance targets and durability constraints, an objective and quantitative initial evaluation result of the design parameters can be generated, achieving direct benchmarking between design parameters and performance indicators. Using this evaluation result as the driving basis, an optimization surrogate model based on deep reinforcement learning iteratively adjusts the initial design parameter set of the permeable pipe, enabling autonomous iterative correction of the design parameters and eliminating the need for manual trial calculations. The iterative adjustment process consistently revolves around the preset performance targets and constraints, gradually optimizing the fit between the design parameters and site conditions. The final optimized permeable pipe design parameter set is highly compatible with the geological and hydrological conditions of the engineering site, enhancing the targetedness and rationality of parameter optimization and further highlighting the match between design parameters and actual engineering needs. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the variable permeability permeable pipe design optimization method based on AI simulation technology described in this invention. Figure 2 A flowchart for generating a multi-source basic dataset for the site through the fusion of cleaning and standardization processes; Figure 3 Flowchart for constructing a three-dimensional geological and hydrological simulation model of the site; Figure 4 A three-dimensional geological and hydrological simulation model diagram; Figure 5 This is a graph showing the convergence curve of deep reinforcement learning iterations. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] See Figure 1 This invention provides a design optimization method for variable permeability permeable pipes based on AI simulation technology. The specific method includes: The process involves acquiring raw geological survey data and historical hydrological monitoring data of the engineering site, cleaning and standardizing these data, and fusing them to generate a multi-source basic dataset for the site. Based on this dataset, a three-dimensional geological and hydrological simulation model of the site is constructed, incorporating information on soil stratification, groundwater level fluctuations, and surrounding loads. An initial design parameter set for permeable pipes is generated according to preset drainage performance targets and durability constraints. A virtual model of permeable pipes, defined by the initial design parameter set, is embedded into the constructed three-dimensional geological and hydrological simulation model. Time-varying groundwater level boundary conditions and surface infiltration conditions are set, and numerical simulation calculations coupling seepage and stress fields are initiated. During this calculation, real-time data on the actual permeability and stress deformation of the virtual model of the permeable pipes at different locations and times are extracted. These data are compared with the preset drainage performance targets and durability constraints to generate an initial evaluation result of the design parameters. Based on the initial evaluation result, a deep reinforcement learning-based optimization surrogate model is driven to iteratively adjust the initial design parameter set of the permeable pipes, thereby generating an optimized design parameter set for the permeable pipes.

[0022] See Figure 2 This study identifies anomalous borehole coordinates, missing soil layer descriptions, and illogical physical and mechanical properties in the original geological survey data, and repairs and corrects them using Kriging geostatistical interpolation. Historical hydrological monitoring data is periodically decomposed to separate long-term trend terms, seasonal periodic terms, and irregular residual terms, and significant outliers caused by equipment malfunctions are removed. A unified spatiotemporal coordinate framework is established, and the corrected geological survey data and processed hydrological monitoring data are correlated and fused at three-dimensional spatial grid nodes. For the fused three-dimensional spatial grid data, the representative soil permeability coefficient, void ratio, and compression modulus of each grid cell are calculated to form a discretized site physical and mechanical property field. The spatiotemporal coordinate framework, three-dimensional spatial grid data, and site physical and mechanical property field are integrated and encapsulated into a structured multi-source basic dataset for the site.

[0023] The implementation method involves cleaning and standardizing the fusion of raw geological survey data and historical hydrological monitoring data to generate a multi-source basic dataset for the site. An example scenario is a municipal road drainage project in a soft soil area. The raw geological survey data includes the coordinates of 12 exploratory boreholes, the depth and description of each soil layer, as well as the water content, void ratio, compression modulus and permeability coefficient measured by indoor geotechnical tests. The historical hydrological monitoring data comes from the groundwater level depth data collected daily from three groundwater monitoring wells within the site over the past five years.

[0024] Anomalies were identified in the original geological survey data. One borehole's planar coordinates showed a significant spatial deviation from other boreholes, which was found to be a data entry error and corrected using coordinates from surrounding boreholes. For missing soil layer descriptions at a certain depth in two boreholes, the name and approximate depth of the soil layer were inferred and filled in using Kriging geostatistical interpolation based on the soil layer sequence of adjacent boreholes. For an illogical physical and mechanical index in a soil sample report—inconsistent data such as extremely high void ratio and extremely high compression modulus—this data was marked as invalid and, similarly, corrected using geostatistical interpolation with data from other valid samples of the same soil layer. For historical hydrological monitoring data, a time series decomposition method was used. First, the water level data of each monitoring well was decomposed into a long-term trend term, a seasonal periodic term, and an irregular residual term, thereby separating the long-term trend of a slow decline in groundwater level over the years and the seasonal periodic pattern of rising water levels during the rainy season and falling water levels during the dry season. Then, among the irregular residual terms decomposed, several obvious outliers were identified that corresponded to the equipment failure time points and deviated from the mean by more than three times the standard deviation. These outlier data were then removed.

[0025] A unified spatiotemporal coordinate framework was established, using a corner of the site as the spatial origin to create a three-dimensional rectangular coordinate system, with the time axis origin set to a fixed date. Corrected geological survey data, including borehole coordinates, soil interface elevations, and various physical and mechanical indicators, were correlated and fused with processed hydrological monitoring data—specifically, water level time series with long-term trends and seasonal cycles removed from outliers—on a three-dimensional spatial grid. This three-dimensional spatial grid was divided into 1-meter by 1-meter units horizontally and into layers ranging from 0.2 meters to 0.5 meters vertically based on the severity of soil changes. Each grid node was associated with a spatial coordinate (X, Y, Z) and a time label T (if needed). For a given time T, if there was no direct monitoring data at a particular grid node (X, Y, Z), its hydrogeological attributes were determined by interpolating and fusing neighboring borehole data and water level monitoring data in the spatiotemporal dimension.

[0026] For the fused 3D spatial grid data, it is necessary to calculate the representative soil physical and mechanical parameters of each grid cell. In practice, for a specific 3D grid cell, its soil permeability coefficient K is determined by integrating all borehole test data falling within the cell's range and the permeability characteristics inverted from hydrological data at the cell's center point. A feasible calculation method is to assign different weights to data from different sources, with direct borehole test data having a higher weight, and data inferred from water level changes serving as a supplement. The void ratio e and compression modulus Es are mainly assigned based on the soil layer category to which the grid cell belongs, as well as the statistical values ​​of test data from surrounding boreholes for that soil layer. After traversing all 3D grid cells and completing the calculations, a discretized site physical and mechanical property field is formed. This property field contains the spatial values ​​of the permeability coefficient K, void ratio e, and compression modulus Es for each grid cell. After completing the above calculations, the site's multi-source basic dataset is encapsulated. The dataset's structure comprises several core components: first, a spatiotemporal coordinate framework defining the spatial extent and grid division; second, three-dimensional spatial grid data recording the coordinates and numbers of each grid node; and third, a field of site physical and mechanical properties containing the permeability coefficient K, void ratio e, and compression modulus Es corresponding to each grid cell. This data is organized into structured arrays or database tables for direct reading and retrieval by subsequent model building functions. For example, a grid cell's data record might include its cell number, center point coordinates (X, Y, Z), soil layer number, permeability coefficient K, void ratio e, and compression modulus Es. An example for calculating the representative permeability coefficient of a grid cell is shown below:

[0027] Where: in the formula This represents the representative permeability coefficient calculated from the current grid cell. This indicates the number of boreholes affecting the current grid cell. Indicates the first The permeability coefficient test values ​​provided by each borehole in the corresponding soil layer Indicates the first The weight of each borehole data point is determined based on the distance from the borehole to the center of the grid cell. This represents a reference value for the permeability coefficient, estimated by inversion from historical water level change data at that location. This indicates the weight of the hydrological inversion data. The formula reflects the fusion process of geological exploration data and hydrological monitoring data. Ultimately, the integrated spatiotemporal coordinate framework, three-dimensional spatial grid data, and site physical and mechanical property fields are collectively packaged into a complete, structured site multi-source basic dataset file.

[0028] See Figure 3This study extracts spatial distribution interfaces and thickness variation information of different soil layers from a multi-source basic dataset of the site. A non-uniform rational B-spline surface fitting method is used to generate a continuous and smooth soil layer structure surface model. Based on the long-term trend term and seasonal periodic term decomposed from historical hydrological monitoring data, a mathematical expression for the time-varying groundwater level is constructed and used as a dynamic boundary condition for the bottom boundary of the soil layer structure surface model. Existing structures and planned loads around the engineering site are identified, and the loads are simplified into equivalent static or dynamic pressures, mapped onto the corresponding surface areas of the soil layer structure surface model. An unsaturated seepage theory framework and an elastoplastic constitutive model are selected to describe the transport law of water in the soil and the deformation behavior of the soil under stress, respectively. The soil layer structure surface model, dynamic water level boundary conditions, surrounding load mapping results, unsaturated seepage theory framework, and elastoplastic constitutive model are integrated and assembled into a three-dimensional geological and hydrological simulation model of the site that can be used for coupled calculations.

[0029] The implementation method involves constructing a three-dimensional geological and hydrological simulation model of the site, incorporating soil stratification, groundwater level fluctuation characteristics, and surrounding load information. The example scenario continues with a municipal road drainage project in a soft soil area. The site's multi-source basic dataset already includes discretized soil interface coordinates, physical and mechanical property fields, and processed five-year groundwater level monitoring data. Spatial distribution interfaces and thickness variation information of different soil layers are extracted from the site's multi-source basic dataset, specifically including the elevation values ​​of the top and bottom plates of three main soil layers—"plain fill," "silty clay," and "silty clay"—on the three-dimensional spatial grid nodes. A non-uniform rational B-spline surface fitting method is used, with the discrete elevation point set of the top or bottom plate of each soil layer as control points. By adjusting the node vectors and control point weights, a soil stratification structure surface model that can smoothly and continuously pass through or approximate these data points is generated. Each surface model is described by a mathematical parametric equation, representing the bottom plate surface of "plain fill," the top and bottom plate surfaces of "silty clay," and the top plate surface of "silty clay," respectively.

[0030] Based on the long-term trend term and seasonal periodic term derived from historical hydrological monitoring data, a mathematical expression for the change of groundwater level over time is constructed. The long-term trend term analysis shows that the groundwater level in the area is slowly decreasing at an average rate of 0.05 meters per year, while the seasonal periodic term exhibits a sinusoidal fluctuation characteristic with a one-year cycle and an amplitude of 0.8 meters. Combining these two factors, a mathematical expression is constructed to describe the change of water head at the bottom boundary of the site model over time. This expression is used as a dynamic boundary condition for the bottom boundary of the soil layered structure surface model, namely the bottom surface of the "silty clay" layer. This expression defines the water head value at any point on the bottom boundary at any simulation time t. Existing structures and planned loads around the engineering site are identified. An existing municipal road exists on one side of the site, and a multi-story building is planned to be constructed on the other side. The traffic load of the existing road is simplified into a uniformly distributed static pressure, and the foundation load of the planned building is simplified into a static pressure acting on the soil surface of a specific area through the diffusion angle principle. These equivalent static pressures are then mapped to the corresponding surface areas of the soil layer structure surface model, that is, specific locations on the top surface of the "plain fill" corresponding to the site surface.

[0031] The unsaturated seepage theory framework and the elastoplastic constitutive model were selected to describe water transport and soil deformation behavior. In practice, the unsaturated seepage theory framework uses the Richards equations as the governing equations to describe water flow in both unsaturated and saturated soils. The soil-water characteristic curves and permeability coefficient function are described using the van Genuchten model, and the required parameters are obtained from the physical and mechanical property fields in the site's multi-source basic dataset. The elastoplastic constitutive model uses the modified Cambridge model to describe the elastoplastic deformation behavior of soil under stress. The initial void ratio, compression index, and rebound index parameters required for this model are also derived from the site's multi-source basic dataset. The soil layered structure surface model generated in the preceding steps, the dynamic water level boundary conditions defined by mathematical expressions, and the peripheral load pressure results mapped onto the surface of the surface model are integrated with the unsaturated seepage theory framework and the elastoplastic constitutive model. The integration process was completed in finite element numerical simulation software. The continuous soil layer structure surface model was discretized into a three-dimensional finite element mesh. Material parameters obtained from the interpolation of physical and mechanical property fields were assigned to each element. Dynamic hydraulic boundary conditions were applied to the bottom boundary of the model. Equivalent static pressure loads were applied to the corresponding elements on the model surface. Richards equations and modified Cambridge models were specified as the governing equations and constitutive relations of the elements. Finally, the model was assembled into a three-dimensional geological and hydrological simulation model of the site that can be used to calculate the coupled seepage field and stress field.

[0032] An example of a mathematical expression used to describe dynamic water level boundary conditions is shown below:

[0033] in: Indicates at simulation time The head value at the bottom boundary of the model. This represents the reference head value at the initial moment. This represents the average annual rate of water level decline determined by the long-term trend term. This represents the amplitude of water level fluctuations determined by the seasonal periodic term. This indicates the frequency of seasonal fluctuations, with a value of 1 / year. This represents the phase angle, used to adjust the initial phase of seasonal fluctuations. This expression is directly assigned to all nodes at the bottom boundary of the site's 3D geological and hydrological simulation model, allowing the boundary head to change dynamically over time, thus more realistically reflecting the patterns revealed by historical hydrological monitoring data.

[0034] In one embodiment of the present invention, the initial design parameter set for the permeable pipe includes pipe diameter, perforation pattern, porosity, material composition, and wall thickness. Engineering design requirements documents are obtained, and the allowable surface water depth and groundwater level control elevation under specific return period rainfall are extracted and quantified as drainage performance targets. Long-term strength decay curves, corrosion resistance indices, and minimum safe thickness requirements for the materials used in the permeable pipe are extracted from a material library and specifications to form durability constraints. Using a parametric modeling method, the pipe diameter, perforation pattern shape, perforation distribution density, and equivalent perforation diameter are defined as variable design variables. For each variable design variable, within its engineering feasibility range, a set of uniformly distributed design parameter samples is generated using the Latin hypercube sampling method. Combining the drainage performance targets and durability constraints, the generated design parameter samples are screened for feasibility, eliminating samples that clearly do not meet the constraints, and combining the remaining samples into the initial design parameter set for the permeable pipe. In the three-dimensional geological and hydrological simulation model of the site, the embedded permeable pipe virtual model is discretized into a finite element mesh, ensuring that its nodes are coordinated and connected with those of the surrounding soil mesh. The total simulation duration and time step are set. Within each time step, the unsaturated seepage control equation and the soil stress balance equation are solved sequentially, considering the influence of volume forces caused by the seepage field on the stress field, and the influence of stress field changes on the soil permeability coefficient. After each time step, the velocity vector data flowing through the pipe wall and the stress tensor data of the pipe wall elements are read from the finite element mesh elements of the permeable pipe virtual model. Based on the velocity vector data, the local actual permeability at each opening location under the current head difference is dynamically calculated. Specifically, for each opening element on the permeable pipe virtual model, the head values ​​of its surrounding soil elements and the inner wall elements are collected at the current time step, and the head difference corresponding to the opening element is calculated. The normal velocity component perpendicular to the opening surface is extracted from the velocity vector data corresponding to the opening element. Based on the geometric information of the opening pattern, the actual opening area represented by the opening element is determined. Multiplying the normal velocity component by the actual orifice area yields the instantaneous flow rate through the corresponding orifice. Dividing the instantaneous flow rate by the head difference and the orifice characteristic area then calculates the local actual permeability at the corresponding orifice location in the current time step. Based on the stress tensor data, the equivalent stress and local deformation of the permeable pipe are calculated and summarized into stress-deformation data.

[0035] In practical implementation, the method involves generating the initial design parameter set for permeable pipes, initiating numerical simulation calculations, and dynamically calculating the actual local permeability. The initial design parameter set for permeable pipes includes pipe diameter, perforation pattern, porosity, material composition, and wall thickness. Engineering design requirements documents are obtained, stipulating that under a 20-year return period rainfall condition, the depth of water accumulation at any point on the ground surface must not exceed 5 cm, and the groundwater level must drop to below 1.0 meter below the surface within 24 hours after rainfall ceases. These requirements are quantified as drainage performance targets. From material libraries and industry standards, the long-term strength decay curve of high-density polyethylene (HDPE), its corrosion resistance index in pH 4-10 environments, and the minimum safe wall thickness for different burial depths are extracted to form durability constraints.

[0036] Design variables for the permeable pipe were defined using a parametric modeling method. The pipe diameter was defined as a continuous variable ranging from 200 mm to 500 mm. The perforation pattern shape was defined as a discrete variable, with options including circular, rectangular, and elliptical perforations. The perforation density was defined as the number of perforations per meter along the pipe axis, ranging from 10 to 30 per meter. The equivalent diameter of the perforation was defined as the diameter of a circular perforation with an area equal to that of a single perforation, ranging from 8 mm to 20 mm. For each variable design variable, within its engineering feasible range, a set of uniformly distributed design parameter samples was generated using the Latin hypercube sampling method. The generated design parameter samples were then subjected to feasibility screening, considering drainage performance objectives and durability constraints. For example, samples with wall thicknesses less than the minimum safe thickness required by specifications were discarded, and the remaining samples were combined to form the initial design parameter set for the permeable pipe. See Table 1: Table 1: Sample Table of Initial Design Parameters for Permeable Pipes

[0037] In practical implementation, the embedded permeable pipe virtual model is discretized into a finite element mesh within the three-dimensional geological and hydrological simulation model of the site. Solid elements are used for the permeable pipe body, and the mesh is locally refined in the perforated area of ​​the pipe wall. It is ensured that the finite element mesh nodes of the permeable pipe virtual model are coordinated and connected with the nodes of the surrounding soil mesh to achieve force and flow transmission. The total simulation duration is set to 30 days, with a time step of 1 hour. Within each time step, the numerical simulation sequentially solves the unsaturated seepage control equation describing water transport and the soil stress balance equation describing soil equilibrium. The calculation process considers the influence of seepage volume forces caused by the seepage field on the stress field, as well as the feedback effect of changes in the soil pore structure caused by stress field changes on the soil permeability coefficient. After each time step, the velocity vector data flowing through the pipe wall and the stress tensor data of the pipe wall elements are read from the finite element mesh elements of the permeable pipe virtual model.

[0038] The local actual permeability at each opening location is dynamically calculated based on velocity vector data under the current head difference. For each opening unit on the permeable pipe virtual model, the head values ​​of the adjacent soil units and the pipe wall units connected to the opening unit are collected at the current time step. The head difference corresponding to the opening unit is calculated, which is the difference between the soil head and the pipe head. The normal velocity component perpendicular to the opening surface is extracted from the velocity vector data corresponding to the opening unit. Based on the geometric information of the opening pattern, the actual opening area represented by the opening unit is determined. An example of calculating the local actual permeability is shown below:

[0039] in: This represents the local actual permeability at the current time step at this opening location. This represents the instantaneous flow rate through the corresponding opening, obtained by multiplying the normal velocity component by the actual opening area. This represents the calculated head difference. This represents the characteristic area of ​​the opening; here, the actual physical area of ​​the opening is used. Based on the stress tensor data, the equivalent stress of the permeable pipe element is calculated, and the displacement of the permeable pipe nodes is calculated to obtain the local deformation, which is then summarized into stress-deformation data.

[0040] In one embodiment of the invention, the implementation involves generating initial evaluation results of design parameters and driving iterative adjustments of an optimization proxy model based on deep reinforcement learning. From the drainage performance target, the required groundwater level control elevation at a specific time is obtained. In the example scenario, the drainage performance target requires the groundwater level to drop below 1.0 meter from the surface within 24 hours after rainfall stops, corresponding to a specific simulation time t=24 hours. The groundwater level distribution at all nodes in the entire site's three-dimensional geological and hydrological simulation model obtained from this simulation is spatially compared with the 1.0-meter control elevation. The percentage of nodes with water levels below the control elevation is calculated as the water level control compliance rate. From the actual permeability data, the average drainage flow per unit length of the permeable pipe virtual model is calculated during a preset high infiltration intensity period. The high infiltration intensity period is set to within 6 hours after the start of rainfall, based on the design rainfall process line. The drainage flow of all open units in the permeable pipe virtual model at each time step within this period is integrated and summed, then divided by the total length of the permeable pipe and the duration of the high infiltration period to calculate the average drainage flow.

[0041] After generating the initial evaluation results of the design parameters, the results drive a deep reinforcement learning-based optimization surrogate model to iteratively adjust the initial design parameter set of the permeable pipe. In this embodiment, the optimization surrogate model adopts a deep deterministic policy gradient algorithm framework suitable for continuous action spaces. This framework includes an Actor network for decision-making and a Critic network for evaluation. The model's state space is defined as a vector concatenated from the permeable pipe design parameters of the current iteration step and the normalized key hydrological state variables extracted from the previous round of simulation from the site's three-dimensional geological and hydrological simulation model. The model's action space is defined as a vector of fine-tuning parameters for each of the above design parameters, with the adjustment range of each action component pre-set based on engineering feasibility.

[0042] Both the Actor and Critic networks employ multilayer perceptrons. The Actor network receives a state vector as input, which passes through two fully connected hidden layers with 256 neurons each, processed using the ReLU activation function. The output layer then uses the tanh activation function to constrain the output to the [-1, 1] interval, linearly mapping it to the actual adjustment range of each design parameter to output a definite action. The Critic network, on the other hand, first inputs the state vector into a fully connected layer, concatenates its output with the action vector, and then processes it through two more fully connected layers. Finally, it outputs a scalar Q-value, used to evaluate the long-term expected benefit of performing a specific action in a given state.

[0043] The model is pre-trained using the initial reward data derived from the initial evaluation results. The agent in the 3D geological and hydrological simulation model interacts with the simulation environment: the agent observes the current state and outputs actions with added exploration noise through the Actor network; these actions are converted into specific design parameter adjustments, driving a new round of numerical simulation; after simulation, the immediate reward is calculated based on the new performance data, and the new state is observed. The experience of each interaction is stored in an experience replay buffer. During training, small batches of data are randomly sampled from the buffer, and the Critic network is updated by minimizing temporal difference errors, while the Actor network is updated using the policy gradient ascent method. Simultaneously, a soft update technique is used to slowly synchronize the parameters of the target network to ensure the stability of the training process. This interaction and update process is continuously repeated, and training terminates when the average reward value within the evaluation period fluctuates less than a preset convergence threshold over multiple iterations. Finally, the optimal action sequence output by the fully trained Actor network is applied to the initial parameter set to generate an optimized set of permeable pipe design parameters.

[0044] The calculated average drainage flow rate is compared with the theoretically required drainage flow rate calculated based on the surface water depth requirement. The surface water depth requirement is no more than 5 cm. The theoretically required drainage flow rate is calculated based on the design rainfall intensity, surface runoff coefficient, and catchment area through hydrological calculations. The ratio of the average drainage flow rate to the theoretically required drainage flow rate is calculated as the drainage capacity satisfaction level. A ratio greater than or equal to 1 indicates satisfaction, while a ratio less than 1 indicates non-satisfaction. From the stress-deformation data, the maximum equivalent stress borne by all pipe elements in the permeable pipe virtual model during the entire 30-day simulation period is identified, along with the maximum local deformation of the pipe wall calculated from all node displacements. The maximum equivalent stress is compared with the long-term strength of high-density polyethylene material (allowable stress considering the safety factor), and the maximum local deformation is compared with the deformation limit allowed by the specification to determine whether durability constraints are violated. Integrating the water level control compliance rate, drainage capacity satisfaction level, and violations of durability constraints, a quantitative, multi-dimensional initial evaluation result of design parameters is generated. For example, for a design parameter sample S-01, its initial evaluation results may include: water level control compliance rate of 92%, drainage capacity satisfaction of 1.15, maximum equivalent stress not exceeding allowable stress, and maximum local deformation not exceeding the limit.

[0045] The initial training sample set is composed of each design parameter sample from the initial design parameter set of the permeable pipe and its corresponding initial evaluation result. See Table 2: Table 2: Partial Data of the Initial Training Sample Set

[0046] Based on the initial training sample set described above, a deep reinforcement learning optimization agent model is constructed. This embodiment employs the Deep Deterministic Policy Gradient (DDPG) model, whose data flow and training mechanism are as follows: The state vector observed by the agent is input into the policy network, which outputs a deterministic action vector. This action vector, along with the current state, is input into the value network, which outputs the corresponding Q-value to evaluate the quality of the decision. The agent applies actions to the environment of the 3D geological and hydrological simulation model of the site. After performing a new round of numerical simulation, the environment returns a new state and an immediate reward calculated based on performance. This interaction experience is stored in an experience replay buffer. During the training phase, batches of historical data are randomly sampled from the buffer to update the value network by minimizing temporal difference errors, and the policy network by using policy gradient ascent, thereby iteratively optimizing the decision policy.

[0047] A deep reinforcement learning network is constructed, whose state space is a combination of design parameters and the current site hydrological conditions. Design parameters include pipe diameter, orifice distribution density, equivalent orifice diameter, and wall thickness. Hydrological conditions may include the average groundwater level depth and average soil saturation at the end of the previous simulation. The action space contains fine-tuning parameters, such as adjustments to the pipe diameter (+10 mm, -5 mm, etc.). The reward function is constructed based on the water level control compliance rate and drainage capacity satisfaction in the initial evaluation results, and penalizes violations of durability constraints. An example reward function is shown below:

[0048] in: Indicates an immediate reward. This indicates the water level control compliance rate. This indicates the degree to which the drainage capacity is satisfied (taken as 1 if greater than or equal to 1, otherwise its actual value is taken). This indicates the durability constraint violation index (0 if there is no violation, 1 if there is a violation). and The weighting coefficients are used to balance the relative importance of the two objectives. This is the penalty coefficient for violating constraints. The deep reinforcement learning network is pre-trained using an initial training sample set, enabling it to learn a preliminary mapping relationship from "states" (design parameters and hydrological conditions) to "evaluation results" (compliance rate, satisfaction level, violation status). Essentially, this involves learning an initial policy and value function.

[0049] In practice, based on the pre-trained network, the deep reinforcement learning agent explores within a simulated 3D geological and hydrological model of the site. The agent observes the current state, which includes the permeable pipe design parameters for the current iteration and key hydrological state variables from the previous simulation, such as average site saturation and average pore pressure, read from the 3D geological and hydrological simulation model. According to its network strategy, the agent outputs an action, which is an adjustment vector for the current permeable pipe design parameters, for example (pipe diameter: +0 mm, perforation density: +2 per meter, equivalent perforation diameter: -0.5 mm, wall thickness: +0.1 mm). Based on this adjustment vector, the design parameters of the permeable pipe virtual model are modified, and a new, 30-day coupled seepage and stress field numerical simulation calculation is initiated within the 3D geological and hydrological simulation model.

[0050] From the new simulation results, the updated actual permeability and stress deformation data are extracted, and the water level control compliance rate and drainage capacity satisfaction are recalculated, while durability constraints are checked. Based on the recalculation and check results, and combined with the predefined reward function calculation formula, the immediate reward obtained by the current state-action pair is calculated. The current state, the executed action, the obtained immediate reward, and the new state entered after executing the action are stored together as a new state-action-reward sample for subsequent network weight updates. The above interaction and update process is continued. When the fluctuation range of the reward function value is less than the pre-set convergence threshold in a preset number of consecutive iterations, training is stopped, and the final design parameter combination corresponding to the action sequence that can obtain the highest cumulative reward given by the deep reinforcement learning agent during the exploration process is applied to the initial design parameter set of the permeable pipe to generate an optimized design parameter set for the permeable pipe.

[0051] In one embodiment of the present invention, the implementation involves substituting the optimized set of permeable pipe design parameters into a three-dimensional geological and hydrological simulation model of the site for verification simulation calculations. From the optimized set of permeable pipe design parameters, the design parameter combination with the highest reward function value is selected as the final candidate scheme. Assuming that after deep reinforcement learning optimization iterations, a design sample with a pipe diameter of 320 mm, a rhomboid perforation pattern, a perforation distribution density of 20 per meter, an equivalent perforation diameter of 11 mm, and a wall thickness of 7.0 mm achieves the highest cumulative reward value, this sample is determined as the final candidate scheme.

[0052] A new virtual model of the permeable pipe is fully reconstructed based on the pipe diameter, perforation pattern, porosity, material composition, and wall thickness parameters defined in the final candidate scheme. In the computer-aided design environment, the 3D geometry of the pipe is generated based on a 320 mm diameter and 7.0 mm wall thickness. Parametric perforation modeling is performed on the pipe wall according to a rhombus pattern, a distribution density of 20 perforations / meter, and the specific rhombus dimensions calculated using an equivalent diameter of 11 mm. The material property is specified as high-density polyethylene, thus generating a completely new virtual 3D digital model of the permeable pipe with fully defined geometry and properties. This new virtual model is then integrated into the original 3D geological and hydrological simulation model of the site, replacing the original virtual model used in the optimization phase. This ensures that the new model re-establishes a coordinated connection with the nodes of the surrounding soil mesh.

[0053] To initiate the final round of coupled seepage and stress field numerical simulations, a longer simulation duration or more stringent hydrological boundary conditions than those used in the optimization phase can be established. In the optimization phase, the simulation duration might be set to 30 days to cover a major rainfall and water level recovery cycle. In the validation simulation, the simulation duration can be extended to 90 days to examine the performance stability of the permeable pipe over a longer time span, especially after multiple wet-dry cycles. Simultaneously, more stringent hydrological boundary conditions can be applied, such as increasing the design rainfall return period from 20 years to 50 years, i.e., using a more intense and longer-lasting rainfall infiltration process as the upper boundary condition of the model, to test the performance of the permeable pipe under design redundancy.

[0054] In practical implementation, verification simulation calculations are initiated. After the calculations are completed, data on the dynamic changes in the site's groundwater level throughout the entire simulation period are extracted. This data, presented as a time series, records the groundwater level values ​​at all nodes or key observation points in the model. Time series data on the drainage flow rate of the permeable pipes is extracted, recording the total flow rate or flow rate per unit length through the virtual model of the permeable pipes at each time step within the 90-day simulation period. Historical stress-strain data throughout the entire life cycle of the permeable pipes is extracted, recording the equivalent stress, principal stresses, and strain components of all elements on the permeable pipe body at each time step. These extracted data collectively constitute the basis for the final design verification and are used to prepare the design verification report. An example of comprehensively evaluating drainage efficiency during the verification phase is shown below:

[0055] in: This represents the comprehensive index of drainage efficiency during the verification simulation period. and These represent the start and end times of the simulation, respectively. Indicates time The average depth of the groundwater level in the site area. This indicates the target depth of the groundwater level to be controlled by the design. This indicates the maximum water level depth (or initial depth) allowed in the simulation. Indicates time The drainage flow rate of the permeable pipe. This index combines the degree of compliance with water level control with drainage flow rate; a higher value indicates better overall drainage performance under validation conditions. Calculated... The value can be compared with the performance of other designs in the optimization phase, but the calculation itself is part of the confirmatory analysis, not the optimization process.

[0056] See Figure 4This is a 3D geological and hydrological simulation model that comprehensively presents the site's strata, hydrology, and key engineering elements in a 3D visualization. The backfill layer on the surface of the plain fill site has a slightly higher center arc shape, simulating the micro-topographical undulations of the real site. The silty clay layer, located beneath the plain fill, is a weak soil layer and a key layer affecting foundation stability. The bearing layer at the bottom of the silty clay site forms the base of the entire model. The groundwater level, located near the silty clay layer, exhibits an arc shape matching the strata trend, simulating the groundwater level pattern of the real site. Variable permeability permeable pipes, engineering components that diagonally penetrate the strata, play a crucial role in actively regulating the site's groundwater by adjusting their own permeability. The seepage direction indicates the direction in which groundwater flows from high to low water levels under the influence of gravity and pressure, visually demonstrating the seepage path of water within the site. Surface loads simulate the loads applied by surface buildings or engineering projects, which are important external conditions that cause strata deformation and affect groundwater seepage.

[0057] See Figure 5 This is a deep reinforcement learning iterative convergence curve, visually demonstrating the performance changes of the algorithm during training. The iteration number (Episode) represents the number of training rounds, ranging from 0 to 500. In each iteration, the agent tries different control strategies in the geological and hydrological simulation model and adjusts its behavior based on feedback. The metric value is a key indicator of algorithm performance, ranging from 0 to 1, with values ​​closer to 1 indicating better performance. All three curves break through the 0.9 threshold around iteration 200 and remain stable above 0.9 in subsequent iterations, indicating that the algorithm has converged at this point and found a stable and effective control strategy. The small fluctuations in the subsequent curves are common exploratory noise in reinforcement learning, which does not affect the overall convergence trend and represents fine-tuning the strategy near the optimal solution. As training progresses, water level control, drainage capacity, and overall reward all continuously improve, eventually stabilizing at a high performance level. The small fluctuations in the later curves at high levels indicate that the algorithm strategy is stable and can maintain reliable control performance under different operating conditions.

[0058] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A design optimization method for variable permeability permeable pipes based on AI simulation technology, characterized in that, Includes the following steps: The original geological survey data and historical hydrological monitoring data of the engineering site are obtained, and the original geological survey data and historical hydrological monitoring data are cleaned, standardized and fused to generate a multi-source basic dataset of the site. Based on the site's multi-source basic dataset, a three-dimensional geological and hydrological simulation model of the site is constructed, which includes soil layering structure, groundwater level fluctuation characteristics, and surrounding load information. Based on the preset drainage performance targets and durability constraints, an initial design parameter set for the permeable pipe is generated; In the three-dimensional geological and hydrological simulation model of the site, the virtual model of the permeable pipe defined by the initial design parameter set of the permeable pipe is implanted, and the groundwater level boundary conditions and surface infiltration conditions that change with time are set. The numerical simulation calculation of coupled seepage and stress field is started. During the numerical simulation calculation, the actual permeability and stress deformation data of the virtual model of the permeable pipe at different locations and at different times are extracted in real time. The actual permeability and stress deformation data are compared with the preset drainage performance targets and durability constraints to generate the initial evaluation results of the design parameters. Based on the initial evaluation results, a deep reinforcement learning-based optimization agent model is driven to iteratively adjust the initial design parameter set of the permeable pipe, generating an optimized design parameter set for the permeable pipe.

2. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 1, characterized in that, The original geological survey data and historical hydrological monitoring data are cleaned, standardized, and fused to generate a multi-source basic dataset for the site, specifically including: Identify anomalous borehole coordinates, missing soil layer descriptions, and illogical physical and mechanical indicators in the original geological exploration data, and use geostatistical interpolation methods for repair and correction. Historical hydrological monitoring data are periodically decomposed to separate long-term trend terms, seasonal periodic terms, and irregular residual terms, and obvious outliers caused by equipment failures are removed. Establish a unified spatiotemporal coordinate framework to associate and fuse the corrected geological exploration data and the processed hydrological monitoring data on the three-dimensional spatial grid nodes; For the fused three-dimensional spatial grid data, the representative soil permeability coefficient, void ratio and compression modulus of each grid cell are calculated to form a discretized site physical and mechanical property field; Integrating spatiotemporal coordinate frameworks, three-dimensional spatial grid data, and site physical and mechanical property fields, it encapsulates them into a structured multi-source basic dataset for the site.

3. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 2, characterized in that, The construction of the site's three-dimensional geological and hydrological simulation model, which includes soil stratification, groundwater level fluctuation characteristics, and surrounding load information, specifically includes: Spatial distribution interface and thickness variation information of different soil layers are extracted from the multi-source basic dataset of the site. A non-uniform rational B-spline surface fitting method is used to generate a continuous and smooth soil layer structure surface model. Based on the long-term trend term and seasonal periodic term decomposed from historical hydrological monitoring data, a mathematical expression for the change of groundwater level over time is constructed, and this expression is used as a dynamic boundary condition to be assigned to the bottom boundary of the soil stratification surface model. Identify existing structures and planned loads around the engineering site, simplify the loads into equivalent static or dynamic pressures, and map them onto the corresponding surface areas of the soil layer structure surface model. The unsaturated seepage theory framework and the elastoplastic constitutive model are selected to describe the transport law of water in soil and the deformation behavior of soil under stress, respectively. The soil layered structure surface model, dynamic water level boundary conditions, surrounding load mapping results, unsaturated seepage theory framework and elastoplastic constitutive model are integrated and assembled into a three-dimensional geological and hydrological simulation model for the site that can be used for coupled calculations.

4. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 3, characterized in that, Based on preset drainage performance targets and durability constraints, an initial design parameter set for the permeable pipe is generated, specifically including: The initial design parameters for the permeable pipe include pipe diameter, perforation pattern, perforation ratio, material composition, and wall thickness. Obtain engineering design requirements documents, extract the allowable surface water accumulation depth and groundwater level control elevation under specific return period rainfall, and quantify them as drainage performance targets; From the material library and specifications, the long-term strength decay curve, corrosion resistance index and minimum safe thickness requirements of the materials used in permeable pipes are extracted to form durability constraints. Using a parametric modeling method, the pipe diameter, perforation pattern shape, perforation distribution density, and equivalent perforation diameter of the permeable pipe are defined as variable design variables. For each variable design variable, within its engineering feasible range, a set of uniformly distributed design parameter samples are generated according to the Latin hypercube sampling method; Combining drainage performance objectives and durability constraints, the generated design parameter samples are screened for feasibility, samples that clearly do not meet the constraints are removed, and the remaining samples are combined into the initial design parameter set for the permeable pipe.

5. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 4, characterized in that, Initiate numerical simulation calculations of coupled seepage and stress fields. During the numerical simulation, extract real-time data on actual permeability and stress deformation of the virtual model of the permeable pipe at different locations and times. Specifically, this includes: In the three-dimensional geological and hydrological simulation model of the site, the virtual model of the implanted permeable pipe is discretized into a finite element mesh, and its nodes are coordinated and connected with those of the surrounding soil mesh. Set the total simulation duration and time step. Within each time step, solve the unsaturated seepage control equation and the soil stress balance equation in sequence, and consider the influence of the volume force caused by the seepage field on the stress field, as well as the influence of the stress field change on the soil permeability coefficient. After each time step of the solution is completed, the velocity vector data flowing through the pipe wall and the stress tensor data of the pipe wall elements are read from the finite element mesh elements of the permeable pipe virtual model. Based on the velocity vector data and the geometric information of the orifice pattern, the local actual permeability at each orifice location under the current head difference is dynamically calculated; Based on the stress tensor data, the equivalent stress and local deformation of the permeable pipe body are calculated and summarized into stress-deformation data.

6. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 5, characterized in that, The process of comparing actual permeability and stress-deformation data with preset drainage performance targets and durability constraints to generate initial evaluation results of design parameters specifically includes: From the drainage performance targets, the groundwater level control elevation required at a specific time is obtained, and it is spatially compared with the groundwater level distribution at the corresponding time obtained from the simulation calculation to calculate the water level control compliance rate. Based on actual permeability data, the average drainage flow per unit length of the permeable pipe virtual model was statistically analyzed during the preset high infiltration intensity period. The average drainage flow rate is compared with the theoretically required drainage flow rate calculated based on the surface water depth requirement to determine the drainage capacity satisfaction. From the stress and deformation data, identify the maximum equivalent stress and the maximum local deformation of the pipe wall that the virtual model of the permeable pipe experiences throughout the entire simulation period. The maximum equivalent stress is compared with the long-term strength of the material, and the maximum local deformation is compared with the allowable deformation limit to determine whether the durability constraint has been violated. By integrating the compliance rate of water level control, the satisfaction of drainage capacity, and the violation of durability constraints, a quantitative, multi-dimensional initial evaluation result of design parameters is generated.

7. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 6, characterized in that, Based on the initial evaluation results, a deep reinforcement learning-based optimization surrogate model is driven to iteratively adjust the initial design parameter set of the permeable pipe, specifically including: The initial training sample set is formed by combining each design parameter sample in the initial design parameter set of the permeable pipe and its corresponding initial evaluation result. A deep reinforcement learning network is constructed, whose state space is a combination of design parameters and the current hydrological state of the site, and whose action space is the fine-tuning amount of the design parameters. The reward function is constructed based on the water level control compliance rate and drainage capacity satisfaction in the initial evaluation results, and the violation of durability constraints is penalized. The deep reinforcement learning network is pre-trained using the initial training sample set to learn the preliminary mapping relationship between the design parameters and the evaluation results. Based on the pre-trained network, the deep reinforcement learning agent explores in the simulated 3D geological and hydrological model of the site, obtains new state-action-reward samples by interacting with the environment, and continuously updates the network weights. When the fluctuation range of the reward function value is less than the pre-set convergence threshold in a series of preset iterations, training stops, and the optimal action sequence given by the deep reinforcement learning agent is applied to the initial design parameter set of the permeable pipe to generate an optimized design parameter set for the permeable pipe.

8. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 7, characterized in that, The process of dynamically calculating the local actual permeability at each opening location under the current head difference based on flow velocity vector data and the geometric information of the opening pattern specifically includes: For each perforated element in the virtual model of the permeable pipe, collect the head values ​​of the surrounding soil elements and the inner wall elements of the pipe at the current time step. Calculate the head difference corresponding to the opening unit, where the head difference is the difference between the water head in the soil and the water head inside the pipe; Extract the normal velocity component perpendicular to the opening surface from the velocity vector data corresponding to the opening unit; Based on the geometric information of the opening pattern, the actual opening area represented by the opening unit is determined; Multiply the normal velocity component by the actual orifice area to obtain the instantaneous flow rate through the corresponding orifice. Then divide the instantaneous flow rate by the head difference and the orifice characteristic area to calculate the local actual permeability of the corresponding orifice location at the current time step.

9. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 8, characterized in that, The exploration within the simulated 3D geological and hydrological model of the site, through interaction with the environment to obtain new state-action-reward samples, specifically includes: The deep reinforcement learning agent observes the current state, which includes the permeable pipe design parameters of the current iteration step, as well as the key hydrological state variables of the previous simulation read from the site's three-dimensional geological and hydrological simulation model. The agent outputs an action based on its network policy, which is an adjustment vector for the current permeable pipe design parameters; Based on the adjustment vector, the design parameters of the permeable pipe virtual model are modified, and a new, full-duration coupled seepage and stress field numerical simulation calculation is initiated in the three-dimensional geological and hydrological simulation model of the site. From the new simulation results, the updated actual permeability and stress deformation data are extracted, and the water level control compliance rate and drainage capacity satisfaction are recalculated, and the durability constraints are checked. Based on the results of the recalculation and inspection, and combined with the predefined reward function calculation formula, calculate the immediate reward obtained by the current state-action pair; The current state, the action performed, the immediate reward obtained, and the new state entered after performing the action are stored together as a new state-action-reward sample for subsequent network weight updates.

10. The variable permeability permeable pipe design optimization method based on AI simulation technology according to claim 9, characterized in that, Also includes: The optimized set of permeable pipe design parameters was substituted into the site's three-dimensional geological and hydrological simulation model for verification simulation calculations, specifically including: From the optimized set of permeable pipe design parameters, the design parameter combination with the highest reward function value is selected as the final candidate scheme; A new virtual model of a permeable pipe is completely reconstructed based on the pipe diameter, opening pattern, opening ratio, material composition, and wall thickness parameters defined in the final candidate scheme. The new permeable pipe virtual model is implanted into the original site 3D geological and hydrological simulation model, replacing the original permeable pipe virtual model. Set a longer simulation duration or more stringent hydrological boundary conditions than in the optimization phase, and initiate the final round of coupled seepage and stress field numerical simulation calculations. After the verification simulation calculation is completed, the dynamic change process of the groundwater level, the time series of drainage flow of the permeable pipe, and the stress and strain history data of the permeable pipe throughout its entire life cycle are extracted and used as the basis for the final design verification.

Images