A new energy vehicle aluminum alloy rear longitudinal beam design method based on topology optimization

Abstract

The application discloses a new energy automobile aluminum alloy rear longitudinal beam design method based on topology optimization, and relates to the technical field of new energy automobile structure design, and comprises the following steps: step S1: inputting rear longitudinal beam initial design parameters and multi-physical field boundary conditions; and step S2: constructing an aluminum alloy material constitutive model.The new energy automobile aluminum alloy rear longitudinal beam design method based on topology optimization solves the problem of low multi-field data interaction efficiency in a traditional tool chain. The characteristic decoupling layer automatically strips redundant characteristics of heterogeneous physical fields, replaces data channel intelligent avoidance of high-delay nodes, and realizes real-time synchronous iteration of structure-collision-thermal coupling simulation. In combination with a self-adaptive reduced-order proxy model and physical constraint embedded verification, the optimization precision is guaranteed, and the single iteration period is shortened to a fraction of the time consumed by the traditional method, thereby effectively accelerating the lightweight development process of key components of new energy automobiles.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle structural design technology, specifically a design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization. Background Technology

[0002] In the field of topology optimization design for aluminum alloy rear longitudinal beams in new energy vehicles, existing technologies heavily rely on electro-digital data processing (EDS) for simulation and optimization. To meet the dual goals of lightweighting and collision safety, the design process needs to integrate multi-physics coupled simulation, including structural mechanics analysis, dynamic collision simulation, and thermal management impact assessment. However, current mainstream methods employ decentralized simulation toolchains, with computational modules for different physics fields running independently, resulting in low data interaction efficiency. For example, structural optimization iterations require repeated transmission of stress, deformation, and thermal field data to the collision simulation module, but interface redundancy and data format differences between commercial software lead to frequent data conversions and transmission delays. This inefficient data collaboration mechanism significantly extends the computation time of a single optimization iteration, with a large amount of resources consumed in redundant data reading and writing rather than solving the core algorithm, severely restricting design efficiency. Especially when dealing with the dynamic nonlinear behavior of aluminum alloy rear longitudinal beams, due to data synchronization lag, optimization results often deviate from actual collision conditions, requiring additional physical experiments for verification, further increasing the R&D cycle and cost. While existing patents have optimized algorithm convergence or material models, they have not fundamentally solved the bottleneck of real-time data flow under multi-field coupling, making it difficult to meet the rapid iteration requirements of key components in new energy vehicles. The technical problem to be solved is: how to improve the real-time performance of data interaction in multi-physics coupled simulations to efficiently coordinate the topology optimization process for lightweighting and collision safety. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: a design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization, comprising the following steps:

[0004] Step S1: Input the initial design parameters and multiphysics boundary conditions for the rear longitudinal beam;

[0005] Step S2: Construct a constitutive model of the aluminum alloy material;

[0006] Step S3: Perform multi-field coupled topology optimization through a unified computation graph architecture;

[0007] Step S4: Embed a physical constraint verification layer at the output of the computation graph;

[0008] Step S5: Verify the optimization results using a dual-channel adversarial network;

[0009] Step S6: Output the topology of the rear longitudinal beam.

[0010] Preferably, the multiphysics boundary conditions in step S1 include structural loads, collision conditions, and thermal management constraints;

[0011] The constitutive model described in step S2 associates the lightweight target with the collision safety threshold.

[0012] Preferably, the unified computation graph architecture described in step S3 is implemented as follows:

[0013] The structural stress field, collision dynamics field, and thermodynamic field data are encoded into a domain adaptive tensor, which includes spatial coordinates, time steps, and physical quantity types.

[0014] Heterogeneous data formats are converted through a physical field feature decoupling layer.

[0015] Preferably, step S3 further includes dynamically compiling and generating a computational graph:

[0016] Integrating a reduced-order collision proxy model with thermal influence factor nodes;

[0017] Nodes are scheduled in parallel based on data dependencies;

[0018] For nonlinear collision calculation nodes, alternative data channels are generated under the condition of satisfying real-time constraints.

[0019] Preferably, the reduced-order collision proxy model is constructed through a pre-trained neural network, with the structural deformation gradient field as input and the collision energy absorption value as output.

[0020] The alternative data channel generates an approximate solution based on a simplified energy conservation equation.

[0021] Preferably, the physical constraint verification layer in step S4 operates as follows:

[0022] The correlation threshold between the strain rate effect and the thermal softening effect of the built-in aluminum alloy;

[0023] When the optimization result deviates from the constitutive equation, local data resampling is triggered.

[0024] Preferably, the basis for setting the association threshold includes:

[0025] The attenuation characteristics of aluminum alloy yield strength under dynamic load;

[0026] Critical temperature conduction gradient in battery thermal runaway scenarios.

[0027] Preferably, the dual-channel adversarial network described in step S5 operates as follows:

[0028] The generator synthesizes extreme collision condition data based on the current optimization results;

[0029] The discriminator compares the statistical distribution differences between simulation data and the historical database of physical experiments.

[0030] Preferably, it also includes confidence level determination and backtracking mechanisms:

[0031] When the confidence level of the discriminator output is lower than the dynamic calibration threshold, the process backtracks to the key coupling node marked in the computation graph.

[0032] The physical test history database indexes new energy vehicle crash test data through a transfer learning framework.

[0033] Preferably, step S6: perform post-processing before output, including manufacturing process constraint mapping and topology smoothness filtering.

[0034] This invention provides a design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization. It has the following beneficial effects:

[0035] This topology-optimized design method for aluminum alloy rear longitudinal beams in new energy vehicles addresses the inefficiency of multi-field data interaction in traditional toolchains by employing dynamic compilation technology and a domain-adaptive tensor transformation mechanism within a unified computational graph architecture. A feature decoupling layer automatically removes redundant features from heterogeneous physical fields, intelligently avoiding high-latency nodes by replacing data channels, and enabling real-time synchronous iteration of structure-collision-thermal coupling simulation. Combined with a condition-adaptive reduced-order proxy model and embedded verification of physical constraints, the method shortens the single iteration cycle to the fragment-level time of traditional methods while ensuring optimization accuracy, effectively accelerating the lightweight development process of key components for new energy vehicles.

[0036] This topology optimization-based design method for aluminum alloy rear longitudinal beams in new energy vehicles achieves engineering-level reliability assurance of the topology optimization results through a dual-channel adversarial network and a confidence level backtracking mechanism. The generator synthesizes extreme working conditions to enhance verification coverage in weak areas, while the discriminator determines the confidence level through distribution moment differences and KL divergence classification. A dual-threshold combination triggering mechanism in the physical verification layer avoids invalid interruptions, and node-level precise backtracking inherits valid data and strengthens input. This system reduces the need for physical experimental verification to a low level, ensuring a fundamental synergistic breakthrough in both collision safety and lightweighting goals for the aluminum alloy rear longitudinal beam, while simultaneously generating constraint mapping rules to directly output a mass-producible topology structure. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the system architecture of a design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to the present invention.

[0038] Figure 2 This is a flowchart illustrating the design method of aluminum alloy rear longitudinal beam for new energy vehicles based on topology optimization according to the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 and Figure 2 This invention provides a technical solution: a design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization, comprising the following steps:

[0041] Step S1: Input the initial design parameters and multiphysics boundary conditions for the rear longitudinal beam;

[0042] Step S2: Construct a constitutive model of the aluminum alloy material;

[0043] Step S3: Perform multi-field coupled topology optimization through a unified computation graph architecture;

[0044] Step S4: Embed a physical constraint verification layer at the output of the computation graph;

[0045] Step S5: Verify the optimization results using a dual-channel adversarial network;

[0046] Step S6: Output the topology of the rear longitudinal beam.

[0047] It should be further explained that, in the specific implementation process, the initial design parameters of the rear longitudinal beam and multiphysics boundary conditions are input, including structural load distribution, collision acceleration pulse waveform, and battery pack thermal management constraints; when constructing the constitutive model of the aluminum alloy material, the upper limit of the mass of the lightweight target and the minimum energy absorption requirement of the collision safety threshold are correlated. The implementation steps are as follows:

[0048] S01: Unified Computational Graph Architecture: Heterogeneous data from structural stress fields, collision dynamics fields, and thermodynamic fields are transformed into isomorphic domain adaptive tensors through a physics feature decoupling layer. These tensors are organized according to spatial coordinate grids, time step sequences, and physical quantity types, such as stress, temperature, and deformation, in three dimensions. When generating the optimized computational graph based on dynamic compilation technology, if the computational delay of a collision dynamics node exceeds the real-time constraint threshold (e.g., single-step iteration timeout), an alternative data channel is automatically generated, including the following:

[0049] Normal operating conditions: Invoke the reduced-order collision proxy model, which maps the structural deformation gradient to the collision energy absorption value by a pre-trained neural network;

[0050] Extreme operating conditions: An approximate solution is generated based on a simplified energy conservation equation to ensure iterative continuity;

[0051] When scheduling thermal influence factor nodes and structural field solution nodes in parallel, the execution order is dynamically adjusted based on data dependencies, such as freezing structural field updates before heat conduction is completed.

[0052] S02: Dual-modal verification mechanism: The physical constraint verification layer incorporates aluminum alloy strain rate-thermal softening correlation thresholds, such as the critical point of yield strength decay under dynamic load and the safety line of battery thermal runaway conduction gradient. When the stress-temperature coupling data in the topology optimization results deviates from the constitutive equation, local data resampling is immediately triggered, targeting only abnormal mesh regions rather than a global restart. The dual-channel adversarial network verification operates according to the following logic, including:

[0053] The generator synthesizes two types of extreme condition data, including the following two dominant types:

[0054] Collision-dominated type: Randomly superimposed road surface excitation and bias collision vectors;

[0055] Thermally dominant type: Coupled variables of battery thermal shock and structural thermal expansion;

[0056] The discriminator compares the statistical distribution of simulation data with the historical physical experiment database: if the KL divergence is lower than the dynamic calibration threshold, it is determined to be verified as passed based on the confidence interval set in the historical database; if the KL divergence exceeds the standard, it backtracks to the key coupling nodes marked in the computation graph, such as the structure-collision interaction layer, and retains the verified valid data branch while only restarting the problem node.

[0057] S03: Before the final output of the longitudinal beam topology, perform manufacturing process constraint mapping and smoothness filtering.

[0058] The multiphysics boundary conditions in step S1 include structural loads, collision conditions, and thermal management constraints.

[0059] In step S2, the constitutive model correlates the lightweight target with the collision safety threshold.

[0060] It should be further explained that, in the specific implementation process, the input of multiphysics boundary conditions includes the following steps:

[0061] Structural loads are applied according to the actual installation points, distinguishing between static loads and dynamic fatigue loads; static loads include the weight of the battery pack, while dynamic fatigue loads include the frequency domain spectrum of road surface excitation; three collision scenarios are preset, and the preset process is as follows:

[0062] Frontal full-width collision: Define uniform deceleration acceleration pulse waveform;

[0063] Side column impact: Specifies the time history curve of the local impact force;

[0064] Rear-end collision: Set up a multi-energy-level trapezoidal wave;

[0065] Thermal management constraints define the critical temperature range for thermal runaway of the battery pack and correlate it with the maximum allowable temperature rise gradient of the rear longitudinal beam heat conduction path.

[0066] When constructing the constitutive model of aluminum alloy materials, the lightweighting objective is transformed into an upper limit threshold of mass, which is dynamically allocated according to the volume density distribution of the rear longitudinal beam topology domain. The collision safety threshold is set as the minimum energy absorption requirement, which is defined differently according to different working conditions. The process is as follows: frontal collisions must meet the lower limit of total energy absorption value; side collisions must meet the lower limit of energy absorption efficiency per unit deformation. When the lightweighting objective conflicts with the collision safety threshold, such as insufficient energy absorption due to weight reduction, the coordinates of the conflict area are marked in the model for priority optimization of the computational graph.

[0067] The implementation method of the unified computation graph architecture in step S3 is as follows:

[0068] The structural stress field, collision dynamics field, and thermodynamic field data are encoded into a domain adaptive tensor, which includes spatial coordinates, time steps, and physical quantity types.

[0069] Heterogeneous data formats are converted through a physical field feature decoupling layer.

[0070] It should be further explained that, in the specific implementation process, the conversion of the original heterogeneous data of the structural stress field, collision dynamics field, and thermodynamic field into a domain adaptive tensor follows the following procedure:

[0071] The spatial coordinate dimension is discretized according to the rear longitudinal beam topology mesh, and each mesh cell is bound to a material property identifier; the time step dimension is dynamically segmented according to the working condition type, and the segmentation process is as follows:

[0072] Collision-dominated segment: High sampling rate sequences are used to capture transient responses;

[0073] Thermally dominant section: Steady-state conduction is recorded using an equally spaced sequence;

[0074] The physical quantity type dimension integrates three types of data. The integration process is as follows: the structural field outputs nodal displacement, equivalent stress and plastic strain; the collision field outputs kinetic energy absorption and peak contact force; and the thermal field outputs temperature field distribution and heat flux density.

[0075] The operational logic of the physical field feature decoupling layer includes feature separation and isomorphic recombination; among which, feature separation involves applying a spatial convolution kernel to the thermal field data to remove pure heat conduction features; and applying a temporal filtering kernel to the collision field data to separate inertial effect features.

[0076] Homogenization and recombination: If a structural-thermal coupling region is detected, such as the connection of a battery bracket, the stripped features are mapped to a shared tensor channel according to the grid coordinates; if the collision-structural data are asynchronous in time, such as stress wave propagation delay, they are aligned to a unified time step sequence and then written into the tensor.

[0077] Step S3 also includes dynamically compiling and generating the computation graph:

[0078] Integrating a reduced-order collision proxy model with thermal influence factor nodes;

[0079] Nodes are scheduled in parallel based on data dependencies;

[0080] For nonlinear collision calculation nodes, alternative data channels are generated under the condition of satisfying real-time constraints.

[0081] It should be further explained that, in the specific implementation process, when dynamically compiling and generating the computation graph, the input and output flow between nodes is analyzed based on data dependencies. The analysis process is as follows:

[0082] If the output of the thermal influence factor node is the input of the structural field solution node, then the thermal node is forced to be scheduled first; if the collision proxy node and the structural field node have no direct data dependency, they are executed in parallel to shorten the iteration cycle; when the estimated duration of the nonlinear collision calculation node is detected to exceed the real-time constraint threshold, an alternative data channel is automatically generated, and the generation process is as follows:

[0083] For conventional energy absorption scenarios, such as low-speed collisions, the reduced-order collision proxy model is invoked to directly output the results. For highly nonlinear scenarios, such as material tearing failure, an approximate solution channel is constructed based on a simplified energy conservation equation. The construction process is as follows: input the structural deformation gradient field, and output the equivalent energy absorption value through the momentum-energy equivalent conversion formula to avoid complex contact calculations. During the activation of the alternative channel, the calculation status of the original node is continuously monitored. If it is completed ahead of schedule, the high-precision data stream is switched back.

[0084] When integrating a reduced-order collision proxy model, neural network weights are preloaded during the compilation phase, and model branches are dynamically selected based on the statistical characteristics of the input deformation gradient field, including:

[0085] Lightweight sub-models are enabled in low-variance regions;

[0086] High-precision sub-models are used in high-variance regions;

[0087] The thermal influence factor node is associated with the distance of the battery heat source based on the spatial position of the rear longitudinal beam, and the attenuation coefficient of local temperature rise on the yield strength of aluminum alloy is calculated; among them, the statistical characteristics include the strain distribution variance.

[0088] The reduced-order collision surrogate model is constructed through a pre-trained neural network. Its input is the structural deformation gradient field, and its output is the collision energy absorption value.

[0089] The alternative data channel generates approximate solutions based on a simplified energy conservation equation.

[0090] It should be further explained that, in the specific implementation process, the reduced-order collision proxy model is built based on a pre-trained neural network, and its training data covers the mapping relationship between deformation gradient fields and energy absorption values ​​for multiple collision scenarios; among these, the multiple collision scenarios include frontal, side, and rear-end collisions; during dynamic compilation and invocation:

[0091] Input preprocessing: The structural deformation gradient field is divided into local blocks according to the spatial grid, and the variance of strain distribution in each block is calculated;

[0092] Dynamic model version selection: If the block variance is lower than the calibration threshold and there are no discontinuous deformation features, a lightweight sub-model with ≥50% compression layers is called; if the block variance exceeds the standard or there are material folding patterns, the model is switched to a high-precision sub-model with added residual deformation feedback channels.

[0093] Output calibration: The proxy model results are compared with the mean of the adversarial network historical database. If the deviation exceeds the limit, the alternative channel is triggered.

[0094] The alternative data channel generates approximate solutions based on a simplified energy conservation equation, including the following:

[0095] The startup conditions include:

[0096] The proxy model outputs a confidence level lower than the dynamic threshold;

[0097] The real-time computational load at the collision node exceeds the hardware processing limit.

[0098] The system is activated if any one of the conditions is met.

[0099] Energy equivalent conversion rule: Decompose the structural deformation gradient field into elastic deformation energy and plastic dissipation energy; correlate the kinetic energy absorption with the plastic dissipation energy increment through the principle of momentum conservation; output the equivalent energy absorption value and add a low confidence mark;

[0100] Channel switching mechanism: During the operation of the alternative channel, the original collision node is switched to background calculation; when the background calculation is completed and the deviation between the result and the approximate solution is less than the tolerance threshold, the system silently switches back to the main channel.

[0101] The operation mode of the physical constraint verification layer in step S4 is as follows:

[0102] The correlation threshold between the strain rate effect and the thermal softening effect of the built-in aluminum alloy;

[0103] When the optimization result deviates from the constitutive equation, local data resampling is triggered.

[0104] It should be further explained that, in the specific implementation process, the physical constraint verification layer monitors the stress-temperature field data output by the topology optimization in real time, and its built-in correlation thresholds for the aluminum alloy strain rate effect and thermal softening effect include the following:

[0105] Dynamic load yield decay threshold: Based on the aluminum alloy grade, a critical curve for strength decay under different strain rates is set. When the actual strength at the monitoring point is lower than the allowable value of the curve, an early warning is triggered.

[0106] Battery thermal runaway conduction gradient threshold: The maximum allowable temperature rise rate is defined according to the battery pack thermal management constraints. If the gradient of the rear longitudinal beam heat conduction path exceeds the limit, it is marked as a high-risk area.

[0107] Local data resampling is triggered only when the same grid cell simultaneously meets the following two conditions:

[0108] Condition 1: Strain rate > calibrated value and actual yield strength < corresponding value of attenuation threshold;

[0109] Condition 2: Temperature gradient > conduction threshold and distance to adjacent battery packs < set value;

[0110] Resampling execution rules include:

[0111] Local control: Data acquisition is restarted only for the trigger grid and its adjacent cells, while the optimization results of other areas are preserved;

[0112] Operating condition adaptation: If the triggering cause is high strain rate, then the collision load input is strengthened during resampling; if it is excessive thermal gradient, then the intensity of heat source simulation is increased.

[0113] Before the resampling results are output, they need to be verified quickly by the adversarial network discriminator.

[0114] The criteria for setting the correlation threshold include:

[0115] The attenuation characteristics of aluminum alloy yield strength under dynamic load;

[0116] Critical temperature conduction gradient in battery thermal runaway scenarios.

[0117] It should be further explained that, in the specific implementation process, the rules for setting the yield strength attenuation characteristics of aluminum alloys under dynamic loads are as follows:

[0118] Material behavior database matching: Based on the specified aluminum alloy grade, the pre-stored test database is indexed to obtain the material's performance in the typical strain rate range of 10. -3 ~10 3 s -1 The yield strength decay curve;

[0119] Operating condition adaptation corrections include the following:

[0120] Frontal collision: using a medium-to-high strain rate range of 10 1 ~10 2 s -1 The attenuation curve focuses on the energy absorption characteristics of uniform deformation.

[0121] Side column impact: using an ultra-high strain rate range >10 2 s -1 Attenuation curve, enhancing local tear resistance threshold;

[0122] When the actual strain rate in the optimization results exceeds the range of the database, the attenuation trend is extrapolated based on the Hollomon hardening model. The extrapolation results need to be tested for material confidence by the adversarial network.

[0123] The rules for setting the critical temperature conduction gradient in battery thermal runaway scenarios include the following:

[0124] Heat source-structure path modeling: Based on the spatial topological relationship between the battery pack and the rear longitudinal beam, the distance ≤200mm is divided into the near field region and the distance >200mm is divided into the far field region; the equivalent conduction gradient of the near field region is calculated according to the heat flux density distribution, and the model of the far field region is simplified according to Fourier's law;

[0125] Critical gradient calibration: Near-field gradient threshold = Battery thermal runaway critical temperature / Shortest heat conduction path length × Safety factor; Far-field gradient threshold = Near-field threshold × Distance attenuation factor;

[0126] Thermal shock superposition rule: If an alternative channel is enabled or resampling is triggered, the gradient threshold is temporarily lowered to enhance the safety margin.

[0127] The dual-channel adversarial network operates as follows in step S5:

[0128] The generator synthesizes extreme collision condition data based on the current optimization results;

[0129] The discriminator compares the statistical distribution differences between simulation data and the historical database of physical experiments.

[0130] It should be further explained that, in the specific implementation process, when the generator synthesizes extreme collision condition data based on the current topology optimization results, it performs spatial weak zone identification and condition enhancement, as follows:

[0131] Identify high-stress concentration areas in the rear longitudinal beam, such as joints and bends, and extract the strain energy density distribution of the mesh elements. If the strain energy density is greater than the historical average, then superimpose two types of strengthening loads on this area, including the following two strengthening methods:

[0132] Collision-driven enhancement: Injecting random road surface excitation time-domain waves into the original working condition and increasing the offset collision vector angle;

[0133] Thermally driven enhancement: A transient thermal shock waveform is added to the area adjacent to the battery to simultaneously increase the intensity of the heat source;

[0134] Before the generated data is output to the discriminator, it is normalized to a standard statistical distribution format.

[0135] When the discriminator compares the simulation data with the physical experiment history database, it performs the following operations:

[0136] Distribution feature extraction: For simulation data and historical data under the same working conditions, calculate the third and fourth moments of strain energy density distribution, i.e. skewness difference and kurtosis difference.

[0137] Difference quantification rules: If the skewness difference is less than the tolerance value and the kurtosis difference is less than the tolerance value, the confidence level is considered to be up to standard; if the difference of either moment exceeds the standard, the KL divergence value is calculated and compared with the dynamic calibration threshold.

[0138] Confidence level output: High confidence, i.e., both moments meet the standard, then the verification is marked as passed; Medium confidence, i.e., KL divergence meets the standard but one moment exceeds the standard, then the physical verification layer local review is triggered; Low confidence, i.e., KL divergence exceeds the standard, then the backtracking mechanism is activated.

[0139] It also includes confidence level determination and backtracking mechanisms:

[0140] When the confidence level of the discriminator output is lower than the dynamic calibration threshold, the process backtracks to the key coupling node marked in the computation graph.

[0141] The physical test history database indexes new energy vehicle crash test data through a transfer learning framework.

[0142] It should be further explained that, in the specific implementation process, when the confidence level of the discriminator output is lower than the dynamic calibration threshold, backtracking is performed, and the process is as follows:

[0143] Global deviation, i.e., when the KL divergence exceeds the limit across the board, backtrack to the initial node of the computation graph and completely restart for optimization;

[0144] Local anomalies, i.e., when a specific physical field interaction node exceeds the limit, only the marked critical coupling node is located and restarted, including:

[0145] Structure-collision interaction layer: when the timing of stress wave transmission and collision kinetic energy absorption is mismatched;

[0146] Thermo-structural coupling layer: When the temperature rise gradient causes the yield strength to decrease beyond the correlation threshold;

[0147] The backtracking scope is controlled to include the problematic node and three directly downstream nodes, while the data of the upstream verified nodes is frozen and inherited.

[0148] The backtracking restart optimization process includes execution condition enhancements, including:

[0149] Data inheritance rules: retain all optimization results for the unbacktracked areas; inherit the stress-temperature field data that was most recently validated in the physical verification layer;

[0150] Input reinforcement rules: If collision data is abnormally backtracked, the collision condition amplitude is increased to the upper limit of the generator's synthesized value; if thermal data is abnormally backtracked, the heat source intensity is superimposed on the transient impact waveform and the duration is extended.

[0151] Validation acceleration mechanism: The first output after restarting skips the full distribution validation and only performs a fast comparison of moment differences; if the validation fails twice in a row, a high-precision surrogate model is forcibly switched.

[0152] When indexing data in the physics experiment history database using a transfer learning framework, priority rules and anomaly filtering are included; among them:

[0153] Priority rule: Historical data of the same vehicle platform > Data of the same battery type > Basic material test data;

[0154] Anomaly filtering: Automatically removes historical records with calibration errors exceeding industry standards.

[0155] Step S6: Perform post-processing before output, including manufacturing process constraint mapping and topology smoothness filtering.

[0156] It should be further explained that, in the specific implementation process, post-processing is performed before outputting the final rear longitudinal beam topology, including the following processes:

[0157] Manufacturing process constraint mapping: Identify unit regions in the topology with wall thickness < minimum allowable value, and expand the local cross section according to the aluminum alloy stamping forming limit; if there is an inward acute angle with a radius of curvature < 3 times the plate thickness, it is automatically filled as a process fillet and inherits the original unit mechanical properties; when the region data retained by the backtracking mechanism conflicts with the newly added process structure, the manufacturing constraints are satisfied first and the physical verification layer is triggered for local review.

[0158] Topological smoothness filtering: For high stress gradient transition regions where the strain energy density change rate is greater than the calibration value, a Gaussian smoothing kernel is applied, and the iteration continues until the curvature is continuous; for low confidence regions marked by adversarial networks, the smoothness level is forcibly increased, and the number of smoothing iterations is increased; the smoothed structure needs to be quickly verified by a reduced-order collision proxy model.

[0159] It should be further explained that, in the specific implementation process, when executing the method, the initial design parameters of the rear longitudinal beam of the new energy vehicle are first input, including: geometric dimension range, material grade identification and installation point coordinates; multi-physics boundary conditions, including static battery pack gravity distribution and dynamic road excitation spectrum in structural loads, three types of standards in collision conditions: acceleration pulse waveform, and battery pack thermal runaway critical temperature and allowable temperature rise gradient in thermal management constraints.

[0160] When constructing the constitutive model of aluminum alloy materials, the upper limit threshold of the mass of the lightweight target is correlated with the minimum energy absorption threshold of the collision safety requirement. When the two conflict in a specific grid area, the spatial coordinates of that area are marked as high-priority optimization targets.

[0161] By integrating and optimizing the process through a unified computational graph architecture, the original heterogeneous data of structural stress field, collision dynamic field and thermodynamic field are transformed into a domain adaptive tensor. This tensor is organized according to the spatial coordinates of the discrete grid of the rear longitudinal beam, the time step sequence dynamically divided according to the working condition type, and the dimensions including multiple physical quantities such as displacement, stress and temperature.

[0162] During the transformation, the physical field feature decoupling layer automatically identifies coupling regions. For example, when a mesh cell simultaneously detects a high temperature gradient and a high stress value, a feature stripping operation is initiated. A spatial convolution kernel is applied to the thermal field to separate pure heat conduction features, and a temporal filtering kernel is applied to the collision field to separate inertial effect features. Subsequently, the data are mapped to a unified tensor channel according to the mesh coordinates. If different physical field data are time-asynchronous, such as stress wave propagation delay exceeding the step size ratio, nonlinear interpolation is used to align the time series.

[0163] When generating the computational graph using dynamic compilation technology, the data dependencies between nodes are analyzed, and nodes forcing thermal influence factors to be executed before structural field solution nodes are prioritized. If there are no direct dependencies between nodes, parallel scheduling is employed. When a nonlinear collision calculation node is detected as potentially causing real-time performance limitations, an alternative data channel is automatically generated: for conventional energy absorption scenarios, a pre-trained reduced-order collision proxy model is invoked, which dynamically selects a lightweight or high-precision sub-version based on the strain distribution variance of the input deformation gradient field; for highly nonlinear failure scenarios, an equivalent energy absorption value is output based on a simplified energy conservation equation, with an added confidence level. During the operation of the alternative channel, the status of the original node is continuously monitored, and the system automatically switches back to the main data stream upon completion.

[0164] The physical constraint verification layer monitors the stress-temperature coupling data in the optimization results in real time. When the same mesh element simultaneously violates the strain rate-related yield strength decay threshold and the thermal management-related temperature rise gradient threshold, local resampling is triggered. The resampling range is limited to the anomalous element and its adjacent meshes: if it is caused by high strain rate, the collision load input is strengthened; if it is caused by excessive thermal gradient, the heat source intensity is increased. The sampling results need to be compared with the distribution mean through a fast verification mode.

[0165] When verifying the optimization results using a dual-channel adversarial network, the generator first identifies the high strain energy density region of the rear longitudinal beam and synthesizes extreme working condition data accordingly: collision-dominated data superimposed with random road surface excitation and offset collision vectors, and thermodynamic-dominated data injected with transient thermal shock waveforms. The discriminator extracts the third-order moment skewness and fourth-order moment kurtosis features from the simulation data and the physical test history database. If the difference between the two moments is below the tolerance value, the verification is passed directly; if a single moment exceeds the limit, the KL divergence is calculated. Based on the confidence level, high-confidence regions are directly output, medium-confidence regions trigger local verification of the physical verification layer, and low-confidence regions initiate a backtracking mechanism.

[0166] The backtracking operation precisely locates the anomaly based on its type: for global deviations, the entire computation graph is restarted; for local mismatches, backtracking is only performed to the marked key coupling nodes, such as the structural collision interaction layer or the thermo-structural coupling layer, and valid upstream data is frozen. After restarting, input parameters are enhanced: for collision anomalies, the load amplitude is increased to the generator's upper limit; for thermal anomalies, the thermal shock duration is extended, and a rapid verification process is enabled. The physical test history library uses a transfer learning framework to prioritize data from the same vehicle platform and automatically filters records exceeding industry error standards.

[0167] Before the final rear longitudinal beam topology is output, manufacturing constraint mapping is performed: areas with insufficient wall thickness are identified, and the cross-section is expanded according to the stamping direction; when concave acute angles are detected, process fillets are filled, and the mechanical properties of the original unit are inherited. Adaptive Gaussian smoothing kernel weights are applied to the high-stress gradient transition region to improve the smoothness level of low-confidence regions marked by the adversarial network. The processed structure needs to be quickly verified by a reduced-order collision proxy model.

[0168] It should be further explained that, in the specific implementation process, data collaboration within the unified computation graph architecture includes the following:

[0169] Feature decoupling self-identification mechanism: Feature stripping is initiated only when the product of the mesh element temperature gradient and the equivalent stress exceeds the material property correlation value to avoid invalid calculations;

[0170] Time-asynchronous processing rules: Linear interpolation is used in the low strain rate region, and nonlinear interpolation based on the material hardening curve is used in the high plastic deformation region;

[0171] Proxy model version switching: Lightweight sub-models are used for regions with uniform strain distribution. High-precision sub-models are automatically activated in meshes with material folding patterns.

[0172] The reliability guarantees of the dual-modal verification mechanism include the following:

[0173] Physical threshold combination triggering: the yield strength decay threshold is dynamically matched with the aluminum alloy grade, and the thermal gradient threshold is set according to the battery distance. Failure is only determined when both thresholds exceed the limit at the same time.

[0174] Adversarial network enhancement synthesis: Increase the bias collision angle for high-stress areas at the joints and improve the thermal shock amplitude in the battery proximity area;

[0175] Confidence level response: The moment difference tolerance value is associated with the historical database statistical standard deviation and the KL divergence threshold, which dynamically fluctuates with the optimization process status.

[0176] Backtracking for precise positioning: The mismatch in the structural collision interaction layer is manifested as the phase difference between the stress wave propagation time history and the collision kinetic energy absorption curve exceeding the allowable error band.

[0177] Engineering manufacturability conversion rules include:

[0178] Process conflict handling: When a newly added rounded corner structure conflicts with backtracked retained data, local mesh verification is triggered first, prioritizing manufacturing constraints over a global restart;

[0179] Smoothness graded control: In the high stress gradient region, Gaussian kernels of different radii are matched according to the strain energy density change rate, and the number of smoothing iterations is multiplied in the low confidence region;

[0180] Quick verification and fault tolerance: When a deviation occurs in the collision-dominant region, it reverts to the pre-smoothing state. The thermal-dominant region maintains output but with additional special monitoring flags.

[0181] It should be further explained that, in the specific implementation process, a design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization includes the following steps:

[0182] Step S1: Input the initial design parameters and multiphysics boundary conditions of the rear longitudinal beam, including the range of geometric dimensions, material grade identification, structural load distribution, three types of collision acceleration waveforms, and battery pack thermal management constraint range.

[0183] Step S2: Construct a constitutive model of the aluminum alloy material, correlate the upper limit threshold of the mass of the lightweight target with the minimum energy absorption threshold of the collision safety, and mark the spatial coordinates of the conflict area.

[0184] Step S3: Perform multi-field coupled topology optimization through a unified computational graph architecture: convert the raw data of structural stress field, collision dynamics field, and thermodynamic field into a domain adaptive tensor, including:

[0185] The spatial dimension is discretized according to the rear longitudinal beam mesh and bound to material properties;

[0186] The time dimension is dynamically segmented and sampled according to the collision-dominant segment and the thermal-dominant segment;

[0187] The physical quantity dimension integrates parameters such as displacement, stress, temperature, and heat flux density;

[0188] The feature decoupling layer automatically identifies coupling regions: when a mesh cell simultaneously detects a high temperature gradient and a high stress value, a spatial convolution kernel is applied to the thermal field to strip away pure heat conduction features; a temporal filtering kernel is applied to the collision field to separate inertial effect features; asynchronous data is aligned to the time series through nonlinear interpolation;

[0189] Dynamically compile and generate computational graphs: enforce thermal nodes to be scheduled before structural field nodes; allow independent nodes to execute in parallel;

[0190] Real-time over-limit generation of alternative data channels: In normal scenarios, a reduced-order collision proxy model is called, and a lightweight or high-precision sub-model is dynamically selected according to the variance of the strain distribution; in highly nonlinear scenarios, the equivalent energy absorption value is output based on the energy conservation equation and a confidence flag is attached.

[0191] Step S4: Embedding a physical constraint verification layer and a dual-channel adversarial network for verification, including the following:

[0192] The verification layer monitors stress-temperature field data in real time: local resampling is triggered only when the same element simultaneously violates the strain rate-related yield decay threshold and the thermal gradient threshold; the sampling range is limited to anomalous elements and their neighborhoods, and the input load or heat source is enhanced according to the anomalous type;

[0193] Adversarial network verification: The generator identifies regions with high strain energy density and synthesizes collision-type or thermal-type extreme conditions;

[0194] The discriminator compares the third-order moment skewness and fourth-order moment kurtosis of the simulation data with the historical database: if the difference between the two moments meets the standard, it passes directly; if the difference between the single moment exceeds the standard, the KL divergence of the whole distribution is calculated.

[0195] Output three levels of confidence response: High confidence: output the result; Medium confidence: trigger local verification of the physical verification layer; Low confidence: activate the backtracking mechanism.

[0196] Step S5: Perform confidence-driven backtracking and optimized restart, including:

[0197] Accurately locate abnormal nodes at low confidence levels: if there is a global deviation, restart the entire computation graph; if there is a local mismatch, only backtrack to the key coupled nodes and freeze the inherited valid upstream data.

[0198] Enhanced input after restart: Increase load to the generator synthesis limit when collision anomalies occur; extend thermal shock duration when thermal anomalies occur; enable fast verification process, only perform moment difference comparison.

[0199] Step S6: Perform post-processing before outputting the topology structure, including:

[0200] Manufacturing constraints are mapped: the cross-section of thin-walled regions is expanded along the stamping direction, and the concave acute angles are filled with process fillets; adaptive Gaussian smoothing is applied to high-stress gradient regions, and the smoothing iteration number is doubled in low-confidence regions; the final structure is output after fast verification through a reduced-order collision proxy model.

[0201] By employing dynamic compilation technology and a domain-adaptive tensor transformation mechanism within a unified computational graph architecture, the inefficiency of multi-field data interaction in traditional toolchains is addressed. The feature decoupling layer automatically removes redundant features from heterogeneous physical fields, intelligently avoiding high-latency nodes in place of data channels, and enabling real-time synchronous iteration of structure-collision-thermal coupling simulations. Combined with a condition-adaptive reduced-order proxy model and embedded physical constraint verification, the single iteration cycle is shortened to the fragment-level time consumption of traditional methods while maintaining optimization accuracy, effectively accelerating the lightweight development process of key components for new energy vehicles.

[0202] Based on a dual-channel adversarial network and a confidence backtracking mechanism, the reliability of topology optimization results is guaranteed to an engineering-level standard. The generator synthesizes extreme working conditions to enhance verification coverage in weak areas, while the discriminator determines the confidence level through distribution moment differences and KL divergence classification. The physical verification layer uses a dual-threshold combination triggering mechanism to avoid invalid interruptions, and node-level precise backtracking inherits valid data and strengthens input. This system reduces the need for physical experimental verification to a low level, ensuring a fundamental synergistic breakthrough in collision safety and lightweighting of the aluminum alloy rear longitudinal beam, while simultaneously generating constraint mapping rules to directly output a mass-producible topology structure.

[0203] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0204] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization, characterized in that, Includes the following steps: Step S1: Input the initial design parameters and multiphysics boundary conditions for the rear longitudinal beam; Step S2: Construct a constitutive model of the aluminum alloy material; Step S3: Perform multi-field coupled topology optimization through a unified computation graph architecture; the implementation method of the unified computation graph architecture in step S3 is as follows: The structural stress field, collision dynamics field, and thermodynamic field data are encoded into a domain adaptive tensor, which includes spatial coordinates, time steps, and physical quantity types. Transform heterogeneous data formats through a physical field feature decoupling layer; Step S4: Embed a physical constraint verification layer at the output of the computation graph; the operation mode of the physical constraint verification layer in step S4 is as follows: The correlation threshold between the strain rate effect and the thermal softening effect of the built-in aluminum alloy; When the optimization result deviates from the constitutive equation, local data resampling is triggered; Step S5: Verify the optimization results using a dual-channel adversarial network; Step S6: Output the topology of the rear longitudinal beam.

2. The design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to claim 1, characterized in that: The multiphysics boundary conditions mentioned in step S1 include structural loads, collision conditions, and thermal management constraints; The constitutive model described in step S2 associates the lightweight target with the collision safety threshold.

3. The design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to claim 1, characterized in that: Step S3 also includes dynamically compiling and generating the computation graph: Integrating a reduced-order collision proxy model with thermal influence factor nodes; Nodes are scheduled in parallel based on data dependencies; For nonlinear collision calculation nodes, alternative data channels are generated under the condition of satisfying real-time constraints.

4. The design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to claim 3, characterized in that: The reduced-order collision proxy model is constructed through a pre-trained neural network, with the structural deformation gradient field as input and the collision energy absorption value as output. The alternative data channel generates an approximate solution based on a simplified energy conservation equation.

5. The design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to claim 1, characterized in that: The criteria for setting the association threshold include: The attenuation characteristics of aluminum alloy yield strength under dynamic load; Critical temperature conduction gradient in battery thermal runaway scenarios.

6. The design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to claim 1, characterized in that: The dual-channel adversarial network described in step S5 operates as follows: The generator synthesizes extreme collision condition data based on the current optimization results; The discriminator compares the statistical distribution differences between simulation data and the historical database of physical experiments.

7. The design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to claim 6, characterized in that: It also includes confidence level determination and backtracking mechanisms: When the confidence level of the discriminator output is lower than the dynamic calibration threshold, the process backtracks to the key coupling node marked in the computation graph. The physical test history database indexes new energy vehicle crash test data through a transfer learning framework.

8. The design method for aluminum alloy rear longitudinal beams of new energy vehicles based on topology optimization according to claim 1, characterized in that: Step S6: Perform post-processing before output, including manufacturing process constraint mapping and topology smoothness filtering.

Images