Complex shock wave capture method, system, device and storage medium based on multi-strategy fusion physical information neural network.
By fusing physical information neural networks with multiple strategies, smoothing initial conditions, improving high-frequency feature representation, and dynamically adjusting loss weights, the problem of insufficient reconstruction capability of complex shock wave shallow water flow fields is solved, achieving higher accuracy and stable shock wave capture.
Patent Information
- Application Number
- CN202611610101.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively coordinate initial conditions, boundary conditions, and physical constraints, resulting in insufficient ability to reconstruct complex shock wave shallow water flow fields. In particular, when dealing with discontinuous initial conditions and high-frequency spatiotemporal variations, network training becomes difficult and shock wave profile reconstruction is inaccurate.
A multi-strategy fusion physical information neural network is adopted. The initial conditions are smoothed by data spatiotemporal configuration and course loading module. The expressive power is improved by Fourier high-frequency feature encoding layer. The shock wave region is dynamically captured by adaptive spatiotemporal resampling evaluator. The loss weight is adjusted by LRAnneal global gradient weight scheduling engine to achieve adaptive adjustment of global composite energy loss function and form closed loop optimization.
It improves the reconstruction capability of complex shock wave shallow water flow fields, ensures training stability and accuracy, reduces excessive smoothing of shock wave profiles, enhances the expression of high-frequency features and coverage of local regions, coordinates the gradient contributions of different physical constraints, and improves the convergence and accuracy of the model.
Smart Images

Figure CN123116874A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and fluid mechanics, and particularly relates to a complex shock wave capture method, system, device and storage medium based on a multi-strategy fusion physical information neural network. Background Technology
[0002] Shallow water equations (SWEs) are a set of governing equations derived by integrating the fluid motion equations along the water depth direction under the assumption of hydrostatic pressure. They are widely used in simulating shallow water dynamic processes such as dam-break wave propagation, tsunami evolution, estuary dynamics, and flood inundation. Accurate reconstruction of shallow water flow fields with discontinuous initial conditions, locally steep wavefronts, and complex bed topography is a crucial requirement in related engineering analyses and disaster process simulations.
[0003] For solutions to the aforementioned shallow water dynamics problems using physical information neural networks, it is necessary to simultaneously fit the initial conditions and boundary conditions while satisfying the constraints of the governing equations. Existing training methods face the following main technical challenges, particularly regarding strongly discontinuous initial conditions and evolving shock waves.
[0004] First, the discontinuous initial water depth has abrupt changes in space, while the network output is usually a continuous function. If a steep step initial value is applied directly in the early stage of training, the network needs to fit local strong gradients and other physical constraints at the same time before a reasonable flow field is formed, which can easily increase the difficulty of co-optimizing the initial conditions and the governing equations.
[0005] Secondly, ordinary fully connected networks are insufficient in representing high-frequency spatiotemporal variations, and tend to fit local steep wavefronts as overly wide smooth transitions, affecting the reconstruction of shock wave profiles and their propagation processes.
[0006] The initial condition loss, boundary condition loss, and physical residual loss contribute differently to the gradient of network parameters and change with the steepening of the initial value and the widening of the collocation. Fixed loss weights are difficult to consistently take into account these constraints, and it is easy for some losses to dominate parameter updates while other constraints are not adequately improved.
[0007] Therefore, within the same training process, how to coordinate the incremental loading of initial conditions, high-frequency feature representation, local collocation configuration, and gradient contributions of multiple physical constraints, so that newly added collocations and updated initial targets can effectively participate in subsequent optimization, thereby improving the reconstruction capability of complex shock wave shallow water flow fields, is a technical problem that urgently needs to be solved. Summary of the Invention
[0008] The purpose of this invention is to provide a method, system, device and storage medium for capturing complex shock waves based on a multi-strategy fusion physical information neural network, and to solve the technical problem of how to improve the reconstruction capability of complex shock wave shallow water flow fields.
[0009] This application provides a complex shock wave capture method based on a multi-strategy fusion physical information neural network, comprising: S01. Construct a complex shock wave capture model based on a multi-strategy fusion physical information neural network; S02. Configure a spatiotemporal baseline point set for training the complex shock wave capture model, smoothing discontinuous initial conditions into a gradually changing transition state. The spatiotemporal baseline point set includes the set of coordinate points within the baseline. Boundary point set With the initial point set ; S03. Use a complex shock wave capture model to perform up-dimensional mapping and complete the forward propagation of the macroscopic state prediction values of shallow water fluids. S04. Using a complex shock wave capture model, obtain the augmented set of fundamental internal collocation points. ; S05, based on , and A global composite energy loss function is constructed, and the loss weights of various physical constraints of the global composite energy loss function are adaptively adjusted. Through hybrid iteration, the complex shock wave capture model is converged and the output global high-frequency spatiotemporal dynamic field is reconstructed.
[0010] Optionally, in some embodiments of this application, the complex shock wave capture model includes a data spatiotemporal configuration and course loading module, a Fourier high-frequency feature encoding layer, a backbone forward physical reconstruction network, an adaptive spatiotemporal resampling evaluator, and an LRAnneal global gradient weight scheduling engine.
[0011] Optionally, in some embodiments of this application, the spatiotemporal configuration and course loading module for the data of the complex shock wave capture model is used to configure the spatiotemporal basis point set for training the complex shock wave capture model, smoothing the discontinuous initial conditions into a gradually changing transition state, including the following steps: S021. Determine the spatial threshold of the study. And the projection time window [0,T], where , It is a two-dimensional real space; the Latin hypercube sampling algorithm is used in the complete spatiotemporal domain. Internally generated set of internal collocation points for physical residual monitoring At the same time along the physical boundary disjunctive boundary point set Determine the initial point set along the time boundary, i.e., the initial time. ; S022, Set the target's true state at time zero as... ,in The initial water depth, and They are respectively direction and Initial velocity in direction, To extrapolate the end time of the time window, we extract the initial conditions that are discontinuous in the simulation of tsunami caused by extreme dam failure or fault displacement. S023, with To ultimately approach the target, in the initial point set We introduce a dynamically steepened hyperbolic tangent function to construct alternative training objectives for course learning, along... The hyperbolic tangent function for the direction is as follows: ; In the formula, Represents the hyperbolic tangent function. For the first Position during round iteration The alternative initial water depth at the location; and These represent the initial water depths on the left and right sides of the step section, respectively. for The center position of the directional step section; For the first Steepness scheduling factor during round iteration ; This is the current training iteration round; In the Cartesian coordinate system The value of direction; along The hyperbolic tangent function for the direction is as follows: ; In the formula, Represents the hyperbolic tangent function. For the first Position during round iteration The alternative initial water depth at the location; and These represent the initial water depths on the left and right sides of the step section, respectively. for The center position of the directional step section. In the Cartesian coordinate system The value of the direction.
[0012] S024. In the initial stage of training, By forcing the initial conditions to a minimum value, the discontinuous initial conditions are smoothed into a gradually changing transition state.
[0013] Optionally, in some embodiments of this application, as the number of training rounds progresses toward a set total number of rounds, the amplification is automatically performed according to an exponential law or a piecewise linear law. The value, until Approaching the preset threshold The hyperbolic tangent function gradually approximates the original true target state, enabling the course learning to dynamically update the alternative training target as the training progresses, and gradually approaching the discontinuous initial conditions.
[0014] Optionally, in some embodiments of this application, the Fourier high-frequency feature encoding layer of the complex shock wave capture model is used for dimensionality upscaling and submitted to the backbone forward physical reconstruction network of the complex shock wave capture model to complete the forward propagation of the macroscopic conservation prediction value of the fluid, including the following steps: S031, For the basic internal collocation set The coordinate vector of any point in Pre-distributed from a zero-mean Gaussian distribution Extract and solidify a non-trainable weight matrix. Inputting low-dimensional spacetime Transform into high-frequency feature tensors ,in , ; In the formula, Each element is composed of Extract and fix independently. ; S032, Provide a layer number of Neuron width is The backbone forward physical reconstruction network employs the Swish activation function, which has a globally smooth analytical solution. This backbone forward physical reconstruction network receives high-frequency feature tensors. Then, the tensor of the continuous macroscopic prediction of shallow water fluid throughout the entire spatiotemporal domain is output. In the formula, The weights and biases of all neurons within the representation system to be optimized. To predict water depth, and They are the predicted ones Directional momentum and Directional velocity, To predict the end time of the time window, In the Cartesian coordinate system The value of direction, In the Cartesian coordinate system The value of direction, These are values on the timeline.
[0015] Optionally, in some embodiments of this application, an adaptive spatiotemporal resampling evaluator for complex shock wave capture models is used to evaluate the macroscopic prediction tensor of shallow water fluids. Real-time physical imbalance measurement is performed, shock wave abrupt change regions in the flow field are dynamically captured, and samples are added to the basic internal colligative point set to obtain the augmented basic internal colligative point set. It includes the following steps: S041, Tracking the shallow water fluid macroscopic prediction tensor output by the forward physical reconstruction network. ,right Regarding the original spacetime coordinates The evolution trajectory is analyzed using a chain rule. Substituting the obtained first-order spatiotemporal derivatives into the shallow water wave control equations, the mass conservation residuals and momentum conservation residuals are reconstructed. The mass conservation residual of the continuity equation is expressed as: ; The momentum conservation residuals in the x and y directions are expressed as follows: ; ; In the formula, The weights and biases of all neurons within the representation system to be optimized. To predict water depth, and They are the predicted ones Directional momentum and Directional momentum The elevation of the bed surface; The mass-conserved residuals for the continuity equation; and They are respectively direction and Momentum conservation residuals in the direction; For gravitational acceleration, the above-mentioned mass conservation residuals and momentum conservation residuals are collectively referred to as physical residuals. They are aggregated point by point to form the full residual norm, which is used in subsequent steps to construct the resampling probability density function and the physical residual loss function. S042. In each training iteration, the above-mentioned mass conservation residual and momentum conservation residual are reconstructed in real time. After each fixed iteration cycle is completed... Subsequently, the adaptive spatiotemporal resampling estimator for the complex shock wave capture model is used in the global extrapolation domain. It features an ultra-high-density evaluation grid and obtains the full residual norm of each point on the evaluation grid through low-cost parallel inference. Based on this full residual norm, a spatiotemporal resampling probability density function is constructed. : ; In the formula, This represents the spatiotemporal resampling probability density function; The norm of the total residual is represented by the point-by-point aggregation of the mass conservation residual and the momentum conservation residual reconstructed by S041. The exponential parameter controls the focusing intensity of the extreme shock wave surface; the denominator represents... Within range Perform integral normalization; S043, According to the probability density function In the global simulation domain Non-uniform Monte Carlo sampling was performed to accurately locate the shock wave rupture point and the nonlinear strong swirling zone at the fluid interface where the mass conservation residual and momentum conservation residual are maximized, and to extract... A new high-error coordinate point was established, and... A new high-error coordinate point dynamic reverse fusion is added to the basic internal point set. In the process, obtain the augmented set of fundamental internal collocations. .
[0016] Optionally, in some embodiments of this application, based on , and A global composite energy loss function is constructed. The LRAnneal global gradient weight scheduling engine is used to calculate and adaptively adjust the loss weights of various physical constraints within the global composite energy loss function. This drives a dual optimizer for hybrid iteration, achieving convergence of the complex shock wave capture model and reconstructing the global high-frequency spatiotemporal dynamic field, including: S051. Construct the global composite energy loss function as follows: ; In the formula, The initial conditional loss function; The boundary condition loss function; The loss function for the mass conservation residual and momentum conservation residual of the shallow water wave equation is the physical residual loss function. These are the weights of the initial conditional loss function; These are the weights of the boundary condition loss function; The weights of the loss functions for the mass conservation residuals and momentum conservation residuals of the shallow water wave equation are the weights of the physical residual loss function. S052. Before performing parameter updates on the global composite energy loss function, select a subset of parameters from the last layer or shared hidden layer of the backbone forward physical reconstruction network. Calculate the loss functions for mass conservation residuals and momentum conservation residuals, initial condition loss function, and boundary condition loss function relative to the shallow water wave equation. gradient ,in i ∈{PDE,IC,BC}; S053, Maximum absolute value of the gradient of the physical residual loss function Based on, Calculate the current option reestimation value at the e-th weight update: ; In the formula, The parameter subset of the last layer or shared hidden layer of the backbone forward physical reconstruction network; This represents the maximum absolute value of the gradient of the physical residual loss function; The physical residual loss function, initial condition loss function, and boundary condition loss function are relative to... The gradient; The loss weights are then updated using an exponentially smoothed moving average: ; in, To control the hyperparameters of memory decay, and The first Second and third The loss weight for the next update; By forcibly scaling, the gradients of various external physical constraints are kept at the same energy level, thus resolving ill-conditioned gradient competition. The updated loss weights are then substituted into the aforementioned global composite energy loss function. S054. Pre-training is performed using the Adam first-order adaptive optimizer. During the Adam phase, after every M preset parameter updates, the adaptive spatiotemporal resampling estimator is invoked to recalculate the physical residual loss function for the current backbone forward physical reconstruction network prediction field and augment the basic internal collocation set. Subsequently based on the augmented Continue Adam training to form a closed-loop process of residual evaluation, collocation augmentation, gradient weight update, and retraining. After the Adam phase reaches the preset M parameter updates, fix the internal collocation set and switch to the L-BFGS quasi-Newton optimizer. Use finite memory to approximate the second-order curvature information for fine optimization. When the L-BFGS iteration count reaches the maximum iteration count, end the model training and fix the final model parameters. S055. Input any desired three-dimensional spatiotemporal coordinate point set (x,y,t), and output the predicted water depth, x-direction flow velocity, and y-direction flow velocity at the corresponding locations in parallel, i.e., the global high-resolution spatiotemporal dynamic field.
[0017] Accordingly, embodiments of this application also provide a complex shock wave capture system based on a multi-strategy fusion physical information neural network, including: A model building module is used to construct a complex shock wave capture model based on a multi-strategy fusion physical information neural network; The configuration module for the spatiotemporal fundamental point set is used to configure the spatiotemporal fundamental point set for training complex shock wave capture models. This smooths discontinuous initial conditions into a gradually changing transition state. The spatiotemporal fundamental point set includes a set of collocation points within the fundamental system. Boundary point set With the initial point set ; The forward propagation module uses a complex shock wave capture model to perform up-dimensional mapping and completes the forward propagation of the macroscopic state prediction values of shallow water fluids. The augmented fundamental internal colligative point set module utilizes a complex shock wave capture model to obtain the augmented fundamental internal colligative point set. ; Output module, based on , and A global composite energy loss function is constructed, and the loss weights of various physical constraints of the global composite energy loss function are adaptively adjusted. Through hybrid iteration, the complex shock wave capture model is converged and the output global high-frequency spatiotemporal dynamic field is reconstructed.
[0018] Accordingly, embodiments of this application also provide a computer device, including a storage device and a processor, wherein the storage device stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.
[0019] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.
[0020] In this application: LRAnneal: Learning Rate Annealing; In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: (1) To address the training initiation difficulties caused by discontinuous initial conditions, the data spatiotemporal configuration and course loading module does not change the given original step initial value, but instead constructs a continuously differentiable course learning alternative training objective on the initial point set. In the early stage of training, a relatively gentle transition is adopted to reduce the fitting difficulty of local mutations, and then the steepness scheduler in the module gradually narrows the transition band, so that the model gradually shifts from a smooth initial field to the original strongly discontinuous initial field, taking into account both training initiation and step profile approximation.
[0021] (2) To address the problem that ordinary fully connected networks are insufficient in representing high-frequency changes, the Fourier high-frequency feature coding layer maps low-dimensional spatiotemporal coordinates into feature tensors containing different frequency changes, which are then passed to the backbone forward physical reconstruction network to predict water depth and flow velocity. This coding method, combined with the asymptotic steepening of the initial value, enables the network to have a corresponding high-frequency representation basis when the target transition gradually becomes steeper, which helps to reduce the situation where the local wavefront is over-smoothed.
[0022] (3) To address the issue of insufficient coverage of the shock wave region by fixed collocation points, the adaptive spatiotemporal resampling evaluator directly calculates the mass conservation residual and momentum conservation residual using the current prediction field, and constructs the resampling probability density function based on their aggregation results. New collocation points are preferentially assigned to spatiotemporal regions with larger physical residuals and merged with existing collocation points to form the current collocation. This allows the limited number of newly added collocation points to more concentratedly constrain local regions that have not yet fully satisfied the governing equations, rather than indiscriminately increasing the number of collocation points across the entire domain.
[0023] (4) To address the issue of inconsistent gradient contributions from different physical constraints, the LRAnneal global gradient weight scheduling engine evaluates the gradient magnitude of each loss term based on the same subset of network parameters, and updates and smooths the weights of each loss term. This scheduling does not simply assign fixed weights to the initial loss, but adjusts the relative contributions of each constraint in parameter updates according to the training state, which helps to reduce the situation where a certain loss term dominates the update for a long time while other constraints are not adequately improved.
[0024] (5) The above strategy forms a continuous connection in terms of data and optimization process: the course learning updates the training objective at the initial point, the Fourier encoder and the backbone network generate the current prediction field, and the physical residual of the current prediction field determines the position of the newly added collocation points in the next round; the updated initial objective, boundary points and augmented internal collocation points jointly participate in the composite loss calculation, and their gradient changes are used to update the loss weights. The updated network further changes the residual distribution and affects subsequent sampling, thereby linking the initial objective, feature representation, collocation distribution and loss weights into the same optimization closed loop, rather than using multiple training techniques independently. Attached Figure Description
[0025] Figure 1 This is a flowchart of the complex shock wave capture method based on a multi-strategy fusion physical information neural network according to the present invention; Figure 2 This is a two-dimensional application example of the present invention—the spatiotemporal evolution of a planar water column descent and wavefront diffusion. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0027] The technical solution of this application is as follows: Firstly, please refer to Figure 1 This application provides a complex shock wave capture method based on a multi-strategy fusion physical information neural network, including: S01. Construct a complex shock wave capture model based on a multi-strategy fusion physical information neural network; S02. Configure a spatiotemporal baseline point set for training the complex shock wave capture model, smoothing discontinuous initial conditions into a gradually changing transition state. The spatiotemporal baseline point set includes the set of coordinate points within the baseline. Boundary point set With the initial point set ; S03. Use a complex shock wave capture model to perform up-dimensional mapping and complete the forward propagation of the macroscopic state prediction values of shallow water fluids. S04. Using a complex shock wave capture model, obtain the augmented set of fundamental internal collocation points. ; S05, based on , and A global composite energy loss function is constructed, and the loss weights of various physical constraints of the global composite energy loss function are adaptively adjusted. Through hybrid iteration, the complex shock wave capture model is converged and the output global high-frequency spatiotemporal dynamic field is reconstructed.
[0028] In S01: In some embodiments, the complex shock wave capture model includes a data spatiotemporal configuration and course loading module, a Fourier high-frequency feature encoding layer, a backbone forward physical reconstruction network, an adaptive spatiotemporal resampling evaluator, and an LRAnneal global gradient weight scheduling engine.
[0029] In S02: In some embodiments, the spatiotemporal fundamental point set includes a fundamental internal collocation set. Boundary point set With the initial point set .
[0030] Furthermore, utilizing the data spatiotemporal configuration and course loading module of the complex shock wave capture model, a spatiotemporal basis point set is configured for training the complex shock wave capture model, smoothing the discontinuous initial conditions into a gradually changing transition state, including the following steps: S021. Determine the spatial threshold of the study. And the projection time window [0,T], where , It is a two-dimensional real space; the Latin hypercube sampling algorithm is used in the complete spatiotemporal domain. Internally generated set of internal collocation points for physical residual monitoring At the same time along the physical boundary disjunctive boundary point set Determine the initial point set along the time boundary, i.e., the initial time. ; S022, Set the target's true state at time zero as... ,in The initial water depth, and They are respectively direction and Initial velocity in direction, To extrapolate the end time of the time window, we extract the initial conditions that are discontinuous in the simulation of tsunami caused by extreme dam failure or fault displacement. S023, with To ultimately approach the target, in the initial point set We introduce a dynamically steepened hyperbolic tangent function to construct alternative training objectives for course learning, along... The hyperbolic tangent function for the direction is as follows: ; In the formula, Represents the hyperbolic tangent function. For the first Position during round iteration The alternative initial water depth at the location; and These represent the initial water depths on the left and right sides of the step section, respectively. for The center position of the directional step section; For the first Steepness scheduling factor during round iteration ; This is the current training iteration round; In the Cartesian coordinate system The value of direction; along The hyperbolic tangent function for the direction is as follows: ; In the formula, Represents the hyperbolic tangent function. For the first Position during round iteration The alternative initial water depth at the location; and These represent the initial water depths on the left and right sides of the step section, respectively. for The center position of the directional step section; In the Cartesian coordinate system The value of the direction.
[0031] S024. In the initial stage of training, By forcing the initial conditions to a minimum value, the discontinuous initial conditions are smoothed into a gradually changing transition state.
[0032] Furthermore, as the number of training rounds approaches the set total number of rounds, the amplification is automatically performed according to the exponential or piecewise linear law. The value, until Approaching the preset threshold The hyperbolic tangent function gradually approximates the original true target state, enabling the course learning to dynamically update the alternative training target as the training progresses, and gradually approaching the discontinuous initial conditions.
[0033] It is understandable that the discontinuous initial conditions are replaced with a hyperbolic tangent function that is continuously differentiable everywhere, which mathematically eliminates the step discontinuity at the initial time step (time boundary) and ensures stable startup of gradient-based training; as Approaching The course learning replaces the training objective to progressively restore the true step shape, so that the model can still approximate the true initial value of the shock wave with high accuracy.
[0034] It's understandable that the initial training phase involves multiple iterations. .
[0035] For example, Forced to a very small value, such as .
[0036] For example, in the 1D dam-break problem, the water depth is 10 when the value is less than 0, and 1 when the value is greater than 0.
[0037] It is understandable that the course learning substitutes the training objective, namely the target value given at each initial point by the dynamically degenerate hyperbolic tangent function constructed by S023.
[0038] In S03: Furthermore, the Fourier high-frequency feature encoding layer of the complex shock wave capture model is used for dimensionality upscaling, and then submitted to the backbone forward physical reconstruction network of the complex shock wave capture model to complete the forward propagation of the macroscopic conservation prediction values of the fluid, including the following steps: S031, For the basic internal collocation set The coordinate vector of any point in Pre-distributed from a zero-mean Gaussian distribution Extract and solidify a non-trainable weight matrix. Inputting low-dimensional spacetime Transform into high-frequency feature tensors ,in , ; In the formula, Each element is composed of Extract and fix independently. ; S032, Provide a layer number of Neuron width is The backbone forward physical reconstruction network employs the Swish activation function, which has a globally smooth analytical solution. This backbone forward physical reconstruction network receives high-frequency feature tensors. Then, the tensor of the continuous macroscopic prediction of shallow water fluid throughout the entire spatiotemporal domain is output. In the formula, The weights and biases of all neurons within the representation system to be optimized. To predict water depth, and They are the predicted ones Directional momentum and Directional velocity, To predict the end time of the time window, In the Cartesian coordinate system The value of direction, In the Cartesian coordinate system The value of direction, These are values on the timeline.
[0039] It's understandable that the variance hyperparameter... By manually or heuristically setting parameters based on the estimated shock wave intensity, computational domain scale, or empirical parameters, the network's ability to express high-frequency spatiotemporal variations and steep gradient structures can be improved, thereby alleviating the spectral bias of ordinary fully connected networks.
[0040] For example, It can be 6 to 10 floors, for example, it can be 6, 7, 8, 9, 10 floors, etc.; It is understandable that providing a layer number is... Neuron width is The fully connected network backbone is used. The ReLU function, whose second derivative is always zero, is abandoned in favor of the Swish activation function, which has a globally smooth analytical solution.
[0041] In S04: Furthermore, an adaptive spatiotemporal resampling evaluator based on a complex shock wave capture model is used to evaluate the macroscopic prediction tensor of shallow water fluids. Real-time physical imbalance measurement is performed, shock wave abrupt change regions in the flow field are dynamically captured, and samples are added to the basic internal colligative point set to obtain the augmented basic internal colligative point set. It includes the following steps: S041, Tracking the shallow water fluid macroscopic prediction tensor output by the forward physical reconstruction network. ,right Regarding the original spacetime coordinates The evolution trajectory is analyzed using a chain rule. Substituting the obtained first-order spatiotemporal derivatives into the shallow water wave control equations, the mass conservation residuals and momentum conservation residuals are reconstructed. The mass conservation residual of the continuity equation is expressed as: ; The momentum conservation residuals in the x and y directions are expressed as follows: ; ; In the formula, The weights and biases of all neurons within the representation system to be optimized. To predict water depth, and They are the predicted ones Directional momentum and Directional momentum The elevation of the bed surface; The mass-conserved residuals for the continuity equation; and These are the momentum conservation residuals in the x and y directions, respectively. For gravitational acceleration, the above-mentioned mass conservation residuals and momentum conservation residuals are collectively referred to as physical residuals. They are aggregated point by point to form the full residual norm, which is used in subsequent steps to construct the resampling probability density function and the physical residual loss function. S042. In each training iteration, the above-mentioned mass conservation residual and momentum conservation residual are reconstructed in real time. After each fixed iteration cycle is completed... Subsequently, the adaptive spatiotemporal resampling estimator for the complex shock wave capture model is used in the global extrapolation domain. It features an ultra-high-density evaluation grid and obtains the full residual norm of each point on the evaluation grid through low-cost parallel inference. Based on this full residual norm, a spatiotemporal resampling probability density function is constructed. : ; In the formula, This represents the spatiotemporal resampling probability density function; The norm of the total residual is represented by the point-by-point aggregation of the mass conservation residual and momentum conservation residual reconstructed by S041 (e.g., taking the square root of the sum of squares of each residual). The exponential parameter controls the focusing intensity of the extreme shock wave surface; the denominator represents... Within range Perform integral normalization; S043, According to the probability density function In the global simulation domain Non-uniform Monte Carlo sampling was performed to accurately locate the shock wave rupture point and the nonlinear strong swirling zone at the fluid interface where the mass conservation residual and momentum conservation residual are maximized, and to extract... A new high-error coordinate point was established, and... A new high-error coordinate point dynamic reverse fusion is added to the basic internal point set. In the process, obtain the augmented set of fundamental internal collocations. .
[0042] It is understandable that we obtain the augmented set of basic internal collocations. This is used to recalculate the mass conservation residual and momentum conservation residual in subsequent training rounds, and then substitute them into the global composite energy loss function constructed in step S05.
[0043] It is understandable that by densely stacking samples in the shock zone, the network is forced to allocate more capacity to characterize the steep contours.
[0044] It is understandable that the exponential parameter Used for adaptive control of the focusing intensity of extreme shock wave surfaces.
[0045] For example, M is 50 times or 100 times.
[0046] In S05: It can be understood that step S05 receives the initial point set generated in step S02. Boundary point set The course learning replaces the training objective, and receives the shallow water fluid macroscopic prediction tensor output from step S03. It also receives the mass conservation residual, momentum conservation residual, and augmented set of fundamental internal collocation points reconstructed in step S04. Step S05 organizes the above data into a closed-loop optimization process: First, S051 defines a global composite energy loss function based on the above input, which serves as the unified optimization objective for subsequent sub-steps; then, S052 evaluates the gradient magnitude of each loss branch with respect to the network parameters in each backpropagation, providing a quantitative basis for the adaptive adjustment of weights; next, S053 corrects and smoothly updates each loss weight based on the evaluation results of S052, which is then applied to the global composite energy loss function; then, S054, driven by the loss function after weight updates, uses a dual optimizer to perform mixed iterations on the network parameters until convergence; finally, S055 uses the converged model to reconstruct the output global high-frequency spatiotemporal dynamic field.
[0047] In some embodiments, based on , and A global composite energy loss function is constructed. The LRAnneal global gradient weight scheduling engine is used to calculate and adaptively adjust the loss weights of various physical constraints within the global composite energy loss function. This drives a dual optimizer for hybrid iteration, achieving convergence of the complex shock wave capture model and reconstructing the global high-frequency spatiotemporal dynamic field, including: S051. Construct the global composite energy loss function as follows: ; In the formula, The initial conditional loss function; The boundary condition loss function; The loss function for the mass conservation residual and momentum conservation residual of the shallow water wave equation is the physical residual loss function. These are the weights of the initial conditional loss function; These are the weights of the boundary condition loss function; The weights of the loss functions for the mass conservation residuals and momentum conservation residuals of the shallow water wave equation are the weights of the physical residual loss function. S052. Before performing parameter updates on the global composite energy loss function, select a subset of parameters from the last layer or shared hidden layer of the backbone forward physical reconstruction network. Calculate the loss functions for mass conservation residuals and momentum conservation residuals, initial condition loss function, and boundary condition loss function relative to the shallow water wave equation. gradient ,in i ∈{PDE,IC,BC}; S053, Maximum absolute value of the gradient of the physical residual loss function Based on, Calculate the current option reestimation value at the e-th weight update: ; In the formula, The parameter subset of the last layer or shared hidden layer of the backbone forward physical reconstruction network; This represents the maximum absolute value of the gradient of the physical residual loss function; The physical residual loss function, initial condition loss function, and boundary condition loss function are relative to... The gradient; The loss weights are then updated using an exponentially smoothed moving average: ; in, To control the hyperparameters of memory decay, and The first Second and third The loss weight for the next update; By forcibly scaling, the gradients of various external physical constraints are kept at the same energy level, thus resolving ill-conditioned gradient competition. The updated loss weights are then substituted into the aforementioned global composite energy loss function. S054. Pre-training is performed using the Adam first-order adaptive optimizer. During the Adam phase, after every M preset parameter updates, the adaptive spatiotemporal resampling estimator is invoked to recalculate the physical residual loss function for the current backbone forward physical reconstruction network prediction field and augment the basic internal collocation set. Subsequently based on the augmented Continue Adam training to form a closed-loop process of residual evaluation, collocation augmentation, gradient weight update, and retraining. After the Adam phase reaches the preset M parameter updates, fix the internal collocation set and switch to the L-BFGS quasi-Newton optimizer. Use finite memory to approximate the second-order curvature information for fine optimization. When the L-BFGS iteration count reaches the maximum iteration count, end the model training and fix the final model parameters. S055. Input any desired three-dimensional spatiotemporal coordinate point set (x,y,t), and output the predicted water depth, x-direction flow velocity, and y-direction flow velocity at the corresponding locations in parallel, i.e., the global high-resolution spatiotemporal dynamic field.
[0048] Understandable. The initial conditional loss function is based on The mean square error between the predicted value at each initial point and the target value given by the course learning alternative training objective at that initial point is obtained; the course learning alternative training objective is not a single numerical value, but a target value given point by point at each initial point by the hyperbolic tangent function constructed by S023 and dynamically updated with each training round. For boundary condition loss function, based on The mean square error between the predicted values of each boundary point and the corresponding physical boundary conditions is obtained. The loss function is the mass conservation residual and momentum conservation residual of the shallow water wave equation, based on... The mean square error of the mass conservation residual and momentum conservation residual of each internal collocation point is obtained.
[0049] It is understandable that after step S054 is completed and the model parameters are fixed, any required three-dimensional spatiotemporal coordinate point set (x, y, t) can be input during the inference phase. Numerical integration at each time step is not required, and the predicted water depth, x-direction velocity, and y-direction velocity at the corresponding locations can be output in parallel. Based on this, the global spatiotemporal dynamic field can be reconstructed to describe dam-break wavefront propagation, reflection, and other strongly discontinuous and high-gradient shallow water dynamic structures. The aforementioned predicted water depth field, x-direction velocity field, and y-direction velocity field constitute the global high-resolution spatiotemporal dynamic field output in step S05.
[0050] For example, the Adam stage can sustain 15,000 to 20,000 parameter updates, allowing the backbone to physically reconstruct the network parameters into a better parameter region.
[0051] For example, M times is between 15,000 and 20,000 times.
[0052] For example, the maximum number of iterations for L-BFGS is set to 5000.
[0053] Secondly, embodiments of this application provide a complex shock wave capture system based on a multi-strategy fusion physical information neural network, including: A model building module is used to construct a complex shock wave capture model based on a multi-strategy fusion physical information neural network; The configuration module for the spatiotemporal fundamental point set is used to configure the spatiotemporal fundamental point set for training complex shock wave capture models. This smooths discontinuous initial conditions into a gradually changing transition state. The spatiotemporal fundamental point set includes a set of collocation points within the fundamental system. Boundary point set With the initial point set ; The forward propagation module uses a complex shock wave capture model to perform up-dimensional mapping and completes the forward propagation of the macroscopic state prediction values of shallow water fluids. The augmented fundamental internal colligative point set module utilizes a complex shock wave capture model to obtain the augmented fundamental internal colligative point set. ; Output module, based on , and A global composite energy loss function is constructed, and the loss weights of various physical constraints of the global composite energy loss function are adaptively adjusted. Through hybrid iteration, the complex shock wave capture model is converged and the output global high-frequency spatiotemporal dynamic field is reconstructed.
[0054] Thirdly, this application provides a computer device including a storage device and a processor. The storage device stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the complex shock wave capture method based on a multi-strategy fusion physical information neural network as described above.
[0055] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.
[0056] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device. Of course, the memory may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the complex shock wave capture method based on a multi-strategy fusion physical information neural network. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.
[0057] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, for example, to run program code for a complex shock wave capture method based on a multi-strategy fusion physical information neural network.
[0058] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the complex shock wave capture method based on a multi-strategy fusion physical information neural network as described above.
[0059] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to enable the at least one processor to perform the steps of the complex shock wave capture method based on a multi-strategy fusion physical information neural network as described above.
[0060] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the complex shock wave capture method based on multi-strategy fusion physical information neural network described in the embodiments of this application.
[0061] This application coordinates the progressive loading of initial conditions, high-frequency feature representation, local collocation configuration, and gradient contributions of multiple physical constraints within the same training process, enabling newly added collocations and updated initial targets to effectively participate in subsequent optimization, thereby improving the reconstruction capability of complex shock wave shallow water flow fields.
[0062] Application examples This application example uses the collapse of a water column and wavefront diffusion on a two-dimensional ideal plane as an example. Initially, there is a step-high water column in the central region. The method of this invention is used to predict the changes in water depth and flow velocity during the descent of the water column and the outward diffusion of the water wave.
[0063] Step 1 involves constructing a complex shock capture model consisting of a data spatiotemporal configuration and course loading module, a Fourier high-frequency feature encoding layer, a backbone forward physical reconstruction network, an adaptive spatiotemporal resampling estimator, and an LRAnneal global gradient weight scheduling engine. The steepness scheduler, located within the data spatiotemporal configuration and course loading module, controls the gradual loading of the initial water depth in this example.
[0064] Step 2, Determine Coverage Figure 2 The spatial domain of the propagation process shown and simulation time window Generate within the domain through Latin hypercube sampling And configure at the physical boundary and the initial time respectively. and For the step change at the boundary between the central high water level and the peripheral low water level, a hyperbolic tangent substitute target is constructed at each initial point based on the actual boundary location; a smaller target is used in the early stage of training. A gradual transition is formed, followed by a gradual increase according to a pre-set pattern. The initial velocity components are those given in this example. and It, together with the alternative initial water depth, participates in the initial condition loss calculation. Figure 2The initial field of the central water column is not constructed by directly adding two axially smooth expressions.
[0065] Step 3: Input the spatiotemporal coordinates of the internal collocation points, boundary points, and initial points into the same Fourier high-frequency feature encoding layer, and then output the results from the backbone feedforward physical reconstruction network using the Swish activation function. , and The predicted values at the internal collocation points are used to construct the residuals of the governing equations, while the predicted values at the initial and boundary points are used to construct the corresponding conditional losses.
[0066] Step 4: Perform automatic differentiation on the current predicted field according to S041 above, and calculate the mass conservation residual and the momentum conservation residual in both directions. For the ideal plane in this example, the bed elevation... As constant, the derivatives of the bed surface elevation in both directions are zero. The residuals are aggregated on the evaluation mesh and a model is constructed. Based on this probability distribution, additional points are added, and the newly added points are incorporated into the current distribution. Whether to add a collocation point near the wavefront and the location of the addition are determined by the current physical residual distribution, and a fixed shock wave collocation point region is not pre-specified.
[0067] Step 5, in The initial conditional loss between the predicted value and the current course learning alternative training objective is calculated. Calculate the boundary condition loss on the current... The physical residual loss is calculated; the gradient contribution of each loss is evaluated and adjusted using the LRAnneal global gradient weight scheduling engine; and the network parameters are updated based on the composite loss.
[0068] For example, the Adam phase performs S04 resampling and collocation augmentation once every 100 parameter updates, i.e. Then, the physical residual loss is recalculated on the augmented collocation set, and training continues. After a total of 15,000 Adam parameter updates, the current parameters are fixed. Then switch to the L-BFGS quasi-Newton optimizer, with a maximum number of iterations set to 5000. The aforementioned 100 iterations are the resampling cycles, and 15000 iterations are the cumulative parameter update counts in the Adam stage. These two control the local collocation update frequency and the pre-training duration, respectively.
[0069] After training, the model parameters are fixed, and Figure 2 The coordinates shown indicate the required spatiotemporal coordinates within the display range, and the corresponding predicted water depth and flow velocity in two directions are obtained in parallel. Figure 2 exhibit , , and The water depth distribution at that time is used to illustrate the process of the central water column descending and the higher water depth areas expanding outward; the figure shows the output water depth field of S055 at the above time.
[0070] Figure 2 This paper presents a sequence of thermal maps showing the water depth distribution as a water column collapses (dam breaks) and evolves outward on a two-dimensional ideal plane. Initially, a step-high water column (dark red) exists in the central region, forming an extremely strong discontinuous shock wave initial field. As time progresses, the water waves exhibit isotropic radial diffusion, forming a steep, circular shock wave front on the periphery, accompanied by the evolution of sparse waves inside. Figure 2 It is intuitively demonstrated that after introducing Fourier feature encoding and adaptive sampling, the model can still sharply capture the strongly nonlinear two-dimensional shock surface in a completely meshless framework, and the flow field contour lines are smooth and natural, completely overcoming the shock smoothing defect of the basic PINN model.
[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A complex shock wave capture method based on a multi-strategy fusion physical information neural network, characterized in that, include: S01. Construct a complex shock wave capture model based on a multi-strategy fusion physical information neural network; S02. Configure a spatiotemporal baseline point set for training the complex shock wave capture model, smoothing discontinuous initial conditions into a gradually changing transition state. The spatiotemporal baseline point set includes the set of coordinate points within the baseline. Boundary point set With the initial point set ; S03. Use a complex shock wave capture model to perform up-dimensional mapping and complete the forward propagation of the macroscopic state prediction values of shallow water fluids. S04. Using a complex shock wave capture model, obtain the augmented set of fundamental internal collocation points. ; S05, based on , and A global composite energy loss function is constructed, and the loss weights of various physical constraints of the global composite energy loss function are adaptively adjusted. Through hybrid iteration, the complex shock wave capture model is converged and the output global high-frequency spatiotemporal dynamic field is reconstructed.
2. The complex shock wave capture method based on a multi-strategy fusion physical information neural network according to claim 1, characterized in that, The complex shock wave capture model includes a data spatiotemporal configuration and course loading module, a Fourier high-frequency feature encoding layer, a backbone forward physical reconstruction network, an adaptive spatiotemporal resampling evaluator, and an LRAnneal global gradient weight scheduling engine.
3. The complex shock wave capture method based on a multi-strategy fusion physical information neural network according to claim 2, characterized in that, Using the data spatiotemporal configuration and course loading module of the complex shock wave capture model, a spatiotemporal basis point set is configured for training the complex shock wave capture model, smoothing the discontinuous initial conditions into a gradually changing transition state, including the following steps: S021. Determine the spatial threshold of the study. And the projection time window [0,T], where , It is a two-dimensional real space; the Latin hypercube sampling algorithm is used in the complete spatiotemporal domain. Internally generated set of internal collocation points for physical residual monitoring At the same time along the physical boundary disjunctive boundary point set Determine the initial point set along the time boundary, i.e., the initial time. ; S022, Set the target's true state at time zero as... ,in The initial water depth, and They are respectively direction and Initial velocity in direction, To extrapolate the end time of the time window, we extract the initial conditions that are discontinuous in the simulation of tsunami caused by extreme dam failure or fault displacement. S023, with To ultimately approach the target, in the initial point set We introduce a dynamically steepened hyperbolic tangent function to construct alternative training objectives for course learning, along... The hyperbolic tangent function for the direction is as follows: ; In the formula, Represents the hyperbolic tangent function. For the first Position during round iteration The alternative initial water depth at the location; and These represent the initial water depths on the left and right sides of the step section, respectively. for The center position of the directional step section; For the first Steepness scheduling factor during round iteration ; This is the current training iteration round; In the Cartesian coordinate system The value of direction; along The hyperbolic tangent function for the direction is as follows: ; In the formula, Represents the hyperbolic tangent function. For the first Position during round iteration The alternative initial water depth at the location; and These represent the initial water depths on the left and right sides of the step section, respectively. for The center position of the directional step section; In the Cartesian coordinate system The value of direction; S024. In the initial stage of training, By forcing the initial conditions to a minimum value, the discontinuous initial conditions are smoothed into a gradually changing transition state.
4. The complex shock wave capture method based on a multi-strategy fusion physical information neural network according to claim 3, characterized in that, As the number of training rounds approaches the set total number of rounds, the system automatically scales up according to either an exponential or piecewise linear law. The value, until Approaching the preset threshold The hyperbolic tangent function gradually approximates the original true target state, enabling the course learning to dynamically update the alternative training target as the training progresses, and gradually approaching the discontinuous initial conditions.
5. The complex shock wave capture method based on a multi-strategy fusion physical information neural network according to claim 2, characterized in that, The Fourier high-frequency feature encoding layer of the complex shock wave capture model is used for dimensionality upscaling, and then submitted to the backbone forward physics reconstruction network of the complex shock wave capture model to complete the forward propagation of the macroscopic conservation prediction values of the fluid. The steps include the following: S031, For the basic internal collocation set The coordinate vector of any point in Pre-distributed from a zero-mean Gaussian distribution Extract and solidify a non-trainable weight matrix. Inputting low-dimensional spacetime Transform into high-frequency feature tensors ,in , ; In the formula, Each element is composed of Extract and fix independently. ; S032, Provide a layer number of Neuron width is The backbone forward physical reconstruction network employs the Swish activation function, which has a globally smooth analytical solution. This backbone forward physical reconstruction network receives high-frequency feature tensors. Then, the tensor of the continuous macroscopic prediction of shallow water fluid throughout the entire spatiotemporal domain is output. In the formula, The weights and biases of all neurons within the representation system to be optimized. To predict water depth, and They are the predicted ones Directional momentum and Directional velocity, To predict the end time of the time window, In the Cartesian coordinate system The value of direction, In the Cartesian coordinate system The value of direction, These are values on the timeline.
6. The complex shock wave capture method based on a multi-strategy fusion physical information neural network according to claim 5, characterized in that, An adaptive spatiotemporal resampling evaluator based on a complex shock wave capture model is used for the macroscopic prediction tensor of shallow water fluids. Real-time physical imbalance measurement is performed, shock wave abrupt change regions in the flow field are dynamically captured, and samples are added to the basic internal colligative point set to obtain the augmented basic internal colligative point set. It includes the following steps: S041, Tracking the shallow water fluid macroscopic prediction tensor output by the forward physical reconstruction network. ,right Regarding the original spacetime coordinates The evolution trajectory is analyzed using a chain rule. Substituting the obtained first-order spatiotemporal derivatives into the shallow water wave control equations, the mass conservation residuals and momentum conservation residuals are reconstructed. The mass conservation residual of the continuity equation is expressed as: ; The momentum conservation residuals in the x and y directions are expressed as follows: ; ; In the formula, The weights and biases of all neurons within the representation system to be optimized. To predict water depth, and They are the predicted ones Directional momentum and Directional momentum The elevation of the bed surface; The mass-conserved residuals for the continuity equation; and They are respectively direction and Momentum conservation residuals in the direction; For gravitational acceleration, the above-mentioned mass conservation residuals and momentum conservation residuals are collectively referred to as physical residuals. They are aggregated point by point to form the full residual norm, which is used in subsequent steps to construct the resampling probability density function and the physical residual loss function. S042. In each training iteration, the above-mentioned mass conservation residual and momentum conservation residual are reconstructed in real time. After each fixed iteration cycle is completed... Subsequently, the adaptive spatiotemporal resampling estimator for the complex shock wave capture model is used in the global extrapolation domain. It features an ultra-high-density evaluation grid and obtains the full residual norm of each point on the evaluation grid through low-cost parallel inference. Based on this full residual norm, a spatiotemporal resampling probability density function is constructed. : ; In the formula, This represents the spatiotemporal resampling probability density function; The norm of the total residual is represented by the point-by-point aggregation of the mass conservation residual and the momentum conservation residual reconstructed by S041. The exponential parameter controls the focusing intensity of the extreme shock wave surface; the denominator represents... Within the range of ×[0,T] Perform integral normalization; S043, According to the probability density function In the global simulation domain Non-uniform Monte Carlo sampling was performed to accurately locate the shock wave rupture point and the nonlinear strong swirling zone at the fluid interface where the mass conservation residual and momentum conservation residual are maximized, and to extract... A new high-error coordinate point was established, and... A new high-error coordinate point dynamic reverse fusion is added to the basic internal point set. In the process, obtain the augmented set of fundamental internal collocations. .
7. The complex shock wave capture method based on a multi-strategy fusion physical information neural network according to claim 6, characterized in that, based on , and A global composite energy loss function is constructed. The LRAnneal global gradient weight scheduling engine is used to calculate and adaptively adjust the loss weights of various physical constraints within the global composite energy loss function. This drives a dual optimizer for hybrid iteration, achieving convergence of the complex shock wave capture model and reconstructing the global high-frequency spatiotemporal dynamic field, including: S051. Construct the global composite energy loss function as follows: ; In the formula, The initial conditional loss function; The boundary condition loss function; The loss function for the mass conservation residual and momentum conservation residual of the shallow water wave equation is the physical residual loss function. These are the weights of the initial conditional loss function; These are the weights of the boundary condition loss function; The weights of the loss functions for the mass conservation residuals and momentum conservation residuals of the shallow water wave equation are the weights of the physical residual loss function. S052. Before performing parameter updates on the global composite energy loss function, select a subset of parameters from the last layer or shared hidden layer of the backbone forward physical reconstruction network. Calculate the loss functions for mass conservation residuals and momentum conservation residuals, initial condition loss function, and boundary condition loss function relative to the shallow water wave equation. gradient ,in i ∈{PDE, IC,BC}; S053, Maximum absolute value of the gradient of the physical residual loss function Based on, Calculate the current option reestimation value at the e-th weight update: ; In the formula, The parameter subset of the last layer or shared hidden layer of the backbone forward physical reconstruction network; This represents the maximum absolute value of the gradient of the physical residual loss function; The physical residual loss function, initial condition loss function, and boundary condition loss function are relative to... The gradient; The loss weights are then updated using an exponentially smoothed moving average: ; in, To control the hyperparameters of memory decay, and The first Second and third The loss weight for the next update; By forcibly scaling, the gradients of various external physical constraints are kept at the same energy level, thus resolving ill-conditioned gradient competition. The updated loss weights are then substituted into the aforementioned global composite energy loss function. S054. Pre-training is performed using the Adam first-order adaptive optimizer. During the Adam phase, after every M preset parameter updates, the adaptive spatiotemporal resampling estimator is invoked to recalculate the physical residual loss function for the current backbone forward physical reconstruction network prediction field and augment the basic internal collocation set. Subsequently based on the augmented Continue Adam training to form a closed-loop process of residual evaluation, collocation augmentation, gradient weight update, and retraining. After the Adam phase reaches the preset M parameter updates, fix the internal collocation set and switch to the L-BFGS quasi-Newton optimizer. Use finite memory to approximate the second-order curvature information for fine optimization. When the L-BFGS iteration count reaches the maximum iteration count, end the model training and fix the final model parameters. S055. Input any desired three-dimensional spatiotemporal coordinate point set (x,y,t), and output the predicted water depth, x-direction flow velocity, and y-direction flow velocity at the corresponding locations in parallel, i.e., the global high-resolution spatiotemporal dynamic field.
8. A complex shock wave capture system based on a multi-strategy fusion physical information neural network, characterized in that, include: A model building module is used to construct a complex shock wave capture model based on a multi-strategy fusion physical information neural network; The configuration module for the spatiotemporal fundamental point set is used to configure the spatiotemporal fundamental point set for training complex shock wave capture models. This smooths discontinuous initial conditions into a gradually changing transition state. The spatiotemporal fundamental point set includes a set of collocation points within the fundamental system. Boundary point set With the initial point set ; The forward propagation module uses a complex shock wave capture model to perform up-dimensional mapping and completes the forward propagation of the macroscopic state prediction values of shallow water fluids. The augmented fundamental internal colligative point set module utilizes a complex shock wave capture model to obtain the augmented fundamental internal colligative point set. ; Output module, based on , and A global composite energy loss function is constructed, and the loss weights of various physical constraints of the global composite energy loss function are adaptively adjusted. Through hybrid iteration, the complex shock wave capture model is converged and the output global high-frequency spatiotemporal dynamic field is reconstructed.
9. A computer device, characterized in that, It includes a storage device and a processor, the storage device storing a computer program that, when executed by the processor, causes the processor to perform the steps of the complex shock wave capture method based on a multi-strategy fusion physical information neural network as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, causes the processor to perform the steps of the complex shock wave capture method based on a multi-strategy fusion physical information neural network as described in any one of claims 1-7.
