Mechanism deep learning dual-drive flow-slip disaster numerical simulation system

Abstract

The invention provides a mechanism deep learning dual-drive flow-slip disaster numerical simulation system, and relates to the technical field of geological disaster numerical simulation. The system integrates a physical mechanism model and a deep learning method. Based on a high-resolution DEM (Digital Elevation Model), by introducing a rock-soil type self-adaptive substrate entrainment mechanism, utilizing a shallow water wave equation to accurately calculate the terrain elevation, the flow direction, the flow velocity and the accumulation thickness of the first two time steps; and then taking the obtained multivariable field as four-channel input, and driving a CNN-GRU coupling network to perform rolling prediction. According to the method, the final accumulation form, the influence range and the dynamic evolution process can be quickly obtained only through a small amount of initial mechanism calculation, the simulation efficiency is remarkably improved on the premise of ensuring physical rationality, and the defects that a traditional pure mechanism model is long in calculation time consumption and a pure data driving model is physically distorted are overcome; the method is suitable for rapid risk assessment and emergency decision support of flow slip disasters such as landslide and debris flow.

Description

Technical Field

[0001] This invention relates to the field of numerical simulation technology for geological disasters, and more specifically, to a numerical simulation system for landslide disasters driven by both deep learning and mechanistic mechanisms. Background Technology

[0002] Landslides are high-speed, long-range geological hazards, often triggered by earthquakes, rainfall, or human engineering activities. Their movement involves complex multi-physics interactions, including solid-liquid coupling, particle collisions, and entrainment of underlying materials. Accurate simulation of their dynamic evolution is crucial for hazard risk assessment and emergency decision-making. Currently, numerical simulation of landslides mainly follows two technical approaches: Purely physical mechanism-driven simulation: Numerical models are constructed based on continuum mechanics, such as shallow water wave equations and granular flow theory, and solved iteratively using numerical methods such as discrete element method, DEM, material point method, MPM or finite volume method, FVM, etc. These methods rely on rigorous physical assumptions and parameter calibration, requiring discrete calculations of momentum and mass conservation equations for each time step, and can better reflect the physical nature of flow slip.

[0003] Purely data-driven deep learning simulations: These methods utilize historical disaster data or samples generated through mechanism simulations to train neural networks that directly predict the spatiotemporal evolution of landslides. They achieve rapid predictions by learning statistical patterns from the data, without requiring explicit physical equations, and are computationally efficient.

[0004] Shortcomings of existing technologies: Efficiency bottleneck of pure physical mechanism simulation: Pure mechanism simulation requires iterative calculation of high-resolution grids for each time step. Especially for large-scale, long-duration landslide events, such as landslides with volumes of millions of cubic meters, the computational load increases exponentially, making it difficult to meet the rapid prediction needs in emergency scenarios. At the same time, some complex physical processes, such as the dynamic threshold of the base material, are difficult to accurately characterize through analytical equations, which can easily lead to deviations between simulation results and actual disasters.

[0005] Physical consistency defects of pure deep learning simulations: Pure data-driven models rely on a large number of high-quality samples, while actual on-site monitoring data of landslide disasters are scarce. If samples are generated solely by mechanism simulation, they are easily limited by the distribution of training data. In addition, neural networks lack built-in constraints on physical laws, and the prediction results may show non-physical phenomena, such as negative accumulation thickness, contradiction between flow velocity direction and terrain slope, and it is difficult to explain the physical logic of the prediction results.

[0006] The disconnect between physical processes and data-driven approaches: Existing technologies often apply physical mechanisms and deep learning independently, failing to achieve deep integration between the two. Pure mechanism simulations cannot leverage the spatiotemporal correlations in the data to accelerate computation, while pure deep learning simulations lack the guidance of physical mechanisms to improve prediction accuracy, making it difficult to balance efficiency and accuracy when dealing with complex flow events.

[0007] The spatiotemporal prediction capability of multivariate fields is insufficient: the evolution of landslide hazards involves the dynamic coupling of multiple variables such as topography, flow velocity, flow direction, and deposition thickness. The synchronous prediction accuracy of traditional mechanistic models for multivariate fields is affected by numerical discretization errors. Pure deep learning models, if using a single network structure, such as only CNN or only GRU, are unable to simultaneously capture the local features of the spatial field and the long-range dependence of the time series, leading to the accumulation of continuous prediction errors for multivariate fields. Therefore, a mechanistic deep learning dual-driven numerical simulation system for landslide hazards is proposed. Summary of the Invention

[0008] The purpose of this invention is to address the problems raised in the existing background technology. To achieve the above-mentioned objective, this invention provides the following technical solution: a mechanism-based deep learning dual-driven numerical simulation system for landslide disasters, comprising the following modules: a data input module for inputting initial topographic data of the landslide disaster triggering zone, DEM data with a resolution of not less than 10m × 10m, physical parameters, and the base friction coefficient. Collision force threshold F th ; The data input module introduces a substrate entrainment mechanism: at each time step For each moving unit, calculate its normal collision stress with the base. ,when Entrainment is triggered at certain times, where the entrainment threshold is... Entrained thickness increment Calculate using the following formula:

[0009] in The density of the fluid is... For local slope angle; updated terrain elevation after inclusion: ; The data input module outputs the multivariate field data for the first N time steps. ; The deep learning prediction module constructs a CNN-GRU coupled network, with two consecutive time steps as input. and Four-channel field data The four channels are: terrain elevation z, flow direction angle, and so on. , ; The CNN contains 3 convolutional layers, with a kernel size of [size missing]. The number of channels is 64, 128, and 256 respectively, and the activation function is ReLU; The GRU hidden state dimension is 128, and the output is upsampled and reconstructed by a convolutional layer. Prediction field ; The iterative prediction and termination determination module is used to... For input rolling prediction Until the termination condition is met:

[0010] in ; The system integration and output module of the numerical simulation system for landslide disaster driven by both mechanistic deep learning is used to output the final accumulation morphology, maximum impact range, stopping time, and dynamic evolution video of the entire process.

[0011] As a preferred technical solution of the present invention, the entrainment threshold Dynamic settings based on soil and rock type: For cohesive soil, take... For sandy soil For crushed stone mixtures .

[0012] As a preferred embodiment of the present invention, the spatial resolution of the multivariable field data is consistent with that of the input DEM, and is not less than 5 meters; the time step is... Satisfying the CFL stability condition:

[0013] in This represents the grid spacing.

[0014] As a preferred embodiment of the present invention, the loss function of the deep learning prediction module is a residual loss function, defined as:

[0015] in , This represents the spatial gradient operator, used to enhance the spatial smoothness and structural consistency of the predicted field.

[0016] As a preferred embodiment of the present invention, the CNN-GRU coupled network uses 100–500 sets of spatiotemporal sequence samples generated by the physical mechanism-driven module for supervised learning during the training phase. Each set of samples contains a four-channel field across four consecutive time steps. The training batch size is 8–16, the optimizer is Adam, and the initial learning rate is... A learning rate decay strategy is adopted.

[0017] As a preferred technical solution of the present invention, the value of N is 2, that is, only the physical mechanism driving module needs to calculate the data at two time steps t=0 and t=Δt, and the deep learning prediction module can be started to predict the entire process from t=2Δt to the cessation of the disaster.

[0018] As a preferred technical solution of the present invention, the flow direction angle Calculated from the flow velocity vector:

[0019] Normalized to [0, 2π] in radians The interval is used as an independent channel for the CNN input.

[0020] As a preferred technical solution of the present invention, the mechanism-deep learning dual-driven numerical simulation system for slippage disasters enforces a non-negativity constraint during the prediction process: for the predicted accumulation thickness... and terrain elevation Perform truncation, that is , ,in This represents the elevation of incorruptible bedrock.

[0021] As a preferred technical solution of the present invention, the physical mechanism driving module uses the finite volume method to discretize the shallow water wave equation, and the flux calculation uses an HLLC approximate Riemann solver, with the friction term... use Model:

[0022] middle This is the turbulence drag coefficient. is the Coulomb friction coefficient.

[0023] As a preferred technical solution of the present invention, the mechanism deep learning dual-driven numerical simulation system for slippage disaster is deployed on a GPU server that supports CUDA 11.0 and above. It uses cuDNN to accelerate CNN inference, and the time for a single rolling prediction is no more than 0.5 seconds (taking a 1024×1024 grid as an example). The overall simulation efficiency is 5–10 times higher than that of traditional all-physics simulation.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention improves the accuracy and physical realism of numerical simulation: by introducing a base entrainment mechanism, it dynamically simulates the interaction between the slip body and the base, accurately updates the changes in terrain elevation, and makes the physical evolution of the slip process more consistent with the actual disaster mechanism, thus solving the problem of prediction deviations in accumulation morphology and movement range caused by the neglect of base material entrainment in traditional models.

[0025] Reduce computational costs and efficiency bottlenecks: Deep learning prediction can be started by calculating the data of the first two time steps in the physical mechanism module, which replaces the traditional pure mechanism simulation mode that requires high time consumption and iteration throughout the process. This greatly shortens the computation time of the whole process simulation of slippage disasters, and is especially suitable for the rapid prediction needs of large-scale, long-duration slippage events.

[0026] Enhancing the ability to predict spatiotemporal sequences: The CNN-GRU coupled network combines the feature extraction capability of convolutional layers for spatial field data with the advantage of GRU in capturing the dynamic evolution of time series, effectively learning the spatiotemporal correlation patterns of landslide terrain, flow direction, flow velocity, and deposition thickness, thereby improving the accuracy of continuous prediction of multivariable fields.

[0027] To ensure the rationality and stability of the prediction results: Non-negative constraints are used to handle the accumulation thickness and terrain elevation to avoid physical invalid values ​​in deep learning predictions; CFL stability conditions constrain the time step, and the residual loss function is combined to enhance the spatial smoothness and structural consistency of the prediction field, ensuring that the simulation results conform to physical logic.

[0028] Adaptable to different geological conditions and scenarios: The base entrainment threshold can be dynamically set according to different rock and soil types such as cohesive soil, sandy soil, and gravel mixture. The input DEM data resolution is no less than 10m×10m, taking into account both simulation accuracy and data applicability, and can cover the simulation needs of landslide disasters under various geological backgrounds.

[0029] The output information is rich and practical: the final output shows the final depositional form, maximum impact range, stopping time, and dynamic evolution video of the entire process of the landslide disaster, providing intuitive and crucial decision-making basis for disaster risk assessment and emergency rescue plan formulation, and helping to improve the scientific nature and timeliness of geological disaster prevention and control. Attached image description: Figure 1 The system module data block diagram provided by this invention; Figure 2 A data parameter block diagram provided for this invention; Figure 3 This is a block diagram of the configuration data for the CNN-GRU coupled network provided by the present invention; Figure 4 A flowchart of physical calculation data provided for this invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are specific implementations of the present invention and are not limited to all embodiments.

[0031] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0032] It should be noted that, in the absence of conflict, the embodiments and features and technical solutions in the embodiments of the present invention can be combined with each other. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0033] Example: A mechanism-based deep learning dual-driven numerical simulation system for landslide disasters includes the following modules: a data input module for inputting initial topographic data of the landslide disaster triggering zone, DEM data with a resolution of not less than 10m×10m, physical parameters, and base friction coefficient. Collision force threshold F th ; The data input module introduces a substrate entrainment mechanism: at each time step For each moving unit, calculate its normal collision stress with the base. ,when Entrainment is triggered at certain times, where the entrainment threshold is... Entrained thickness increment Calculate using the following formula:

[0034] in The density of the fluid is... For local slope angle; updated terrain elevation after inclusion: ; The data input module outputs the multivariate field data for the first N time steps. ; The deep learning prediction module constructs a CNN-GRU coupled network, with two consecutive time steps as input. and Four-channel field data The four channels are: terrain elevation z, flow direction angle, and so on. , ; The CNN contains 3 convolutional layers, with a kernel size of [missing value]. The number of channels is 64, 128, and 256 respectively, and the activation function is ReLU; The GRU hidden state has a dimension of 128, and the output is upsampled and reconstructed by a convolutional layer. Prediction field ; The iterative prediction and termination determination module is used to... For input rolling prediction Until the termination condition is met:

[0035] in ; The system integration and output module of the numerical simulation system for landslide disaster driven by both mechanistic deep learning is used to output the final accumulation morphology, maximum impact range, stopping time, and dynamic evolution video of the entire process.

[0036] Entrainment threshold Dynamic settings based on soil and rock type: For cohesive soil, take... For sandy soil For crushed stone mixtures .

[0037] The spatial resolution of the multivariable field data is consistent with the input DEM and is no less than 5 meters; the time step... Satisfying the CFL stability condition:

[0038] in This represents the grid spacing.

[0039] The loss function for the deep learning prediction module is the residual loss function, defined as:

[0040] in , This represents the spatial gradient operator, used to enhance the spatial smoothness and structural consistency of the predicted field.

[0041] The CNN-GRU coupled network uses 100–500 sets of spatiotemporal sequence samples generated by the physical mechanism-driven module for supervised learning during the training phase. Each set of samples contains a four-channel field across four consecutive time steps. The training batch size is 8–16, the optimizer is Adam, and the initial learning rate is... A learning rate decay strategy is adopted.

[0042] With N set to 2, the deep learning prediction module can be activated to predict the entire process from t=2Δt until the disaster stops, as the physical mechanism-driven module only needs to calculate the data at two time steps: t=0 and t=Δt.

[0043] Flow direction angle Calculated from the flow velocity vector:

[0044] Normalized to [0, 2π] in radians The interval is used as an independent channel for the CNN input.

[0045] A mechanism-based deep learning-driven numerical simulation system for landslide hazards enforces non-negativity constraints during the prediction process: for the predicted deposition thickness... and terrain elevation Perform truncation, that is , ,in This represents the elevation of incorruptible bedrock.

[0046] The physical mechanism-driven module uses the finite volume method to discretize the shallow water wave equations, and the flux calculation uses the HLLC approximate Riemann solver. The friction term... use Model:

[0047] middle This is the turbulence drag coefficient. is the Coulomb friction coefficient.

[0048] A mechanism-based deep learning-driven numerical simulation system for slipway disasters is deployed on a GPU server supporting CUDA 11.0 and above. It utilizes cuDNN to accelerate CNN inference, and the time for a single rolling prediction is no more than 0.5 seconds (taking a 1024×1024 grid as an example). The overall simulation efficiency is 5–10 times higher than that of traditional all-physics simulation.

[0049] Working principle: I. Physical mechanism driven stage: Constructing a realistic initial evolution process The system first uses a high-precision digital elevation model (DEM) and initial parameters of the slip source (such as volume, density, and location) to call an improved two-dimensional shallow water wave equation (SWE) solver to simulate the motion of the slip body in the early stage of the disaster (the first N time steps, N=2 or 3).

[0050] The key innovation lies in the introduction of a substrate entrainment mechanism: At each time step Δt (typically 0.1–2.0 s), the system calculates the normal collision stress between each moving unit (or particle) and the underlying substrate. ; When τ n Exceeding the entrainment threshold related to soil and rock type When the pressure is 3–5 kPa for cohesive soil and 7–10 kPa for gravel mixture, entrainment is considered to have occurred. Entrainment thickness increment Δh e Calculated by the following formula:

[0051] in This is an empirical coefficient. The density of the fluid is... This refers to the local slope angle; After being carried over, the terrain elevation at that location is dynamically updated: This enables bidirectional coupling between terrain and flow.

[0052] This stage outputs four-dimensional physical field data for each time step (stored in GeoTIFF format): Topographic elevation (zt); Flow direction angle Flow velocity amplitude ; Deposition thickness ; Deep learning modeling phase: Building a spatiotemporal evolution predictor: The system builds a Coupled neural networks are used to learn the evolution of fluidic bodies in the spatiotemporal dimensions: Spatial modeling : The four-channel field at each time step Input multi-layer convolutional network Extract the spatial structural features of local terrain and flow; Time modeling : Two consecutive time steps The spatial feature vectors are input into the GRU unit (hidden dimension 128) to model temporal dependencies; Output reconstruction: The GRU output is upsampled and convolutionally processed to reconstruct a four-channel prediction field with the same resolution as the input. .

[0053] The model is trained end-to-end using a residual loss function:

[0054] The second term introduces spatial gradient regularization to ensure that the prediction results have good spatial continuity and physical rationality.

[0055] III. Rolling Prediction Phase: Efficiently Simulating the Entire Evolution Process After the model training is complete, the system enters rolling prediction mode: Data from the first and second time steps generated by the physical mechanism module As initial input; Deep learning module predicts the 3rd time step: ; Update the prediction results to the new sequence: To predict ; Repeat the above process to form an autoregressive prediction chain: .

[0056] This process eliminates the need to call the time-consuming physics solver again, relying solely on lightweight neural network inference, which significantly improves computational efficiency.

[0057] IV. Termination Judgment and Result Output Stage The system monitors and predicts the velocity field in real time, and determines that the disaster has stopped when the following termination conditions are met:

[0058] in This is the flow rate stopping threshold.

[0059] Once terminated, the system outputs: final accumulation thickness distribution map; maximum impact range, flooded / accumulated area; disaster cessation time; hazard zoning recommendations for emergency response.

[0060] This system achieves the following through a dual-drive architecture of mechanism fidelity and learning acceleration: Physical consistency: the initial evolution is controlled by strict physical equations, and the entrainment mechanism reflects the real erosion process; Computational efficiency: only 2-3 steps of physical calculation are required to predict hundreds of evolution steps, improving efficiency by 5-10 times; High prediction accuracy: the CNN-GRU model captures nonlinear spatiotemporal coupling relationships, and the error is controlled within 5%. Work Process: Step 1: Input Data Preparation: Obtain a high-precision digital elevation model of the study area. Spatial resolution no less than 5 meters, format as follows or ; Define the initial parameters of the slip source region, including the slip source location, latitude and longitude or grid coordinates; initial volume. ; fluid density ; Soil and rock type, used to determine entrainment threshold and friction parameters; Set simulation parameters: time step Grid size Stability conditions are automatically verified.

[0061] Step 2: The physical mechanism drives the initial simulation. The self-developed two-dimensional shallow water wave equation (SWE) solver is called. The finite volume method is used to discretize the control equations, and the flux calculation uses the HLLC approximate Riemann solver. Perform base entrainment calculations at each time step: calculate the normal collision stress for each moving element. ;like Then according to the formula

[0062] Calculate the entrainment thickness and update the terrain elevation: ; Record number (i.e., N=3) Four-variable field data at three time points, each data set containing four channels: Channel 1: Topographic elevation Channel 2: Flow direction angle; Channel 3: Flow velocity amplitude Channel 4: Stacking thickness The above data is saved as a multi-channel GeoTIFF image sequence and used as the training and initial prediction input for the deep learning model.

[0063] This stage only runs 2–3 time steps and typically takes 1–3 minutes (depending on the size of the region), which is much shorter than traditional full-time domain simulation.

[0064] Step 3: Deep learning model training (offline or online), constructing a CNN-GRU coupled network: Input: two consecutive time steps tensor; CNN part: 3 convolutional layers (3×3 kernels, 64→128→256 channels, ReLU activation); GRU part: 128 hidden units, receiving spatiotemporal features extracted by CNN; Output layer: Upsampling + 1×1 convolution, output Prediction field; Supervised training is performed using multiple sets of spatiotemporal sequences generated in step 2, such as 100–500 sets. Optimization using residual loss function:

[0065] After training, the model weights are saved for subsequent rolling predictions. The model can be pre-trained for general scenarios or fine-tuned for high-precision requirements in specific regions.

[0066] Step 4: Rolling Predictive Evolution, Core Efficiency Stage: Using data from the first and second time steps obtained from physical simulations. As initial input; the trained CNN-GRU model is invoked to predict the 3rd time step:

[0067] Imposing physical constraints on the prediction results: Stack thickness truncation: ; Terrain lower limit constraint: ; Update the input sequence: predict ; Forming an autoregressive prediction chain:

[0068] Each prediction takes ≤0.5 seconds (GPU accelerated, 1024×1024 grid), and can continuously predict hundreds of time steps.

[0069] Step 5: Termination Judgment and Result Output: After each prediction, calculate the maximum flow velocity across the entire field:

[0070] If two consecutive time steps satisfy:

[0071] If the slippery surface has stopped moving, the prediction is terminated. The system automatically generates and outputs the following results: final accumulation thickness distribution map (GeoTIFF format); Maximum impact area of ​​the disaster (vector boundary or grid mask); stopping time Dynamic video of the entire process; Hazard level zoning map (based on accumulation thickness and flow velocity threshold).

[0072] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or equivalent substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. A mechanism-based deep learning dual-driven numerical simulation system for landslide disasters, comprising the following modules: a data input module for inputting initial topographic data of the landslide disaster triggering zone, DEM data with a resolution of not less than 10m×10m, physical parameters, and base friction coefficient. Collision force threshold F th ; The data input module introduces a substrate entrainment mechanism: at each time step For each moving unit, calculate its normal collision stress with the base. ,when Entrainment is triggered at certain times, where the entrainment threshold is... Entrained thickness increment Calculate using the following formula: ； in The density of the fluid is... For local slope angle; updated terrain elevation after inclusion: ; The data input module outputs the multivariate field data for the first N time steps. ; The deep learning prediction module constructs a CNN-GRU coupled network, with two consecutive time steps as input. and Four-channel field data The four channels are: terrain elevation z, flow direction angle, and so on. , ; The CNN contains 3 convolutional layers, with a kernel size of [size missing]. The number of channels is 64, 128, and 256 respectively, and the activation function is ReLU; The GRU hidden state dimension is 128, and the output is upsampled and reconstructed by a convolutional layer. Prediction field ; The iterative prediction and termination determination module is used to... For input rolling prediction Until the termination condition is met: ； in ; The system integration and output module of the numerical simulation system for landslide disaster driven by both mechanistic deep learning is used to output the final accumulation morphology, maximum impact range, stopping time, and dynamic evolution video of the entire process.

2. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The entrainment threshold Dynamic settings based on soil and rock type: For cohesive soil, take... For sandy soil For crushed stone mixtures .

3. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The spatial resolution of the multivariable field data is consistent with that of the input DEM, and is no less than 5 meters; the time step... Satisfying the CFL stability condition: ； in This represents the grid spacing.

4. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The loss function of the deep learning prediction module is the residual loss function, defined as: ； in , This represents the spatial gradient operator, used to enhance the spatial smoothness and structural consistency of the predicted field.

5. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The CNN-GRU coupled network employs supervised learning during the training phase using 100–500 sets of spatiotemporal sequence samples generated by the physical mechanism-driven module. Each set of samples contains a four-channel field across four consecutive time steps. The training batch size is 8–16, the optimizer is Adam, and the initial learning rate is [missing information]. A learning rate decay strategy is adopted.

6. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The value of N is 2, meaning that the deep learning prediction module can be activated to predict the entire process from t=2Δt to the cessation of the disaster by only needing the physical mechanism driving module to calculate the data at two time steps, t=0 and t=Δt.

7. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The flow direction angle Calculated from the flow velocity vector: ； Normalized to [0, 2π] in radians ; The interval is used as an independent channel for the CNN input.

8. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The aforementioned mechanism-based deep learning dual-driven numerical simulation system for slippage disasters enforces a non-negativity constraint during the prediction process: for the predicted accumulation thickness... and terrain elevation Perform truncation, that is , ,in This represents the elevation of incorruptible bedrock.

9. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The physical mechanism-driven module uses the finite volume method to discretize the shallow water wave equations, and the flux calculation uses an HLLC approximate Riemann solver. The friction term... use Model: ； middle This is the turbulence drag coefficient. is the Coulomb friction coefficient.

10. The mechanism-deep learning dual-driven numerical simulation system for slippage disasters according to claim 1, characterized in that, The aforementioned mechanism-driven deep learning-based numerical simulation system for slipway disasters is deployed on a GPU server supporting CUDA 11.0 and above. It utilizes cuDNN to accelerate CNN inference, and the time for a single rolling prediction is no more than 0.5 seconds (taking a 1024×1024 grid as an example). The overall simulation efficiency is 5–10 times higher than that of traditional all-physics simulation.

Images