A neural operator fast prediction method for multi-stage explosion simulation

Abstract

The application provides a neural operator fast prediction method for multi-stage penetration and explosion simulation, and belongs to the technical field of impact dynamics simulation and artificial intelligence agent modeling. The method constructs a conditional vector carrying working condition and penetration and explosion stage distinguishing information, and embeds the conditional vector into a potential query vector of multi-head cross attention. The network can distinguish the alternating penetration and explosion stages, autonomously learn the exclusive damage evolution law of each stage, and relieve the continuous accumulation of prediction errors in the multi-round alternating simulation process, so as to adapt to the cyclic penetration and explosion multi-stage simulation scene. At the same time, since the conditional vector is directly superimposed on the learnable potential query vector and explicitly participates in the core operation of the attention layer, the risk of weakening of the conditional vector in the deep transmission process can be reduced. Using the application can effectively improve the prediction accuracy and calculation efficiency of structure damage evolution in the multi-stage penetration and explosion scene.

Description

Technical Field

[0001] This invention relates to the fields of impact dynamics simulation and artificial intelligence agent modeling technology, specifically to a fast prediction method for neural operators for multi-stage blast simulation. Background Technology

[0002] Multi-stage blasting is an important technical method in engineering such as rock and soil blasting and concrete fracturing. It involves a multi-stage process of penetration, explosion, and re-penetration. The entire process is accompanied by complex physical effects such as high-speed response, large structural deformation, material nonlinearity, local failure, strain concentration, and cross-stage damage transmission.

[0003] In existing engineering analysis, explicit dynamic finite element software such as LS-DYNA is usually used to perform high-fidelity numerical simulations of the above process. However, it has high computational cost and small time step. When performing batch analysis for different impact parameters, a large number of repeated simulations are usually required, which makes it difficult to meet the needs of rapid evaluation of multiple working conditions, parameter optimization design and real-time auxiliary decision-making.

[0004] In recent years, deep learning surrogate models have been gradually used to replace finite element simulation calculations and accelerate simulation result prediction. Mainstream models include convolutional neural networks, graph neural networks, Transformer models, and neural operator models. Among these, neural operator models are mostly used for regular mesh fields and continuous function mapping problems. However, they have significant shortcomings when dealing with stage transitions in the blasting process, damage concentration areas, and large-scale irregular node fields, and cannot adapt to alternating, multi-stage blasting simulation scenarios. Specific shortcomings are as follows: 1. The model uses a single network to fit all alternating time-series data, without distinguishing whether the current calculation step is in the penetration stage or the explosion stage, and cannot implement differentiated constraints on the prediction logic of different stages; after multiple rounds of alternating superposition, the prediction error continues to accumulate, making it difficult to accurately describe the differentiated damage evolution law of each stage in the multi-stage penetration and explosion process.

[0005] 2. Traditional latent space neural operators only express distinguishing information based on the working conditions and use it as shallow input. After being transformed by multiple layers of networks, it is sent to inner modules such as multi-head attention to participate in model optimization. Multi-layer mapping weakens the ability to represent distinguishing information and cannot effectively highlight the inherent differences between different working conditions and different stages. As a result, the multi-head attention mechanism cannot match the feature extraction logic specific to each blasting stage. The model has difficulty learning the damage evolution law specific to each stage. After multiple rounds of alternating blasting simulation, the error accumulates, and the model's generalization ability and prediction accuracy decrease significantly.

[0006] 3. Existing methods employ a globally uniform node sampling strategy, failing to differentiate between penetration and explosion processes, and using the same set of sampling rules to select structural units and nodes for prediction. However, the main damage mechanisms controlling each stage of multi-stage penetration and explosion are completely different: the penetration stage is mainly characterized by the generation of penetration channels along the projectile's forward direction; the explosion stage is characterized by the violent release of energy in the near-field region of the explosion core, the radial spread of the shock wave, and a significant expansion of the existing damage zone. If uniform random sampling is used throughout, the model's computational power will be consumed by a large number of low-damage background nodes, resulting in insufficient learning of features in high-strain, high-gradient, and stage-interaction core regions, ultimately reducing the prediction accuracy of key damage areas.

[0007] 4. Existing prediction networks all adopt a single-network integrated output architecture, relying on a single network to synchronously output two types of physical fields: equivalent plastic strain and nodal displacement. They do not consider that plastic strain is a preliminary intermediate field that induces structural geometric deformation. Strain prediction errors will be directly transmitted to displacement prediction results without obstruction, and the prediction accuracy of the two types of physical fields will be coupled and interfere with each other.

[0008] In summary, existing prediction methods have several shortcomings: First, they cannot distinguish between alternating penetration and explosion processes, lack differentiated constraints for each stage, and the characteristics of working conditions and stage identifiers decay after being mapped through multiple layers of networks. Furthermore, multi-head attention mechanisms struggle to extract the unique evolutionary features of each process. Second, the use of an indiscriminate node sampling strategy throughout the process dilutes the core damage features with redundant background nodes and consumes significant model computational resources. Third, the single-network integrated output architecture causes prediction errors in plastic strain and nodal displacement fields to propagate and couple with each other. These multiple problems collectively result in low overall prediction accuracy, making the model unsuitable for multi-step alternating penetration and explosion simulations and unable to provide high-precision, efficient, and rapid predictions of structural damage evolution. Summary of the Invention

[0009] In view of this, the present invention provides a fast prediction method for neural operators for multi-stage blast simulation, which can effectively improve the prediction accuracy and computational efficiency of structural damage evolution in multi-stage blast scenarios.

[0010] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0011] A fast neural operator prediction method for multi-stage blast simulation is proposed for structural damage evolution analysis in multi-stage blast processes. The method includes: Step 1: Under different working conditions, obtain structural damage evolution data for each frame of the multi-stage blasting process through explicit dynamic simulation; Step 2: Extract the node coordinates, coordinate increments, plastic strain, and strain increments of each node from the single-frame structural damage evolution data to construct damage features; stitch together the damage features from multiple frames to obtain the feature vector. Based on overall and single-frame structural damage evolution data, a conditional vector containing information distinguishing between working conditions and blast stages is constructed. ; using feature vectors and condition vector Use these samples as input to the prediction model to build a training sample set. Step 3: Train a two-level prediction network based on latent space neural operators; the first-level prediction network uses feature vectors and condition vector As input, it outputs the predicted strain increment and synthesizes the predicted plastic strain; in the eigenvector Add strain increment prediction and plastic strain prediction to form a new eigenvector. The second-level prediction network uses feature vectors and condition vector As input, the network outputs coordinate increment predictions and synthesizes node coordinate predictions; moreover, both levels of the prediction network add condition vectors to the learnable latent query vectors. To increase the distinguishing information between working conditions and blasting stages, latent space marking with conditional constraints is obtained through multi-head cross-attention processing; Step 4: Inference stage, use the trained prediction model to predict structural damage evolution data.

[0012] Preferably, in step 2, the step of extracting the node coordinates, coordinate increments, plastic strain, and strain increments of each node from the single-frame structural damage evolution data to construct damage features specifically involves: According to the node sampling rules adapted to different blasting stages, sampling nodes are extracted from single-frame structural damage evolution data; the node coordinates, coordinate increments, plastic strains and strain increments of the sampling nodes are used to construct the damage features of a single frame. The node sampling rules include the sampling ratio and sampling principles, specifically: During the penetration phase, sampling nodes are extracted from the plastic strain concentration region, strain gradient localization region, penetration channel region, and other regions according to a set extraction ratio. Specifically, the extraction ratio of nodes in the plastic strain concentration region is 30% of the total number of sampling points S, the extraction ratio in the strain gradient localization region is 25% of the total number of sampling points S, the extraction ratio in the penetration channel region is 25% of the total number of sampling points S, and the extraction ratio in other regions is 20% of the total number of sampling points S. During the explosion phase, sampling nodes are extracted from the plastic strain concentration region, the strain gradient localization region, the explosion impact influence region, and other regions according to a set extraction ratio. Specifically, the extraction ratio of the plastic strain concentration region is 30% of the total number of sampling points S, the extraction ratio of the strain gradient localization region is 25% of the total number of sampling points S, the extraction ratio of the explosion impact influence region is 25% of the total number of sampling points S, and the extraction ratio of other regions is 20% of the total number of sampling points S. The principle for extracting the plastic strain concentration region is: to extract it according to the nodal plastic strain value from large to small; The principle for extracting the strain gradient localization region is: to extract it from large to small according to the comprehensive plastic strain gradient of the node neighborhood; The principle for extracting the penetration channel region is: nodes whose radial distance from the node to the penetration axis and axial distance along the penetration direction both meet the set range. The principle for selecting the impact zone of the explosion is: to select from nodes whose distance from the explosion center is less than or equal to the explosion radius.

[0013] Preferably, the condition vector middle: The operating condition differentiation information includes the operating condition identifier, the normalized value of the operating condition parameter, and the normalized value of the number of simulation nodes of the current frame's operating condition; The information distinguishing the intrusion phase includes the intrusion phase identifier of the current frame, the normalized value of the frame interval used by the current frame in this intrusion phase, and the normalized value of the duration of the intrusion phase of the current frame in the entire multi-stage intrusion process.

[0014] Preferably, in the two-level prediction network based on latent space neural operators, both the first-level prediction network and the second-level prediction network include a dimensionality-upgrading module, a latent space compression module, a latent space temporal evolution and decoding module, and an update module; The feature vectors and conditional vectors input to this level of the prediction network are both up-mapped by the up-dimensionality module and then added together to obtain the fused features of the latent space dimension. ; The latent space compression module introduces learnable latent query vectors. Add the condition vector after the dimensionality upgrade mapping to the latent query vector. This yields a condition constraint query vector with constraints on operating conditions and blasting stages. ; with fusion features Simultaneously serving as the key and value vectors for multi-head cross-attention, along with the conditional constraint query vector... Multi-head cross-attention computation is performed together. After multi-head concatenation, layer normalization, and residual connection, the loss features of the irregular node field in the current frame are compressed into a fixed number of latent space labels. Combined with historical window data, the latent space labels including the current frame and M-1 historical frames are combined to obtain the historical latent space label sequence. The data is sent to the latent space temporal evolution and decoding module; M is the length of the history window. The latent space temporal evolution and decoding module is used to learn the inter-frame evolution patterns of latent space marker sequences and to process the input historical latent space marker sequences. Mapped to the evolved latent space label sequence After decoding, the damage representation increment predicted by the current prediction network is output. The update module is used to add the damage representation increment predicted by the current prediction network to the absolute damage representation of the previous frame to obtain the absolute damage representation predicted by the current prediction network. For a first-level prediction network, the absolute value of the damage representation is plastic strain, and the damage representation increment is the strain increment; for a second-level prediction network, the absolute value of the damage representation is node coordinates, and the damage representation increment is the coordinate increment.

[0015] Preferably, the dimensionality-upgrading module adopts a multilayer perceptron (MLP); the latent space temporal evolution and decoding module adopts a Transformer network or a Mamba network.

[0016] Preferably, step 3 specifically includes: Step 301: Construct the feature vector of the current frame and condition vector Input the first-level prediction network and output the predicted strain increment for the next frame. ; Step 302: Calculate the predicted strain increment. Plastic strain in the previous frame Add them together to obtain the predicted amount of plastic strain for the next frame. Among them, the plastic strain in the previous frame Plastic strain obtained using historical predictions; Step 303: Calculate the predicted strain increment. and the predicted amount of plastic strain The value is then fused with the corresponding true value to obtain the fused strain increment prediction. And the predicted amount of plastic strain after fusion The weights for fusion of true values ​​gradually decrease from 1 to 0 as training progresses. Step 304: The feature vector corresponding to the current frame Add the fused strain increment prediction And the predicted amount of plastic strain after fusion To obtain new feature vectors ; to feature vector With condition vector The input is a two-level prediction network, which outputs the coordinate increment prediction for the next frame. ; Step 305: Calculate the predicted coordinate increment. Compared to the node coordinates of the previous frame Add them together to get the predicted node coordinates for the next frame. Among them, the node coordinates of the previous frame Node coordinates obtained using historical predictions.

[0017] Preferably, the first-level prediction network and the second-level prediction network are jointly trained and optimized, and the total loss consists of strain prediction loss and geometric prediction loss; The strain prediction loss is obtained by weighted fusion of strain Smooth L1 loss and monotonic penalty loss. and strain adaptive weighted loss get: The monotonic penalty loss Penalize instances where the plastic strain of adjacent frame nodes decreases; The strain adaptive weighted loss The absolute value of the prediction error of the nodal plastic strain is used as the basic constraint term. A first dynamic penalty weight is introduced that increases linearly with the true value of the nodal plastic strain, and the first dynamic penalty weight is increased for nodes with higher plastic strain. The geometric prediction loss is calculated using weighted fusion coordinate Smooth L1 loss and acceleration smoothing loss. and strain-geometric coupling consistency loss get: The acceleration smoothing loss The second-order difference magnitude of the predicted node coordinate increment over three consecutive frames is used to represent the value. The strain-geometric coupling consistency loss The absolute error between the predicted coordinate increment output by the secondary prediction network and the actual value of the corresponding coordinate increment is used as the basic constraint term. A second dynamic penalty weight is introduced that increases linearly with the plastic strain of the node, and the second dynamic penalty weight is increased for nodes with higher plastic strain.

[0018] Preferably, in step 4, during the inference stage, the damage evolution data of the first few frames obtained from the simulation under the working condition to be predicted is first read and placed into the history window; First, the first-level prediction network is used to predict the plastic strain and strain increment of the next frame. Then, the predicted plastic strain and strain increment are used to drive the second-level prediction network to predict the node coordinates and coordinate increments. The prediction results of the two-level prediction networks are jointly filled into the history window to continue predicting subsequent frames.

[0019] Preferably, the method further includes: after obtaining the predicted structural damage evolution data in step 4, mapping the predicted plastic strain and nodal coordinates back to the finite element mesh structure, and using the set chromatographic mapping rules, physical quantity display upper and lower limits and transparency parameters to generate a three-dimensional cloud map that presents the spatial distribution and evolution process of the damage area.

[0020] Beneficial effects: (1) Realize differentiated modeling of multi-stage processes and suppress the accumulation of temporal errors: This invention constructs a conditional vector carrying information on the working conditions and the differentiation of the penetration and explosion stages, and embeds the conditional vector into the latent query vector of multi-head cross-attention. In the feature compression processing stage, the network can distinguish the alternating penetration and explosion stages, autonomously learn the damage evolution law of each stage, avoid the accumulation of prediction errors in the process of multiple alternating simulations, and adapt to multi-stage penetration and explosion simulation scenarios.

[0021] (2) Ensure the complete transmission of working condition and stage constraint information and improve the model's generalization ability: Unlike traditional networks that only place working condition labels in the input layer, this invention directly superimposes the mapped stage and working condition information onto the learnable latent query vector. The constraint information participates in the core operation of the attention layer and will not suffer feature decay due to the transformation of multiple network layers, thus intuitively reflecting the inherent differences between working conditions and stages. Multi-head attention can selectively extract the deformation features of different working conditions and different stages, and the model's generalization performance is significantly improved.

[0022] (3) Differentiated screening of key damage nodes at each stage to enhance the prediction accuracy of high deformation areas: This invention abandons the unified sampling method of all nodes and designs a node sampling method in a targeted manner for the different damage manifestations of the penetration stage and the explosion stage. While avoiding irrelevant data diluting key deformation features and reducing ineffective computing power consumption, it accurately captures the two types of differentiated damage mechanisms, namely, local concentrated deformation during penetration and global cracking during explosion, thereby improving the prediction accuracy.

[0023] (4) Two-level step-by-step decoupled prediction to block cross-interference of physical quantity errors: The present invention adopts a two-level network architecture that first predicts plastic strain and then uses the strain as a priori input to solve for nodal displacement. It clarifies the physical sequence of plastic strain driving geometric deformation and isolates the transmission of strain prediction errors to the geometric field. The two types of physical quantities are optimized independently and do not interfere with each other, and the prediction accuracy of the two sets of output results of plastic strain and nodal coordinates is improved simultaneously. In addition, the strain prediction of each frame drives the geometric field prediction, and the geometric field prediction of each frame participates in the strain prediction of the next frame. It can completely reproduce the mechanical evolution logic of damage deformation in the previous stage of the penetration and explosion process and inherit it from the subsequent steps. It accurately depicts the damage accumulation effect between multiple rounds of alternating penetration and explosion, which is in line with the temporal evolution law of real multi-stage coupled simulation.

[0024] (5) Hybrid training strategy to alleviate the problem of early error amplification in cascaded model training: When inputting the predictions from the first-level prediction network into the second-level prediction network, this invention first mixes the predictions with the true values ​​before providing them to the second-level prediction network. If unstable predictions are used directly to drive geometric field prediction in the early stages of training, the geometric prediction network may learn incorrect mappings caused by strain prediction errors; while if only true values ​​are used throughout the training process, the inconsistency between training and inference distributions will occur when switching to predictions during the inference phase. This invention gradually reduces the weight coefficients, allowing the model to first establish a stable strain-displacement correlation mapping under true strain conditions, and then gradually adapt to the error disturbances caused by predicted strain. This strategy effectively reduces the data distribution offset between training and inference on the one hand, and significantly reduces the error accumulation rate in the process of multi-round cyclic alternating intrusion autoregressive inference on the other hand, significantly improving the long-term stability and global prediction accuracy of time-series damage evolution autoregressive prediction.

[0025] (6) Latent space compression adapts to large-scale irregular grids while taking into account prediction efficiency: This invention completes the latent space compression of massive irregular node fields through dimensional fusion and multi-head cross attention, transforming high-dimensional grid features into low-dimensional temporal latent variables, greatly reducing the computational overhead of long-term multi-stage simulation, and realizing the rapid deduction of damage evolution while ensuring prediction accuracy. Compared with traditional explicit dynamic simulation, the computational efficiency is greatly improved. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the neural operator-based fast prediction method for multi-stage erosion simulation of the present invention.

[0027] Figure 2 This is a schematic diagram illustrating the processing steps of the first-level prediction network and the second-level prediction network. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] This invention provides a fast prediction method for neural operators in multi-stage invasion and explosion simulation, such as... Figure 1 As shown, the method includes the following steps: Step 1: Under different working conditions, obtain structural damage evolution data for each frame of the multi-stage blasting process through explicit dynamic simulation.

[0030] Among them, the operating parameters include at least the initial penetration velocity and penetration location of the multi-stage penetration process.

[0031] This step includes the following sub-steps: Step 101: Obtain structural damage evolution data through simulation.

[0032] This step establishes a finite element model of the multi-stage detonation process, including the geometric model of the target and projectile, mesh generation, material model, contact relationship, initial velocity, boundary constraints, and failure criteria. The information from this finite element model is then written into a k-file (the standard ASCII text input file for the LS-DYNA explicit dynamics solver), and explicit dynamics solution control parameters, termination time, time step control method, and result output control parameters are set in the k-file. The completed k-file is then used to perform calculations using the LS-DYNA solver. After calculation, the corresponding d3plot result file is obtained. This type of file records multi-frame structural damage evolution data from the initial moment of each stage to the termination moment, serving as the data source for subsequent extraction of nodal coordinate fields, plastic strain fields, and construction of neural network training samples.

[0033] Step 102: Data filtering of valid structural units.

[0034] Extract the node coordinates, coordinate increments, plastic strain, strain increments, and part numbers for each node in each frame from the d3plot results file. The part numbers are used to distinguish different structural components. To avoid interference from non-structural regions during the training process, solid structural elements belonging to the concrete structure and their associated nodes are selected based on their part numbers, eliminating elements containing air, explosives, and arbitrary Lagrange operations. Euler (ALE) region.

[0035] Step 103: Determining the sampling nodes for the differentiated invasion and explosion stage.

[0036] To address the issues of small spatial proportion of high-damage areas and easy dilution of key damage information by random sampling in penetration explosion simulation, this invention adopts a stage-adaptive sampling mechanism for the multi-stage penetration explosion process.

[0037] The sampling mechanism for the penetration stage is divided into three regions based on plastic strain values, strain gradients, and the penetration path: a plastic strain concentration region (high damage zone), a strain gradient localization region (damage front region), a penetration channel region, and other regions. Mixed sampling is performed according to a preset node sampling rule to improve the model's ability to learn about penetration damage. Similarly, the sampling mechanism for the explosion stage is divided into three regions based on plastic strain values, strain gradients, and the impact of the explosion: a plastic strain concentration region, a strain gradient localization region, an explosion impact region, and other regions. Mixed sampling is performed according to a preset node sampling rule.

[0038] Specifically, the above node sampling rules involve the sampling ratio and sampling principles: During the penetration stage, the sampling ratio of the plastic strain concentration region is 30% of the total number of sampling points S, the sampling ratio of the strain gradient localization region is 25% of S, the sampling ratio of the penetration channel region is 25% of S, and the sampling ratio of other regions is 20% of S.

[0039] During the explosion phase, the extraction ratio of the plastic strain concentration region is 30% of S, the extraction ratio of the strain gradient localization region is 25% of S, the extraction ratio of the explosion impact influence region is 25% of S, and the extraction ratio of other regions is 20% of S.

[0040] The extraction principles for the plastic strain concentration region are as follows: extraction is performed according to the plastic strain value of the node from largest to smallest; the extraction principle for the strain gradient localization region is as follows: extraction is performed according to the comprehensive plastic strain gradient of the node's neighborhood from largest to smallest; the extraction principle for the penetration channel region is as follows: extraction is performed from nodes whose radial distance from the node to the penetration axis and axial distance from the node along the penetration direction both meet the set range; the extraction principle for the explosion impact influence region is as follows: extraction is performed from nodes whose distance from the node to the explosion center is less than or equal to the explosion radius.

[0041] The above node sampling rules can be expressed by the following formula: The total number of sampling points is S, and the set of all nodes is... .

[0042] The node sampling rule for the penetration phase can be expressed as: S1=0.3 +0.25 +0.25 +0.2( ) The node sampling rule for the explosion phase can be expressed as: S2=0.3 +0.25 +0.25 0.2( ) In the formula, S1 is the set of sampling nodes during the penetration stage, and S2 is the set of sampling nodes during the explosion stage; value Indicates in Sampling nodes that extract a corresponding proportion of values ​​within the region, "Pick , , , At times, these represent the regions of concentrated plastic strain, localized strain gradient, penetration channels, and the impact zone of an explosion, respectively. For example, 0.3 Indicates in 30% of the sampling nodes within the area were selected. ) indicates the penetration phase, except , , Other areas outside, ( ) indicates the explosion stage, except , , Other areas outside of [the designated area].

[0043] The aforementioned region of concentrated plastic strain can be defined as:

[0044] in, This represents the quantile threshold. For example, in implementation, select... The above formula means that the nodes are sorted from largest to smallest plastic strain value, and the top 10% of nodes with high plastic strain are included in the plastic strain concentration region. . This indicates that when the current time is t, the plastic strain at time t-1 is selected to participate in the determination of the sampling nodes in the plastic strain concentration region. Indicates the node sequence number.

[0045] The strain gradient localization region can be defined as:

[0046] in ， Represents the node at time t-1 The equation represents the sum of plastic strain gradients in the neighborhood of a node. The above equation indicates that the nodes with the highest sum of plastic strain gradients in their neighborhoods, sorted from largest to smallest, are selected and included in the strain gradient localization region. .

[0047] The penetration channel area can be defined as:

[0048]

[0049]

[0050] Where, let any node The coordinates are , These are the three-dimensional components of the node coordinates; Let the penetration incident direction be the unit vector. The initial impact point of the projectile (a known quantity). This represents the axial distance of the node along the penetration direction. Let be the radial distance from the node to the penetration axis. Take 1.5 times the instantaneous penetration depth calculated by the Young penetration formula. To determine the threshold for the radial radius of the penetration channel, a value of 10 times the projectile diameter can be used. The above formula indicates that... Belongs to 0~ Between, and Less than or equal to Nodes are extracted from the set number of nodes and included in the intrusion penetration channel area. .

[0051] In the above formula, Young's formula is:

[0052] in, ; ; ; In the formula, Indicates the depth of penetration; Indicates bullet weight; This represents the cross-sectional area of ​​the projectile. Indicates the velocity of the projectile upon impact; Indicates the size scaling effect coefficient; Indicates the performance coefficient of the projectile's head; Indicates the penetration index of the protective layer material; Indicates the correlation coefficient of the protective layer material; This indicates the percentage of the steel content in the protective layer to the total volume of the protective layer. Indicates the setting time of the protective layer pouring; Indicates the thickness of the protective layer; Indicates the diameter of the projectile; Indicates the strength of the protective layer; This represents the ratio of the protective layer width to the projectile diameter; This represents the coefficient of materials without reinforced concrete protective layers.

[0053] The area affected by the explosion's impact can be defined as:

[0054]

[0055] In the formula, The coordinates of the explosion center. The blast radius is... This represents the charge quantity. The above formula indicates that a predetermined number of nodes are selected from those nodes whose distance from the blast center is less than the blast radius and included in the blast impact zone. .

[0056] This completes the selection of sampling nodes.

[0057] The stage-adaptive sampling mechanism in this step extracts sampling nodes from the single-frame structural damage evolution data according to the node sampling rules adapted to different blasting stages. Using the node coordinates, coordinate increments, plastic strains, and strain increments corresponding to the extracted sampling nodes, the damage features of a single frame are constructed in step 2.

[0058] Step 2: Construct training samples based on the structural damage evolution data of the sampling nodes.

[0059] This step extracts node coordinates, coordinate increments, plastic strain, and strain increments from single-frame structural damage evolution data to construct single-frame damage features; then, it stitches together damage features from multiple frames to obtain a feature vector. Based on overall and single-frame structural damage evolution data, a conditional vector containing information distinguishing between working conditions and blast stages is constructed. eigenvectors and condition vector The inputs to the prediction model are used to construct training samples.

[0060] Specifically, this step includes the following sub-steps: Step 201: Construct feature vectors .

[0061] Extract the data of the sampling nodes determined in step 1 from the single-frame structural damage evolution data, including node coordinates, coordinate increments, plastic strain, and strain increments, and construct damage features; these data can be normalized for subsequent splicing. The base value for normalization is the value of the four quantities at a set quantile, such as the 99.5th quantile.

[0062] The feature vector is obtained by concatenating the damage features of all sampling nodes across multiple consecutive frames. This invention sets up a history window, which contains data from the current time and M-1 previous times. Time of the first Feature vectors of each node Represented as:

[0063] in, Indicates splicing, 、 、 、 They represent Time of the first The node coordinates, coordinate increments, plastic strain, and strain increments of each node.

[0064] Step 202: Construct the condition vector .

[0065] Based on overall and single-frame structural damage evolution data, a conditional vector containing information distinguishing between working conditions and blast stages is constructed. In addition to operating condition differentiation information, the condition vector of this invention also includes newly added explosion stage differentiation information.

[0066] The operating condition differentiation information includes operating condition identifier g1, normalized value of operating condition parameter g2, and normalized value of the number of simulation nodes of the current operating condition to which the current frame belongs g3. Here, the normalized value of the number of simulation nodes of the current operating condition to which the current frame belongs g3 is calculated by dividing the number of sampling nodes of the current operating condition by the maximum number of nodes of all operating conditions.

[0067] The information distinguishing the explosion stages includes the explosion stage identifier b1 for the current frame, the normalized value b2 of the frame interval used by the current frame within this explosion stage, and the normalized value b3 of the duration of the explosion stage for the current frame in the entire multi-stage explosion process. The explosion stage identifier b1 for the current frame can be obtained from the d3plot result file. The normalized value b2 of the frame interval used by the current frame within this explosion stage is obtained by dividing the frame interval used by the current frame within this explosion stage by the total duration of this explosion stage. The normalized value b3 of the duration of the explosion stage for the current frame in the entire multi-stage explosion process is obtained by dividing the total duration of the current frame within this stage by the total simulation duration of all explosion stages.

[0068] Step 203: Feature Vector and condition vector The input to the prediction model consists of the plastic strain and its increment in the next frame, and the node coordinates and their increments as labels, forming the training samples.

[0069] Slide the history window to obtain multiple sets of training samples.

[0070] Step 3: Train a two-level prediction network based on latent space neural operators.

[0071] The two-level prediction network based on latent space neural operators of this invention includes a first-level prediction network and a second-level prediction network. The first-level prediction network is a strain prediction network used for strain and its increment prediction, and the second-level prediction network is a geometric prediction network used for coordinate and its increment prediction. The first-level prediction network uses feature vectors... and condition vector The input is the strain increment prediction, which is then used to synthesize the plastic strain prediction. The second-level prediction network uses feature vectors as input. Based on this, eigenvectors for plastic strain prediction and strain increment prediction were added. and condition vector The function takes a node coordinate increment as input and outputs the node coordinate increment prediction, and then synthesizes the node coordinate prediction.

[0072] This invention features a unique design for the prediction network: firstly, it employs a two-stage, decoupled prediction mechanism, first determining which regions will experience plastic accumulation and damage concentration in the next time step, and then inferring the evolution direction and amplitude of the structural geometry based on this plastic strain distribution; secondly, both stages of the prediction network incorporate conditional vectors when introducing learnable latent query vectors. To increase the distinction between working conditions and blasting stages, the constraint information participates in the core operation of the attention layer and will not suffer feature attenuation due to multi-layer network transformation. It intuitively reflects the inherent differences between working conditions and process stages. Multi-head attention can selectively extract deformation features of different working conditions and different processes, and the generalization performance of the model is significantly improved.

[0073] See also Figure 1 The two-stage prediction process in this step is as follows: Step 301: In the first-level prediction network, construct the feature vector of the current frame. and condition vector Input the first-level prediction network; the first-level prediction network outputs the strain increment prediction for the next frame. Here, for the sake of simplicity, the physical quantities are not distinguished between different nodes, and the operation of each node is the same.

[0074] Step 302: Predict the strain increment Plastic strain in the previous frame Add them together to obtain the predicted amount of plastic strain for the next frame. :

[0075] here, The plastic strain obtained using historical predictions already exists within a historical window. During the training phase... Using the plastic strain prediction from the previous frame (See step 303) (Explanation); In the reasoning phase, if there is no true value, then Directly use the prediction from the previous frame As historical prediction values, if a period of true values ​​was collected at the beginning of the inference phase, then in the phase where true values ​​are available, Use the actual values, the same as in the training phase.

[0076] Step 303: Predict the strain increment and plastic strain prediction The value is then fused with the corresponding true value to obtain the fused strain increment prediction. And the predicted amount of plastic strain after fusion During training, the true values ​​are obtained from the structural damage evolution data obtained from the simulation in step 1.

[0077] This invention incorporates a fusion step between predicted and true values ​​between steps 302 and 304, enabling a hybrid training strategy to balance training stability and inference consistency. In the initial training phase, the geometric prediction network primarily uses the true plastic strain of the next frame. With plastic strain increment As a condition for geometric prediction, the system first learns the basic mapping relationship between plastic strain distribution and displacement increment; as training progresses, the prediction capabilities of the first-level network are gradually improved. and The proportion of geometric prediction conditions allows the geometric prediction network to gradually adapt to the actual situation where only predictive strains can be obtained during the inference stage.

[0078] The method for integrating predicted values ​​and actual values ​​is as follows:

[0079]

[0080] in, for The true value of the frame strain increment. for The true value of the frame's plastic strain. To fuse weights; as training progresses, It gradually decays from 1 to 0; when When = 1, the geometric prediction network is entirely driven by real strain; when When = 0, the geometric prediction network is entirely driven by predictive strain, consistent with the actual inference process. During inference, the data is directly applied... Alternatively, set the weight to 0, or, if there are real values ​​in the first few frames, gradually decrease the fusion weight from 1.

[0081] Hybrid training strategies can mitigate the error amplification problem in the early stages of cascaded model training. However, if an unstable training method is used from the very beginning... , Driving geometric field prediction, the geometric prediction network may learn incorrect mappings caused by strain prediction errors; however, if training always uses real data... 、 Then the reasoning phase switches to 、 This can lead to inconsistencies between training and inference distributions. By gradually reducing... The model can first establish a stable strain-displacement relationship under real strain conditions, and then gradually adapt to error disturbances under predicted strain conditions, thereby improving the long-term stability of autoregressive prediction.

[0082] Step 304: The feature vector corresponding to the current frame Add the fused strain increment prediction And the predicted amount of plastic strain after fusion To obtain new feature vectors ; to the new feature vector With condition vector The input is a two-level prediction network, which outputs the coordinate increment prediction for the next frame. .

[0083] The feature vector in this step The prediction results were incorporated into the first-level prediction network; middle Time of the first The components of each node are expressed as follows:

[0084] Step 305: Predict the coordinate increment Compared to the node coordinates of the previous frame Add them together to get the predicted node coordinates for the next frame. : = .

[0085] in, Node coordinates obtained from historical predictions, i.e., node coordinates predicted in the previous frame. It already exists in the historical window.

[0086] See Figure 2 The figure shows the architecture of the first-level prediction network and the second-level prediction network. They have the same composition, both of which include a dimensionality-upgrading module, a latent space compression module, a latent space temporal evolution and decoding module, and an update module.

[0087] The feature vectors and conditional vectors input to this level of the prediction network are respectively subjected to dimensionality-upgrading mapping by the dimensionality-upgrading module and then added together to achieve feature fusion, thereby obtaining fused features in the latent space dimension. The dimensionality-upgrading module can use an MLP to map to a D-dimensional latent space. As mentioned earlier, the feature vector input to the first-level prediction network is... The feature vector input to the second-level prediction network is .

[0088] The latent space compression module introduces learnable latent query vectors. Add the condition vector after the dimensionality upgrade mapping to the latent query vector. This yields a condition constraint query vector with constraints on operating conditions and blasting stages. .

[0089] Then, multi-head cross-attention is performed to fuse features. Simultaneously serving as a key vector for multi-head cross-attention. Sum value vector Together with conditional constraints query vector Multi-head cross-attention computation is performed together. After multi-head concatenation, layer normalization, and residual connection, the loss features of the irregular node field in the current frame are compressed into a fixed number of latent space labels. Combined with historical window data, the latent space labels including the current frame and M-1 historical frames are combined to obtain the historical latent space label sequence. The data is then fed into the latent space temporal evolution and decoding module. M is the length of the history window.

[0090] Multi-head cross-attention processing is represented as: ； ;

[0091] in, For layer normalization operation, This indicates a cross-attention operation.

[0092] Specifically, note that the number of heads is h , For each attention head : , ,

[0093]

[0094]

[0095] in, , , The first The header contains conditional constraints on the query vector, key vector, and value vector. for t The compressed latent space marker of the current frame at time 1. All are learnable matrices. Represents the normalized exponential function, Indicates the first The self-attention calculation results for each attention head.

[0096] This operation is performed independently for each frame, combining the latent space tags of the current frame and M-1 historical frames to ultimately obtain a sequence of historical latent space tags with a length equal to the length M of the historical window. It is then fed into the latent space temporal evolution and decoding module.

[0097] The latent space temporal evolution and decoding module is used to learn the inter-frame evolution patterns of latent space marker sequences and to process the input historical latent space marker sequences. Mapped to the evolved latent space label sequence The damage representation increment is then decoded and output as predicted by the current-level prediction network. For the first-level prediction network, the absolute value of the damage representation is plastic strain, and the damage representation increment is the strain increment; for the second-level prediction network, the absolute value of the damage representation is node coordinates, and the damage representation increment is the coordinate increment.

[0098] The latent space temporal evolution and decoding module can employ either a Transformer network or a Mamba network. Taking Transformer as an example, it will... After inputting the multi-head self-attention layer, the latent space sequence after multi-head attention fusion is obtained. Then, the original input latent label sequence With multi-head attention output Perform residual connections and layer normalization to obtain... :

[0099] The input is fed into a feedforward network, followed by residual connections and layer normalization. After passing through multiple stacked Transformer encoders, the resulting latent space label sequence is output. By decoding from The system reads the temporal information associated with each node, and then outputs the damage characterization increments of each node predicted by the current-level prediction network through the strain decoder.

[0100] The update module is used to add the damage representation increment predicted by the current prediction network to the absolute damage representation of the previous frame to obtain the absolute damage representation of each node predicted by the current prediction network.

[0101] The two-stage prediction network is jointly trained and optimized. The total training loss consists of strain prediction loss and geometric prediction loss. Specifically, the geometric prediction loss includes the strain-geometric coupling consistency loss. This method is used to improve the accuracy of geometric field prediction in regions of concentrated plastic strain. By assigning higher weights to the coordinate increment errors in high-strain regions, the geometric deformation at penetration channels, explosion wavefronts, and damage propagation boundaries is learned in a focused manner.

[0102] Among them, the strain prediction loss is obtained by weighted fusion strain Smooth L1 loss. Monotonous penalty loss and strain adaptive weighted loss Obtain. Strain-adaptive weighted loss This allows nodes located in regions of concentrated plastic strain to receive greater weights in the calculations, thereby improving prediction accuracy. Specifically, this is achieved as follows:

[0103]

[0104] ,

[0105]

[0106] In the formula: To predict losses for strain; , , They are respectively , , The Middle The loss of each node is calculated. For a single loss term, the average loss of all sampled nodes is taken as the global value of that loss term. The global values ​​of the three losses are then weighted to obtain the strain prediction loss. . for Time of the first Predicted plastic strain at each node. for Time of the first The true value of plastic strain at each node. for Time of the first The predicted amount of plastic strain at each node. The parameters are for the Smooth L1 loss.

[0107] Monotonic penalty loss A penalty is imposed when the plastic strain of adjacent frame nodes decreases, when a certain node When the plastic strain decreases, the decrease is taken as The loss value, as plastic strain increases, The loss value is set to 0. Adaptive weighted loss for strain. The absolute value of the prediction error of nodal plastic strain As a basic constraint, a first dynamic penalty weight is introduced that grows linearly with the true value of the nodal plastic strain. Increase the weight of the first dynamic penalty for nodes with higher plastic strain. .

[0108] Geometric prediction loss is achieved through weighted fused coordinate Smooth L1 loss. Acceleration smoothing loss and strain-geometric coupling consistency loss Obtain. The specific implementation is as follows:

[0109]

[0110] ,

[0111]

[0112] In the formula: Geometric prediction loss; , , They are respectively , , The Middle The loss for each node is calculated. For a single loss term, the average loss across all sampled nodes is taken as the global value for that term. The geometric prediction loss is obtained by weighting the global values ​​of the three losses. . for Time of the first The predicted coordinates of each node. for Time of the first The actual values ​​of the node coordinates of each node. for Time of the first The predicted coordinate increment of each node. For the output of the second-level prediction network Time of the first The predicted coordinate increment of each node. for Time of the first The actual value of the coordinate increment of each node. The parameters are the coordinate Smooth L1 loss parameters.

[0113] Acceleration smoothing loss It is a smoothing loss term to prevent the predicted displacement increment from jittering drastically between adjacent frames; it is represented by the second-order difference magnitude of the predicted node coordinate increment over three consecutive frames. The larger the second-order difference magnitude, the greater the acceleration smoothing loss.

[0114] Strain-geometric coupling consistency loss It is a sensing term for the concentrated plastic strain region, with the absolute error between the predicted coordinate increment output by the two-level prediction network and the corresponding true coordinate increment as the basic constraint term, and introduces a second dynamic penalty weight that increases linearly with the plastic strain of the nodes. Increase the weight of the second dynamic penalty for nodes with higher plastic strain. .

[0115] Finally, the strain prediction loss and geometric prediction loss are weighted and calculated to obtain the total loss. :

[0116] The advantage of this cascaded network scheme lies in the fact that the strain prediction network focuses on learning the accumulation of plasticity and the evolution of damage distribution in the material, while the geometry prediction network uses the predicted strain field to further learn the damage-driven structural displacement response. The two form a cascaded closed loop in reasoning, making the plastic strain field an explicit intermediate condition for the evolution of the geometric field. Compared to directly outputting strain and coordinates simultaneously, this scheme better aligns with the physical understanding that "local plastic damage development drives geometric deformation" during penetrating explosions, and it also makes it easier to set strain monotonicity constraints, displacement smoothness constraints, and high-damage region weighted constraints separately in the loss function.

[0117] Step 4: Inference stage, use the trained prediction model to predict structural damage evolution data.

[0118] During the inference phase, the overall prediction process is as follows: first, the first-level prediction network is used to predict the plastic strain and strain increment of the next frame; then, the predicted plastic strain and its increment are used to drive the second-level prediction network to predict the node coordinates and coordinate increments; the prediction results of the two-level prediction networks are jointly backfilled into the history window to continue rolling the prediction of subsequent frames.

[0119] The specific reasoning and prediction steps include: Before inference begins, the damage evolution data from the first few frames obtained from the simulation under the predicted working condition are read and placed in the history window. These known frames come from the completed LS-DYNA calculation results. Simultaneously, the working condition information for constructing the condition vector is read, including the working condition identifier, normalized values ​​of working condition parameters, and normalized values ​​of the number of simulation nodes for the current working condition; all of these can be obtained from the known frames. Stage information uses preset values: the identifier of the current explosion stage can be determined through the d3plot directory structure, which is pre-built but contains only a few frames of real data; the normalized value of the frame interval used by the current frame in this explosion stage is a preset value, calculated based on the pre-determined time interval and total duration of each stage; the normalized value of the duration of the current explosion stage in the entire multi-stage explosion process is also a preset value, determined based on the pre-determined total duration of each stage and the total duration of the entire simulation.

[0120] For the frame data at time t to be predicted, node-level input features are first constructed using node coordinates, plastic strain, coordinate increments, and strain increments from the historical window. The input features from each node in M ​​frames are then combined to obtain the feature vector. This is used to describe the positional state, plastic damage state, and local evolution trend of each node at a historical moment. The feature vectors are then used... and condition vector Its input is a first-level prediction network, namely the strain prediction network. The strain prediction network first processes the feature vector. and condition vector Encoding is performed, and a large-scale irregular node field is compressed into a fixed number of latent space labels through conditional cross-attention pooling. Then, a temporal processing network models the temporal evolution of the historical latent space label sequence. Subsequently, the strain prediction network outputs the predicted strain value for the next time step through a node-level decoder. This prediction uses a non-negative incremental form; that is, the strain increment is output first, and then added to the plastic strain of the previous time step to obtain the predicted strain for the current frame. To ensure the cumulative characteristics of plastic strain, a monotonicity constraint is applied to the prediction result, ensuring that the predicted strain of the current frame is not lower than the predicted strain of the previous frame.

[0121] After obtaining the predicted strain for the current frame, the predicted plastic strain and strain increment are further added to the feature vector as important conditions for the second-level prediction network, namely the geometric prediction network. In the process, the feature vector is obtained. The second-level prediction network is based on feature vectors. and condition vector The system predicts how the structural geometry should evolve under the current plastic damage distribution, obtaining the predicted nodal coordinates and coordinate increments. After obtaining the predicted strain, strain increment, predicted coordinates, and coordinate increments for the current frame, these are added to the history window as the latest frame data to participate in the prediction of the next frame. In this way, the cascaded network forms a complete autoregressive closed loop: the strain prediction of each frame drives the geometry prediction, and the geometry prediction results of each frame participate in the strain prediction of the next frame.

[0122] Step 5: Visualize the spatial distribution and evolution of the damaged area.

[0123] The predicted plastic strain and nodal coordinates are mapped back to the finite element mesh structure. Using the set chromatographic mapping rules, upper and lower limits of physical quantity display, and transparency parameters, a three-dimensional cloud map is generated to intuitively present the spatial distribution and evolution process of the damage area.

[0124] Specifically, after the prediction is completed, the predicted plastic strain and nodal coordinates can be mapped back to the original finite element mesh structure. The visualization process employs jet mapping, consistent with or approximating LS-DYNA, and uniformly sets the upper and lower limits for displaying physical quantities and transparency parameters to generate a 3D cloud map that visually presents the spatial distribution and evolution of the damage zone. Based on the Unity 3D system, the prediction results can be used as node-by-node and element-by-element visualization attributes to play out the continuous penetration evolution process.

[0125] The penetration effect can also be quantified based on visualized data. For example, the following methods can be used: identify pit nodes based on plastic strain and nodal displacement thresholds; calculate pit depth based on the maximum depth of pit nodes in the penetration direction; calculate pit width based on the maximum span of the pit area in the transverse projection; estimate pit volume based on the volume integral of finite element mesh elements; and estimate damage scale based on the number of nodes or elements exceeding the threshold.

[0126] The specific embodiments described above only illustrate the design principles of the present invention. The shapes and names of the components in this description may differ and are not limited. Therefore, those skilled in the art can modify or make equivalent substitutions to the technical solutions described in the foregoing embodiments; and these modifications and substitutions do not depart from the inventive spirit and technical solutions of the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A fast prediction method using neural operators for multi-stage blast simulation, characterized in that, Structural damage evolution analysis for multi-stage blasting processes; the method includes: Step 1: Under different working conditions, obtain structural damage evolution data for each frame of the multi-stage blasting process through explicit dynamic simulation; Step 2: Extract the node coordinates, coordinate increments, plastic strain, and strain increments of each node from the single-frame structural damage evolution data to construct damage features; stitch together the damage features from multiple frames to obtain the feature vector. Based on overall and single-frame structural damage evolution data, a conditional vector containing information distinguishing between working conditions and blast stages is constructed. ; using feature vectors and condition vector Use these samples as input to the prediction model to build a training sample set. Step 3: Train a two-level prediction network based on latent space neural operators; the first-level prediction network uses feature vectors and condition vector As input, it outputs the predicted strain increment and synthesizes the predicted plastic strain; in the eigenvector Add strain increment prediction and plastic strain prediction to form a new eigenvector. The second-level prediction network uses feature vectors and condition vector As input, the network outputs coordinate increment predictions and synthesizes node coordinate predictions; moreover, both levels of the prediction network add condition vectors to the learnable latent query vectors. To increase the distinguishing information between working conditions and blasting stages, latent space marking with conditional constraints is obtained through multi-head cross-attention processing; Step 4: Inference stage, use the trained prediction model to predict structural damage evolution data.

2. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in claim 1, characterized in that, In step 2, the extraction of node coordinates, coordinate increments, plastic strain, and strain increments of each node from single-frame structural damage evolution data to construct damage features specifically involves: According to the node sampling rules adapted to different blasting stages, sampling nodes are extracted from single-frame structural damage evolution data; the node coordinates, coordinate increments, plastic strains and strain increments of the sampling nodes are used to construct the damage features of a single frame. The node sampling rules include the sampling ratio and sampling principles, specifically: During the penetration phase, sampling nodes are extracted from the plastic strain concentration region, strain gradient localization region, penetration channel region, and other regions according to a set extraction ratio. Specifically, the extraction ratio of nodes in the plastic strain concentration region is 30% of the total number of sampling points S, the extraction ratio in the strain gradient localization region is 25% of the total number of sampling points S, the extraction ratio in the penetration channel region is 25% of the total number of sampling points S, and the extraction ratio in other regions is 20% of the total number of sampling points S. During the explosion phase, sampling nodes are extracted from the plastic strain concentration region, the strain gradient localization region, the explosion impact influence region, and other regions according to a set extraction ratio. Specifically, the extraction ratio of the plastic strain concentration region is 30% of the total number of sampling points S, the extraction ratio of the strain gradient localization region is 25% of the total number of sampling points S, the extraction ratio of the explosion impact influence region is 25% of the total number of sampling points S, and the extraction ratio of other regions is 20% of the total number of sampling points S. The principle for extracting the plastic strain concentration region is: to extract it according to the nodal plastic strain value from large to small; The principle for extracting the strain gradient localization region is: to extract it from large to small according to the comprehensive plastic strain gradient of the node neighborhood; The principle for extracting the penetration channel region is: nodes whose radial distance from the node to the penetration axis and axial distance along the penetration direction both meet the set range. The principle for selecting the impact zone of the explosion is: to select from nodes whose distance from the explosion center is less than or equal to the explosion radius.

3. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in claim 1, characterized in that, The condition vector middle: The operating condition differentiation information includes the operating condition identifier, the normalized value of the operating condition parameter, and the normalized value of the number of simulation nodes of the current frame's operating condition; The information distinguishing the intrusion phase includes the intrusion phase identifier of the current frame, the normalized value of the frame interval used by the current frame in this intrusion phase, and the normalized value of the duration of the intrusion phase of the current frame in the entire multi-stage intrusion process.

4. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in claim 1, characterized in that, In the two-level prediction network based on latent space neural operators, both the first-level prediction network and the second-level prediction network include a dimensionality-upgrading module, a latent space compression module, a latent space temporal evolution and decoding module, and an update module. The feature vectors and conditional vectors input to this level of the prediction network are both up-mapped by the up-dimensionality module and then added together to obtain the fused features of the latent space dimension. ; The latent space compression module introduces learnable latent query vectors. Add the condition vector after the dimensionality upgrade mapping to the latent query vector. This yields a condition constraint query vector with constraints on operating conditions and blasting stages. ; with fusion features Simultaneously serving as the key and value vectors for multi-head cross-attention, along with the conditional constraint query vector... Multi-head cross-attention computation is performed together. After multi-head concatenation, layer normalization, and residual connection, the loss features of the irregular node field in the current frame are compressed into a fixed number of latent space labels. Combined with historical window data, the latent space labels including the current frame and M-1 historical frames are combined to obtain the historical latent space label sequence. It is then fed into the latent space temporal evolution and decoding module; M is the length of the history window; The latent space temporal evolution and decoding module is used to learn the inter-frame evolution patterns of latent space marker sequences and to process the input historical latent space marker sequences. Mapped to the evolved latent space label sequence After decoding, the damage representation increment predicted by the current prediction network is output. The update module is used to add the damage representation increment predicted by the current prediction network to the absolute damage representation of the previous frame to obtain the absolute damage representation predicted by the current prediction network. For a first-level prediction network, the absolute value of the damage representation is plastic strain, and the damage representation increment is the strain increment; for a second-level prediction network, the absolute value of the damage representation is node coordinates, and the damage representation increment is the coordinate increment.

5. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in claim 4, characterized in that, The dimensionality enhancement module employs a multilayer perceptron (MLP); the latent space temporal evolution and decoding module employs a Transformer network or a Mamba network.

6. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in claim 1, characterized in that, Step 3 specifically includes: Step 301: Construct the feature vector of the current frame and condition vector Input the first-level prediction network and output the predicted strain increment for the next frame. ; Step 302: Calculate the predicted strain increment. Plastic strain in the previous frame Add them together to obtain the predicted amount of plastic strain for the next frame. Among them, the plastic strain in the previous frame Plastic strain obtained using historical predictions; Step 303: Calculate the predicted strain increment. and the predicted amount of plastic strain The value is then fused with the corresponding true value to obtain the fused strain increment prediction. And the predicted amount of plastic strain after fusion The weights for fusion of true values ​​gradually decrease from 1 to 0 as training progresses. Step 304: The feature vector corresponding to the current frame Add the fused strain increment prediction And the predicted amount of plastic strain after fusion To obtain new feature vectors ; to feature vector With condition vector The input is a two-level prediction network, which outputs the coordinate increment prediction for the next frame. ; Step 305: Calculate the predicted coordinate increment. Compared to the node coordinates of the previous frame Add them together to get the predicted node coordinates for the next frame. Among them, the node coordinates of the previous frame Node coordinates obtained using historical predictions.

7. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in claim 1, characterized in that, The first-level prediction network and the second-level prediction network are jointly trained and optimized, and the total loss consists of strain prediction loss and geometric prediction loss. The strain prediction loss is obtained by weighted fusion of strain Smooth L1 loss and monotonic penalty loss. and strain adaptive weighted loss get: The monotonic penalty loss Penalize instances where the plastic strain of adjacent frame nodes decreases; The strain adaptive weighted loss The absolute value of the prediction error of the nodal plastic strain is used as the basic constraint term. A first dynamic penalty weight is introduced that increases linearly with the true value of the nodal plastic strain, and the first dynamic penalty weight is increased for nodes with higher plastic strain. The geometric prediction loss is calculated using weighted fusion coordinate Smooth L1 loss and acceleration smoothing loss. and strain-geometric coupling consistency loss get: The acceleration smoothing loss The second-order difference magnitude of the predicted node coordinate increment over three consecutive frames is used to represent the value. The strain-geometric coupling consistency loss The absolute error between the predicted coordinate increment output by the secondary prediction network and the actual value of the corresponding coordinate increment is used as the basic constraint term. A second dynamic penalty weight is introduced that increases linearly with the plastic strain of the node, and the second dynamic penalty weight is increased for nodes with higher plastic strain.

8. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in claim 1, characterized in that, In step 4, during the inference stage, the damage evolution data of the first few frames obtained from the simulation under the working condition to be predicted is first read and placed into the history window. First, the first-level prediction network is used to predict the plastic strain and strain increment of the next frame. Then, the predicted plastic strain and strain increment are used to drive the second-level prediction network to predict the node coordinates and coordinate increments. The prediction results of the two-level prediction networks are jointly filled into the history window to continue predicting subsequent frames.

9. The fast prediction method for neural operators in multi-stage invasion and explosion simulation as described in any one of claims 1-8, characterized in that, The method further includes: after obtaining the predicted structural damage evolution data in step 4, mapping the predicted plastic strain and nodal coordinates back to the finite element mesh structure, and using the set chromatographic mapping rules, physical quantity display upper and lower limits and transparency parameters to generate a three-dimensional cloud map that presents the spatial distribution and evolution process of the damage area.

Images