Bionic driving soil infiltration process simulation optimization method

Abstract

The invention discloses a bionic-driven soil infiltration process simulation optimization method, which comprises the following steps: preprocessing and standardizing multi-source data, and enabling a root system voxel unit to correspond to a hydraulic response of a specific time step through a time-space registration unified coordinate system; a model architecture fusing root system features and infiltration dynamics is constructed, and a cross-modal fusion cross attention mechanism module realizes bidirectional mapping of spatial features and time features and embeds physical constraints; training and optimizing the model by adopting a staged course learning strategy, adjusting the learning rate through a physical perception optimizer, and improving the generalization ability of the model in combination with importance sampling and adversarial training; and outputting a priority flow prediction result based on the trained model, and analyzing the prediction result from microcosmic, mesoscopic and macroscopic scales through an interpreter. Therefore, the method has the advantages of high prediction precision, strong applicability, low parameter acquisition cost, introduction of a dynamic mechanism, physical rationality and the like.

Description

Technical Field

[0001] This invention relates to the field of machine learning, specifically to a biomimetic-driven method for simulating and optimizing soil infiltration processes. Background Technology

[0002] Soil water infiltration is a key link in the terrestrial hydrological cycle, directly affecting groundwater recharge, agricultural water use efficiency, and pollutant migration processes. Its research is of great significance for improving agricultural irrigation efficiency, scientific water resource management, non-point source pollution control, and climate change response assessment.

[0003] Currently, research methods for soil water infiltration mainly fall into three categories: experimental observation, numerical simulation, and remote sensing inversion. In terms of experimental observation, in-situ testing commonly uses single-ring and double-ring infiltration tests (with double-ring tests reducing lateral leakage errors), while indoor methods often employ soil column tests or automatic weighing methods, supplemented by sensors such as TDR to monitor profile water dynamics. Regarding theoretical simulation, traditional infiltration theories are based on the assumption of a homogeneous medium, grounded in Darcy's law, and centered on the Richards equation derived from the fluid continuity equation, forming classic infiltration models such as Green-Ampt, Philip, and Kostiacov. Unsaturated areas are often simulated using numerical platforms such as HYDRUS and SWAP. In recent years, remote sensing and data-driven methods have become hot topics, utilizing remote sensing products such as NDVI and surface temperature combined with machine learning models to achieve regional-scale infiltration estimation. Multi-source data fusion, high-resolution observation technologies, and research on infiltration response under extreme climates have also provided new directions for improving accuracy and applicability.

[0004] However, preferential flow in soil water infiltration presents significant challenges to research. Preferential flow refers to the process by which water bypasses the soil matrix and rapidly moves downward along highly conductive channels such as cracks, root canals, or macropores. It is characterized by its suddenness and spatial heterogeneity, accounting for approximately 11%–85% of total infiltration, and has a significant impact on surface runoff, groundwater recharge, and soil moisture redistribution. Under field conditions, the pore channels formed by soil wetting and drying cycles, animal activity, plant root growth, freeze-thaw cycles, chemical dissolution, and human cultivation can cause the saturated hydraulic conductivity of undisturbed soil to be 1–3 orders of magnitude higher than that of undisturbed soil. Current methods for studying preferential flow include experimental observation (such as staining tracing), numerical simulation (such as dual-pore or dual-conduction models), structural reconstruction (such as X-ray CT), and data-driven analysis. However, there are still significant limitations: experimental observations are mostly endpoint-based staining tracings, which are difficult to capture dynamic spatiotemporal evolution; point monitoring is limited to small scales; long-term continuous field observations are costly and difficult to implement in complex terrains; and they do not adequately consider structural factors such as biological pores. Traditional models are based on the assumption of homogeneous media, and Richards equations and derived models cannot simulate the discontinuous and sudden characteristics of preferential flow. Dual-pore / dual-conduction models rely on high-requirement parameters, lack experimental evidence, have high uncertainty, and do not fully incorporate the influence of root systems on fracture systems. At the same time, the models are scale-dependent, and small-scale parameters are difficult to extend to large scales. Summary of the Invention

[0005] The present invention aims to at least partially solve the technical problems in the above-mentioned technologies.

[0006] Therefore, this invention discloses a biomimetic-driven method for simulating and optimizing soil infiltration processes, comprising the following steps:

[0007] S1: Preprocess and standardize the multi-source data, which includes three-dimensional root structure data and infiltration time series data. By registering a unified coordinate system in time and space, the root voxel unit corresponds to the hydraulic response of a specific time step.

[0008] S2: Construct a model architecture that integrates root features and infiltration dynamics. The model architecture includes a root feature extraction module, a temporal processing module, and a cross-modal fusion cross-attention mechanism module. The root feature extraction module uses a 3D convolutional neural network to extract root topological features. The temporal processing module combines a long short-term memory network with time coding to process the infiltration dynamics. The cross-modal fusion cross-attention mechanism module realizes bidirectional mapping between spatial and temporal features and embeds physical constraints.

[0009] S3: A phased learning strategy is adopted to train and optimize the model. The learning rate is adjusted through a physical perception optimizer, and the model's generalization ability is improved by combining importance sampling and adversarial training.

[0010] S4: Based on the trained model, output the priority flow prediction results and analyze the prediction results from micro, meso and macro scales through an interpreter.

[0011] The soil infiltration framework model that replicates preferential flow by incorporating a root system model, as disclosed in this invention, has at least the following beneficial effects:

[0012] (1) Improve prediction accuracy and break through the assumption of homogeneous medium in traditional models. Realize root-specific flow channel network modeling. By integrating three-dimensional root structure data and time-series infiltration dynamic data, and combining 3D-CNN to extract root features, LSTM to process time-series information and cross-modal fusion mechanism, accurately capture the suddenness and spatial heterogeneity of priority flow. Solve the problem of low simulation accuracy and easy underestimation of flow rate in traditional models. The model performance index (RMSE=4.51, MAE=0.64, R²=0.89) verifies its high accuracy.

[0013] (2) Enhanced applicability, applicable to complex underlying surface topography, can simulate meteorological-soil-vegetation continuum for different soil types, vegetation cover, landforms and climate characteristics, overcomes the limitations of traditional models at small scale and scale dependence, can be extended to field, watershed and even regional scale, and can predict the dynamic changes of infiltration rate during different vegetation phenological periods, thus expanding the application scenarios.

[0014] (3) Reduce the cost of parameter acquisition. Through multi-source data preprocessing and standardization process, combined with physical constraint mechanisms (such as the Poiseuille equation and Richards equation verification), the dependence on high-cost measured parameters is reduced, and the uncertainty caused by indirect fitting and inversion of parameters in models such as dual-pore is avoided. In addition, the use of intelligent dual-ring infiltration equipment and standardized root sampling and measurement process in the experimental design reduces the cost of data acquisition.

[0015] (4) Revealing the dynamic mechanism, introducing time evolution operators to establish machine learning representation of transient flow, solving the problem that static parameter system is difficult to describe dynamic infiltration process, can capture the whole stage evolution of preferential flow from initial infiltration, unsteady flow to quasi-steady flow, and at the same time, through the SHP interpreter, analyze the root configuration regulation mechanism of infiltration from micro, meso and macro scales, providing key technical means for the study of soil water infiltration performance, groundwater runoff generation mechanism, and runoff generation.

[0016] (5) To ensure physical rationality, a triple physical constraint mechanism including mass conservation equation, non-negativity constraint and boundary conditions is designed. Combined with the differential-algebraic equation verification module, the model output is ensured to conform to the laws of fluid mechanics, breaking the dilemma of the black box model being separated from physical laws and improving the reliability and interpretability of the model.

[0017] In addition, the soil infiltration framework model construction that replicates preferential flow by incorporating root system model disclosed in this invention may also have the following additional technical features:

[0018] In one embodiment of the present invention, step S1, the preprocessing of the three-dimensional root system structure data includes the following steps:

[0019] S1.1: Mesh repair is performed on the root system model to fill topological defects, and the Laplacian smoothing algorithm is used to eliminate surface noise;

[0020] S1.2: Perform anisotropic voxelization treatment, maintain the first resolution in the horizontal plane, and set the second resolution along the root growth direction;

[0021] S1.3: The improved MarchingCubes algorithm is applied to generate a voxel model with pore connectivity, and a multi-channel feature tensor containing structural density, hydraulic conductivity and biological age is constructed.

[0022] In one embodiment of the present invention, step S1, the preprocessing of the infiltration time series data includes the following steps:

[0023] S1.4: Perform wavelet denoising on the original signal and expand the features based on the Richards equation to obtain derived variables;

[0024] S1.5: Perform non-stationarity testing and segmented standardization, and align the time based on the arrival time of the initial wetting peak;

[0025] S1.6: Construct a sliding time window matrix to complete the standardization processing of time series data.

[0026] In one embodiment of the present invention, step S2, the construction of the root feature extraction module includes the following steps:

[0027] S2.1: An improved 3D-CNN architecture is adopted as the encoder. The encoder includes five stages of downsampling. Each stage sets up parallel paths of standard convolution and anisotropic convolution, and features are fused through an attention gating mechanism.

[0028] S2.2: The decoder uses deformable convolution instead of conventional upsampling to adapt to the irregular growth morphology of the root system;

[0029] S2.3: Introduce a root continuity loss function and use Laplacian smoothing constraints to ensure topological consistency between the feature map and the real root anatomy.

[0030] In one embodiment of the present invention, step S2, the construction of the timing processing module includes the following steps:

[0031] S2.4: Combining Long Short-Term Memory Network with Time Encoding Module;

[0032] S2.5: Set up a bidirectional gating unit and a dynamic time warping algorithm. The bottom layer processes the original signal sequence, and the top layer receives the geometric features output by the root feature extraction module as the initial hidden state.

[0033] S2.6: Configure a fast circulation layer and a slow circulation layer, which run at a first time step and a second time step, respectively, to capture infiltration dynamics at different scales.

[0034] In one embodiment of the present invention, step S2, the construction of the cross-modal fusion cross-attention mechanism module includes the following steps:

[0035] S2.7: Introduce spatial location coding and time coding;

[0036] S2.8: Construct a bidirectional attention flow from space to time and from time to space;

[0037] S2.9: Embedded physical constraint system, including feature regularization based on the Poiseuille equation, loss function constraints with mass conservation and energy dissipation terms, and post-processing verification based on the Richards equation.

[0038] In one embodiment of the present invention, step S3, the phased course learning strategy includes the following steps:

[0039] S3.1: In the first stage, the parameters of the root feature extraction module are locked to optimize the modeling capability of the time series processing module for dynamic processes;

[0040] S3.2: The second stage fixes the parameters of the temporal encoder to improve the accuracy of root feature extraction;

[0041] S3.3: The third stage involves end-to-end joint fine-tuning, using the physically-aware AdamW variant optimizer, whose learning rate is automatically adjusted based on the kinetic gradient of the current batch of data.

[0042] In one embodiment of the present invention, step S4, the analysis process of the interpreter includes the following steps:

[0043] S4.1: At the microscale, the Monte Carlo path integral method is used to calculate the Shapley value of each voxel, and root physiological constraints are introduced to ensure biological rationality;

[0044] S4.2: At the mesoscale, key root system parameters are identified through a hierarchical attribution algorithm, and the joint contribution and independent influence of each subsystem are calculated.

[0045] S4.3: A time-dependent explanatory model is constructed at the macro scale. Through dynamic weight decomposition, the differences in the regulation of the infiltration process by different root configurations are revealed, and the distinguishing features of the initial infiltration, unsteady flow and quasi-steady flow stages are identified.

[0046] Additional features and advantages of this invention will be set forth in the description which follows, or may be learned by practicing the invention. Attached Figure Description

[0047] The technical solution and beneficial effects of the present invention will become apparent and readily understood from the following description in conjunction with the accompanying drawings, wherein:

[0048] Figure 1 This is a technical architecture diagram of the biomimetic-driven soil infiltration process simulation and optimization method of the present invention.

[0049] Figure 2 This is an experimental design diagram of the biomimetic-driven soil infiltration process simulation and optimization method of the present invention;

[0050] Figure 3 This is a schematic diagram of the fitting curves of seven model parameters for the biomimetic-driven soil infiltration process simulation and optimization method of the present invention.

[0051] Figure 4 This is a schematic diagram of the Horton basic model results of the biomimetic-driven soil infiltration process simulation and optimization method of the present invention;

[0052] Figure 5 This is a schematic diagram of the preferential flow model results of the biomimetic-driven soil infiltration process simulation optimization method of the present invention. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0054] The biomimetic-driven soil infiltration process simulation and optimization method disclosed in this invention will now be described with reference to the accompanying drawings.

[0055] like Figure 1 and Figure 2 As shown, a biomimetic-driven method for simulating and optimizing soil infiltration processes includes the following steps:

[0056] S1: Preprocess and standardize the multi-source data, which includes three-dimensional root structure data and infiltration time series data. By registering a unified coordinate system in time and space, the root voxel unit corresponds to the hydraulic response of a specific time step.

[0057] S2: Construct a model architecture that integrates root features and infiltration dynamics. The model architecture includes a root feature extraction module, a temporal processing module, and a cross-modal fusion cross-attention mechanism module. The root feature extraction module uses a 3D convolutional neural network to extract root topological features. The temporal processing module combines a long short-term memory network with time coding to process the infiltration dynamics. The cross-modal fusion cross-attention mechanism module realizes bidirectional mapping between spatial and temporal features and embeds physical constraints.

[0058] S3: A phased learning strategy is adopted to train and optimize the model. The learning rate is adjusted through a physical perception optimizer, and the model's generalization ability is improved by combining importance sampling and adversarial training.

[0059] S4: Based on the trained model, output the priority flow prediction results and analyze the prediction results from micro, meso and macro scales through an interpreter.

[0060] In step S1, the preprocessing of the three-dimensional root system structure data includes the following steps:

[0061] S1.1: Mesh repair is performed on the root system model to fill topological defects, and the Laplacian smoothing algorithm is used to eliminate surface noise;

[0062] S1.2: Perform anisotropic voxelization treatment, maintain the first resolution in the horizontal plane, and set the second resolution along the root growth direction;

[0063] S1.3: The improved MarchingCubes algorithm is applied to generate a voxel model with pore connectivity, and a multi-channel feature tensor containing structural density, hydraulic conductivity and biological age is constructed.

[0064] In step S1, the preprocessing of the infiltration time series data includes the following steps:

[0065] S1.4: Perform wavelet denoising on the original signal and expand the features based on the Richards equation to obtain derived variables;

[0066] S1.5: Perform non-stationarity testing and segmented standardization, and align the time based on the arrival time of the initial wetting peak;

[0067] S1.6: Construct a sliding time window matrix to complete the standardization processing of time series data.

[0068] In step S2, the construction of the root feature extraction module includes the following steps:

[0069] S2.1: An improved 3D-CNN architecture is adopted as the encoder. The encoder contains five stages of downsampling. Each stage sets up parallel paths of standard convolution and anisotropic convolution, and features are fused through an attention gating mechanism.

[0070] S2.2: The decoder uses deformable convolution instead of conventional upsampling to adapt to the irregular growth morphology of the root system;

[0071] S2.3: Introduce a root continuity loss function and use Laplacian smoothing constraints to ensure topological consistency between the feature map and the real root anatomy.

[0072] In step S2, the construction of the timing processing module includes the following steps:

[0073] S2.4: Combining Long Short-Term Memory Network with Time Encoding Module;

[0074] S2.5: Set up a bidirectional gating unit and a dynamic time warping algorithm. The bottom layer processes the original signal sequence, and the top layer receives the geometric features output by the root feature extraction module as the initial hidden state.

[0075] S2.6: Configure a fast circulation layer and a slow circulation layer, which run at a first time step and a second time step, respectively, to capture infiltration dynamics at different scales.

[0076] In step S2, the construction of the cross-modal fusion cross-attention mechanism module includes the following steps:

[0077] S2.7: Introduce spatial location coding and time coding;

[0078] S2.8: Construct a bidirectional attention flow from space to time and from time to space;

[0079] S2.9: Embedded physical constraint system, including feature regularization based on the Poiseuille equation, loss function constraints with mass conservation and energy dissipation terms, and post-processing verification based on the Richards equation.

[0080] In step S3, the phased course learning strategy includes the following steps:

[0081] S3.1: In the first stage, the parameters of the root feature extraction module are locked to optimize the modeling capability of the time series processing module for dynamic processes;

[0082] S3.2: The second stage fixes the parameters of the temporal encoder to improve the accuracy of root feature extraction;

[0083] S3.3: The third stage involves end-to-end joint fine-tuning, using the physically-aware AdamW variant optimizer, whose learning rate is automatically adjusted based on the kinetic gradient of the current batch of data.

[0084] In step S4, the interpreter's analysis process includes the following steps:

[0085] S4.1: At the microscale, the Monte Carlo path integral method is used to calculate the Shapley value of each voxel, and root physiological constraints are introduced to ensure biological rationality;

[0086] S4.2: At the mesoscale, key root system parameters are identified through a hierarchical attribution algorithm, and the joint contribution and independent influence of each subsystem are calculated.

[0087] S4.3: A time-dependent explanatory model is constructed at the macro scale. Through dynamic weight decomposition, the differences in the regulation of the infiltration process by different root configurations are revealed, and the distinguishing features of the initial infiltration, unsteady flow and quasi-steady flow stages are identified.

[0088] Specifically:

[0089] For the multi-source data preprocessing and standardization process, the raw data input is divided into two main channels: three-dimensional root structure data and time-series dynamic data. For the root model generated by OpenSimRoot, the preprocessing process includes six key steps: First, mesh repair is performed to fill the topological defects in the digitization process, and the Laplacian smoothing algorithm is used to eliminate surface noise; then, anisotropic voxelization is performed, maintaining a resolution of 200 μm in the XY plane while setting a resolution of 300 μm along the root growth direction to capture longitudinal expansion features; then, the improved MarchingCubes algorithm is applied to generate a voxel model with guaranteed pore connectivity; finally, a multi-channel feature tensor is constructed, including a structure density channel (local voxel occupancy), a hydraulic conductivity channel (Hagen-Poiseuille calculation based on diameter distribution), and a biological age channel (colored according to growth time). For infiltration time series data, a five-stage processing procedure is performed: wavelet denoising of the original signal, feature expansion based on the Richards equation (deriving derived variables such as equivalent hydraulic conductivity), non-stationarity testing and piecewise standardization, time alignment processing (using the arrival time of the initial wetting peak as the reference point), and construction of a sliding time window matrix. The two types of data achieve coordinate system unification through spatiotemporal registration units, ensuring that each root voxel unit corresponds to the hydraulic response at a specific time step.

[0090] In the systematic design of the technical architecture in this embodiment, the root feature extraction module adopts the EfficientNet3D architecture, an improvement on 3D-CNN. By introducing a channel-space dual attention mechanism, it maintains 90.2% accuracy in root topological feature extraction while reducing the number of parameters to 35% of the traditional 3DResNet. Its encoder contains five downsampling stages, each employing a parallel path of 3×3×3 standard convolutions and 1×3×3 anisotropic convolutions, dynamically fusing different features through an attention gating mechanism. The decoder uses deformable convolutions instead of conventional upsampling to adapt to the irregular growth morphology of the root system. A specially designed root continuity loss function applies a Laplacian smoothing constraint to the feature space, ensuring that the feature maps learned by the network maintain topological consistency with the real root anatomy.

[0091] It should be noted that:

[0092] Channel attention components are specifically as follows Where σ is the Sigmoid activation function, δ is the ReLU activation function, GAP is the global average pooling, W1∈RC / r×C, W2∈RC×C / r (compression ratio r=8).

[0093] The spatial attention components are specifically as follows ,in, It has a 7×7×3 convolution kernel. For channel-dimensional splicing;

[0094] Dual attention feature enhancement specifically refers to , where ⊙ represents element-wise multiplication;

[0095] The Laplacian smoothing algorithm is specifically... ,in, For fixed location, As a smoothing factor, For the set of adjacent vertices;

[0096] The parallel convolution path is specifically as follows ;

[0097] The specific gating weights are as follows: ;

[0098] The total loss function is specifically as follows: .

[0099] The timing processing module innovatively combines LSTM with time coding, reducing the root mean square error by 28% compared to ordinary LSTM on the standard test set. The LSTM uses a bidirectional hierarchical structure with bidirectional gating units and dynamic time warping algorithm. The bottom layer processes the original signal sequence, while the top layer receives geometric features from CNN as the initial hidden state, operating on two time scales: a fast recurrent layer (time step 0.5s) captures infiltration fluctuation details, and a slow recurrent layer (time step 5s) tracks macroscopic trend changes.

[0100] The time encoding module is specifically as follows: timestamp Mapped to dense vectors ,in, For frequency parameters, This represents the maximum time range.

[0101] The enhanced LSTM unit (including time coding) is specifically as follows:

[0102] Input gate is ;

[0103] The Gate of Oblivion ;

[0104] Output gate is ;

[0105] Candidate memories are ;

[0106] Memory update to ;

[0107] Hidden state is ;

[0108] In this context, the symbol ⊕ represents vector concatenation, ⊙ represents element-wise multiplication, and σ is the sigmoid function.

[0109] The inter-layer alignment loss function is specifically as follows: .

[0110] The cross-modal fusion cross-attention mechanism module is the core breakthrough of this research. The proposed dynamic feature gating mechanism includes three key technologies: spatial location encoding based on root distribution, weighting of time-varying feature importance, and attention correction matrix based on physical constraints. This technology enables a spatial-to-temporal attention flow that uses root voxel features as query vectors to find the most relevant hydraulic response patterns in the temporal dimension; the temporal-to-spatial attention flow, in turn, maps back to locate root hotspots that trigger specific infiltration behaviors; the entire framework embeds a three-layer physical constraint system:

[0111] (1) In the data preprocessing stage, the theoretical hydraulic conductivity calculated by the Poiseuille equation is used as prior knowledge to regularize the structural features extracted by CNN;

[0112] (2) During model training, a mass conservation term and an energy dissipation term are added to the loss function, and the constraint strength is dynamically adjusted using the Lagrange multiplier method;

[0113] (3) In the post-processing stage, the solver verifies whether the prediction results satisfy the Richards equation and automatically filters out predictions that are physically inconsistent. A differential-algebraic equation verification module is specially designed to substitute the neural network output into the discretized control equation and calculate the residual norm as a model credibility index. Attention masks are generated to filter feature combinations that do not conform to the laws of fluid dynamics. The fusion process uses gated residual connections while retaining low-level geometric features and high-level functional features. To enhance cross-modal consistency, a contrastive learning loss function is designed to minimize the distance ratio between root-infiltration feature pairs from the same sample and heterogeneous sample pairs in the latent space. Finally, the dynamic mapping relationship between spatial and temporal features is established by calculating the mutual information entropy between the CNN feature map and the LSTM hidden state.

[0114] Spatial location coding (root distribution) is specifically as follows ;

[0115] The specific allocation of time-varying feature weights is as follows: ;

[0116] The physical constraint attention correction matrix is ​​specifically as follows: ;

[0117] Spatial → Temporal Attention Specifically ;

[0118] Temporal → Spatial Attention Specifically ;

[0119] The specific gated residual is ;

[0120] The contrastive learning loss function is specifically as follows: ;

[0121] Information entropy mapping is specifically as follows .

[0122] The SHP interpreter comprises a three-tiered analytical framework: (1) at the microscale, the Monte Carlo path integral method is used to calculate the Shapley value of each voxel, and the biological rationality is enhanced by introducing root physiological constraints (such as requiring a continuous path of at least 5 pixels during Monte Carlo sampling and allowing contributions to propagate only along the direction of the vascular bundle); (2) at the mesoscale, a hierarchical attribution algorithm is developed, first identifying key root parameters such as different diameters, and then calculating the joint contribution and independent influence of each subsystem; (3) at the macroscale, a time-dependent explanatory model is constructed, revealing the differentiated regulation of the infiltration process by different root configurations through dynamic weight decomposition; and identifying the distinguishing features of the three stages of the infiltration process (initial infiltration, unsteady flow, and quasi-steady flow) through time-series decomposition technology. The mathematical rigor of the feature importance analysis is maintained while conforming to physical laws. To enhance the visualization effect of the explanatory results, the system integrates the VTK rendering engine to realize the overlay display of three-dimensional heat maps, and develops a mechanism diagram generation module based on L-system to automatically map important features to root physiological process descriptions.

[0123] The training optimization and adaptive adjustment mechanism employs a phased course learning strategy, setting up three training phases: the first phase locks the CNN parameters, focusing on optimizing the LSTM's ability to model dynamic processes; the second phase fixes the temporal encoder, emphasizing the improvement of spatial feature extraction accuracy; and finally, end-to-end joint fine-tuning is performed. The optimizer is configured with an innovative physically-aware AdamW variant, whose learning rate is automatically adjusted based on the kinetic gradient of the current batch of data. To address the sparsity of root data, an importance sampling algorithm was developed, assigning 3-5 times the sampling weight to key water-conducting regions (such as the 15mm range from the root tip) during loss calculation. To prevent overfitting, in addition to conventional Dropout, adversarial training based on hydrodynamic similarity is implemented, forcing the model to learn universal seepage patterns rather than the apparent characteristics of specific specimens.

[0124] In an embodiment of the present invention, in order to verify the accuracy of the model, Maobulakongdui in Ordos City, Inner Mongolia, was selected as the study area. Using the national 1:1,000,000 soil type map and vegetation type map, four soil types (chestnut brown soil, brown calcareous soil, aeolian sandy soil, and alluvial soil) and three vegetation types (cultivated land, grassland, and wasteland) were divided in the study area. Combined with the national DEM elevation map, different soil-shrub / grass community continuum zoning maps and contour line distribution maps were meticulously drawn. Based on this, according to the area ratio of each soil-shrub / grass community continuum in the study area and considering spatial uniformity, a total of 70 representative sample plot test points were set up, including 56 modeling points and 14 verification points.

[0125] The soil infiltration test was conducted from July 20th to November 31st, 2024, under clear weather conditions. A handheld GPS device was used to record the latitude, longitude, and elevation of the experimental sites. After removing surface soil and vegetation, the soil infiltration was tested. The double-ring infiltration method was employed, using a double ring with an inner ring diameter of 40cm and a buffer index of 0.33. In-situ field tests were conducted at 70 test sites within the study area, with repeated experiments performed at 10-meter intervals. A constant head measurement method was used during the infiltration process, maintaining a stable water level of approximately 3cm between the inner and outer rings. A smart double-ring infiltration device, designed using a self-made Mauritian bottle, vacuum pressure sensing technology, and a solenoid valve system, was employed for soil infiltration rate testing. Measurement indicators included time, temperature, initial infiltration rate, stable infiltration rate, average infiltration rate, and cumulative infiltration volume. Before infiltration, the initial soil moisture content of the test point and surrounding soil needs to be determined. The testing method involves dividing a 10*10m square area centered on the test point, and using a 16-point method with an aluminum box to collect soil samples from the top 0-20cm layer to determine the initial soil moisture content. The average value represents the initial surface soil moisture content of the area. After the infiltration test, root sampling tests are conducted at the double-ring infiltration points. After infiltration, soil cores (5cm in diameter and 5cm in height) are collected at each point using a soil drill. A 32-point method is used, with soil samples taken in layers every 20cm until no roots are found in the soil core. After the soil core was removed, it was immediately sieved using a 0.05mm soil sieve to quickly select all roots. It was then rinsed thoroughly with gauze, wiped with kitchen paper, placed in resealable bags, numbered, and stored in a freezer (4℃). Indoors, images were scanned using a desktop scanner (Epson V700PHOTO) at 300dpi. After obtaining the images, the root length (RL, cm) and root surface area (RSA, cm²) were measured using a root analysis system (WinRhizo). 2 ) and root volume (RV, cm 3 Finally, the roots were placed in an oven and dried at 75°C for 36 hours. After drying, they were removed and weighed to obtain the root biomass (RM, mg). Root length density (RLD, cm / cm²) was also measured. 3 ), root surface area density (RSAD, cm² / cm²) 3 Root volume density (RVD, cm³ / cm³) 3 ) and root biomass density (RMD, mg / cm³) 3 ).

[0126] Taking into account the mechanism of preferential flow in soil, parallel experiments were conducted at the above-mentioned test sites. At intervals of 3 meters, a control experiment was added in bare soil areas whenever possible. Standard sand was laid in the inner and outer rings, and brilliant blue solution was added to the inner ring to calibrate the preferential flow path in large pores. After the experiment, the profile was excavated to verify whether preferential flow occurred, and photographs were taken.

[0127] Using the data obtained above, the models were trained for 150 epochs. The loss values ​​of the three models stabilized, indicating convergence. Furthermore, both training and validation losses decreased simultaneously throughout the training process, suggesting that the models did not experience fitting. The model performance metrics, RMSE, MAE, and R... 2 The values ​​of 4.51, 0.64, and 0.89 indicate that the model has good simulation performance.

[0128] It should also be noted that, such as Figure 3 As shown in Table 1, the obtained matrix flow data were compared with the model. The results of parameter fitting and model fitting performance index verification showed that Horton was the optimal model. Table 1 only shows the simulation results using matrix flow data.

[0129] Evaluation indicators Philip Green-Ampt Mein-Larson kostiakov Horton Holtan Mezencev Adj-R² 0.924 0.874 0.902 0.896 0.980 0.876 0.913 RMSE 0.706 0.435 0.254 0.425 0.054 0.125 0.178

[0130] Table 1

[0131] The priority flow data obtained from the experiments were used to drive the model, which was then fed into both the Horton model and the priority flow model developed in this study. After 150 epochs of training, the model loss value tended to stabilize, indicating convergence. Furthermore, during the entire training process, both training and validation losses decreased simultaneously, indicating that the model did not experience fitting. The performance metrics of our model, RMSE, MAE, and R² values, were 4.51, 0.64, and 0.89, respectively, which are significantly improved compared to the results of the basic (Horton) model (7.67, 0.53, 0.76), demonstrating that the priority flow model framework has good simulation performance.

[0132] In summary, the biomimetic-driven soil infiltration process simulation and optimization method disclosed in this invention can:

[0133] (1) Improve prediction accuracy and break through the assumption of homogeneous medium in traditional models. Realize root-specific flow channel network modeling. By integrating three-dimensional root structure data and time-series infiltration dynamic data, and combining 3D-CNN to extract root features, LSTM to process time-series information and cross-modal fusion mechanism, accurately capture the suddenness and spatial heterogeneity of priority flow. Solve the problem of low simulation accuracy and easy underestimation of flow rate in traditional models. The model performance index (RMSE=4.51, MAE=0.64, R²=0.89) verifies its high accuracy.

[0134] (2) Enhanced applicability, applicable to complex underlying surface topography, can simulate meteorological-soil-vegetation continuum for different soil types, vegetation cover, landforms and climate characteristics, overcomes the limitations of traditional models at small scale and scale dependence, can be extended to field, watershed and even regional scale, and can predict the dynamic changes of infiltration rate during different vegetation phenological periods, thus expanding the application scenarios.

[0135] (3) Reduce the cost of parameter acquisition. Through multi-source data preprocessing and standardization process, combined with physical constraint mechanisms (such as the Poiseuille equation and Richards equation verification), the dependence on high-cost measured parameters is reduced, and the uncertainty caused by indirect fitting and inversion of parameters in models such as dual-pore is avoided. In addition, the use of intelligent dual-ring infiltration equipment and standardized root sampling and measurement process in the experimental design reduces the cost of data acquisition.

[0136] (4) Revealing the dynamic mechanism, introducing time evolution operators to establish machine learning representation of transient flow, solving the problem that static parameter system is difficult to describe dynamic infiltration process, can capture the whole stage evolution of preferential flow from initial infiltration, unsteady flow to quasi-steady flow, and at the same time, through the SHP interpreter, analyze the root configuration regulation mechanism of infiltration from micro, meso and macro scales, providing key technical means for the study of soil water infiltration performance, groundwater runoff generation mechanism, and runoff generation.

[0137] (5) To ensure physical rationality, a triple physical constraint mechanism including mass conservation equation, non-negativity constraint and boundary conditions is designed. Combined with the differential-algebraic equation verification module, the model output is ensured to conform to the laws of fluid mechanics, breaking the dilemma of the black box model being separated from physical laws and improving the reliability and interpretability of the model.

[0138] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A biomimetic-driven method for simulating and optimizing soil infiltration processes, characterized in that, Includes the following steps: S1: Preprocess and standardize the multi-source data, which includes three-dimensional root structure data and infiltration time series data. By registering a unified coordinate system in time and space, the root voxel unit corresponds to the hydraulic response of a specific time step. S2: Construct a model architecture that integrates root features and infiltration dynamics. The model architecture includes a root feature extraction module, a temporal processing module, and a cross-modal fusion cross-attention mechanism module. The root feature extraction module uses a 3D convolutional neural network to extract root topological features. The temporal processing module combines a long short-term memory network with time coding to process the infiltration dynamics. The cross-modal fusion cross-attention mechanism module realizes bidirectional mapping between spatial and temporal features and embeds physical constraints. S3: A phased learning strategy is adopted to train and optimize the model. The learning rate is adjusted through a physical perception optimizer, and the model's generalization ability is improved by combining importance sampling and adversarial training. S4: Based on the trained model, output the priority flow prediction results and analyze the prediction results from micro, meso and macro scales through an interpreter.

2. The biomimetic-driven soil infiltration process simulation and optimization method according to claim 1, characterized in that, In step S1, the preprocessing of the three-dimensional root system structure data includes the following steps: S1.1: Mesh repair is performed on the root system model to fill topological defects, and the Laplacian smoothing algorithm is used to eliminate surface noise; S1.2: Perform anisotropic voxelization treatment, maintain the first resolution in the horizontal plane, and set the second resolution along the root growth direction; S1.3: The improved MarchingCubes algorithm is applied to generate a voxel model with pore connectivity, and a multi-channel feature tensor containing structural density, hydraulic conductivity and biological age is constructed.

3. The biomimetic-driven soil infiltration process simulation and optimization method according to claim 2, characterized in that, In step S1, the preprocessing of the infiltration time series data includes the following steps: S1.4: Perform wavelet denoising on the original signal and expand the features based on the Richards equation to obtain derived variables; S1.5: Perform non-stationarity testing and segmented standardization, and align the time based on the arrival time of the initial wetting peak; S1.6: Construct a sliding time window matrix to complete the standardization processing of time series data.

4. The biomimetic-driven soil infiltration process simulation and optimization method according to claim 3, characterized in that, In step S2, the construction of the root feature extraction module includes the following steps: S2.1: An improved 3D-CNN architecture is adopted as the encoder. The encoder includes five stages of downsampling. Each stage sets up parallel paths of standard convolution and anisotropic convolution, and features are fused through an attention gating mechanism. S2.2: The decoder uses deformable convolution instead of conventional upsampling to adapt to the irregular growth morphology of the root system; S2.3: Introduce a root continuity loss function and use Laplacian smoothing constraints to ensure topological consistency between the feature map and the real root anatomy.

5. The biomimetic-driven soil infiltration process simulation and optimization method according to claim 4, characterized in that, In step S2, the construction of the timing processing module includes the following steps: S2.4: Combining Long Short-Term Memory Network with Time Encoding Module; S2.5: Set up a bidirectional gating unit and a dynamic time warping algorithm. The bottom layer processes the original signal sequence, and the top layer receives the geometric features output by the root feature extraction module as the initial hidden state. S2.6: Configure a fast circulation layer and a slow circulation layer, which run at a first time step and a second time step, respectively, to capture infiltration dynamics at different scales.

6. The biomimetic-driven soil infiltration process simulation and optimization method according to claim 5, characterized in that, In step S2, the construction of the cross-modal fusion cross-attention mechanism module includes the following steps: S2.7: Introduce spatial location coding and time coding; S2.8: Construct a bidirectional attention flow from space to time and from time to space; S2.9: Embedded physical constraint system, including feature regularization based on the Poiseuille equation, loss function constraints with mass conservation and energy dissipation terms, and post-processing verification based on the Richards equation.

7. The biomimetic-driven soil infiltration process simulation and optimization method according to claim 6, characterized in that, In step S3, the phased course learning strategy includes the following steps: S3.1: In the first stage, the parameters of the root feature extraction module are locked to optimize the modeling capability of the time series processing module for dynamic processes; S3.2: The second stage fixes the parameters of the temporal encoder to improve the accuracy of root feature extraction; S3.3: The third stage involves end-to-end joint fine-tuning, using the physically-aware AdamW variant optimizer, whose learning rate is automatically adjusted based on the kinetic gradient of the current batch of data.

8. The biomimetic-driven soil infiltration process simulation and optimization method according to claim 7, characterized in that, In step S4, the interpreter's analysis process includes the following steps: S4.1: At the microscale, the Monte Carlo path integral method is used to calculate the Shapley value of each voxel, and root physiological constraints are introduced to ensure biological rationality; S4.2: At the mesoscale, key root system parameters are identified through a hierarchical attribution algorithm, and the joint contribution and independent influence of each subsystem are calculated. S4.3: A time-dependent explanatory model is constructed at the macro scale. Through dynamic weight decomposition, the differences in the regulation of the infiltration process by different root configurations are revealed, and the distinguishing features of the initial infiltration, unsteady flow and quasi-steady flow stages are identified.

Images