Reverse design method for stealth structure parameters based on diffusion model and multi-modal generation

Abstract

The application discloses a stealth structure parameter reverse design method based on a diffusion model and multi-modal generation, comprising the following steps: S1, a bidirectional generation model is constructed; wherein the bidirectional generation model comprises a diffusion model and a generative adversarial network; S2, a text description is input into a CLIP model for mapping and unification to obtain a multi-modal demand vector, the multi-modal demand vector is spliced to obtain spliced modal demand features, and the spliced modal demand features are input into the diffusion model to obtain a structure parameter combination; S3, the structure parameter combination and the generative adversarial network are input into a simulation-generation closed-loop optimization platform, and the simulation-generation closed-loop optimization platform outputs a candidate parameter set and a gradient of the generative adversarial network; and S4, the generative adversarial network in the step S1 is updated according to the gradient of the generative adversarial network, and the step S3 is returned. The application realizes reverse generation from performance requirements to parameter combination, and improves the design efficiency and innovation capability of the stealth structure.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of electromagnetic stealth, generative deep learning, and structural topology optimization, and particularly relates to a method for reverse design of stealth structure parameters based on diffusion models and multimodal generation. Background Technology

[0002] Traditional stealth structure parameter design suffers from three core pain points: First, forward optimization is inefficient. Simulation-based forward optimization (such as genetic algorithms) requires hundreds of iterations to find a feasible solution, with a single full-band electromagnetic simulation taking hours and design cycles lasting weeks or even months. Second, innovative design capabilities are insufficient. Traditional methods rely on empirical parameter adjustments, making it difficult to overcome the limitations of existing structural configurations (such as honeycomb and stepped cone structures) and generate stealth structures with novel topologies. Third, balancing multiple objectives is challenging. When simultaneously optimizing broadband stealth (e.g., 2-40GHz reflectivity ≤-10dB) and lightweighting (weight ≤0.5kg), traditional algorithms are prone to getting trapped in local optima, leading to mutual constraints on performance indicators.

[0003] In existing technologies, some studies have attempted to use deep learning for parameter prediction, but these can only achieve a forward mapping between parameters and performance, and cannot reverse-engineer parameter combinations that meet specific performance requirements. For example, when the target reflectivity bandwidth is known to be 15 GHz, traditional neural networks struggle to reverse-engineer the corresponding material dielectric constant and lattice configuration parameters. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a reverse design method for stealth structure parameters based on diffusion model and multimodal generation, so as to realize the reverse generation from performance requirements to parameter combinations, thereby improving the design efficiency and innovation capability of stealth structures.

[0005] The objective of this invention is achieved through the following technical solution: a method for reverse design of stealth structural parameters based on a diffusion model and multimodal generation, comprising: Step S1: constructing a bidirectional generative model; wherein the bidirectional generative model includes a diffusion model and a generative adversarial network (GAN); Step S2: inputting text descriptions into a CLIP model for mapping and unification to obtain multimodal demand vectors, concatenating the multimodal demand vectors to obtain concatenated modal demand features, and inputting the concatenated modal demand features into a diffusion model to obtain a combination of structural parameters; Step S3: inputting the combination of structural parameters and the GAN into a simulation-generative closed-loop optimization platform, and the simulation-generative closed-loop optimization platform outputting a set of candidate parameters and the gradient of the GAN; Step S4: updating the GAN in step S1 according to the gradient of the GAN, and returning to step S3.

[0006] In the above-mentioned reverse design method for stealth structure parameters based on diffusion model and multimodal generation, the candidate parameter set is refined by cellular automata algorithm to obtain refined parameters, and the real sample is obtained by 3D printing based on the refined parameters.

[0007] In the aforementioned method for reverse design of stealth structure parameters based on diffusion model and multimodal generation, the diffusion model is used as the core generator and the generative adversarial network is used as the discriminator in the bidirectional generative model.

[0008] In the aforementioned reverse design method for stealth structure parameters based on diffusion models and multimodal generation, the structural parameters include material parameters, geometric parameters, and topology type parameters.

[0009] In the aforementioned reverse design method for stealth structure parameters based on diffusion model and multimodal generation, the simulation-generation closed-loop optimization platform uses reflectivity data and mechanical performance indicators as feedback signals to update the loss function of the bidirectional generation model. The simulation-generation closed-loop optimization platform introduces a reinforcement learning mechanism and designs a reward function that includes electromagnetic performance compliance, mechanical performance compliance, and weight constraints to guide the bidirectional generation model to prioritize the generation of highly feasible parameters.

[0010] In the aforementioned reverse design method for stealth structure parameters based on diffusion model and multimodal generation, the multimodal requirements include electromagnetic performance requirements and mechanical performance requirements; among them, electromagnetic performance includes reflectivity curve, effective bandwidth, and reflectivity threshold of specific frequency bands; mechanical performance includes bending strength, tensile strength, and weight constraints.

[0011] In the aforementioned reverse design method for stealth structure parameters based on diffusion model and multimodal generation, the candidate parameters include candidate material parameters, candidate geometric parameters, and candidate topology type parameters.

[0012] In the aforementioned inverse design method for stealth structure parameters based on diffusion model and multimodal generation, the cellular automata algorithm generates innovative topological structures with fractal characteristics and multi-level porosity by customizing seed parameters and iteration count.

[0013] A stealth structural parameter inverse design system based on a diffusion model and multimodal generation includes: a generative model module for constructing a bidirectional generative model, wherein the bidirectional generative model includes a diffusion model and a generative adversarial network (GAN); a requirement parsing module for mapping and unifying textual descriptions into a CLIP model to obtain multimodal requirement vectors, concatenating the multimodal requirement vectors to obtain concatenated modal requirement features, and inputting the concatenated modal requirement features into the diffusion model to obtain a combination of structural parameters; a co-simulation module for inputting the combination of structural parameters and the GAN into a simulation-generation closed-loop optimization platform, wherein the simulation-generation closed-loop optimization platform outputs a set of candidate parameters and the gradient of the GAN; and an update module for generating the GAN according to the gradient update steps of the GAN.

[0014] An electronic device includes: a memory for storing computer-readable instructions; and a processor for executing the computer-readable instructions to perform a reverse design method for stealth structure parameters based on a diffusion model and multimodal generation.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] This invention enables reverse generation from performance requirements to parameter combinations, improving the design efficiency and innovation capabilities of stealth structures. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0018] Figure 1 This is a flowchart of the reverse design method for stealth structure parameters based on diffusion model and multimodal generation provided in this embodiment of the invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Figure 1This is a flowchart of the reverse design method for stealth structure parameters based on diffusion models and multimodal generation provided in this embodiment of the invention. Figure 1 As shown, the reverse design method for stealth structure parameters based on diffusion model and multimodal generation includes:

[0021] Step S1: Construct a bidirectional generative model; wherein, the bidirectional generative model includes a diffusion model and a generative adversarial network;

[0022] Step S2: Input the text description into the CLIP model for mapping and unification to obtain a multimodal demand vector. Concatenate the multimodal demand vectors to obtain concatenated modal demand features. Input the concatenated modal demand features into the diffusion model in the bidirectional generative model to obtain a combination of structural parameters. The structural parameters include material parameters, geometric parameters, and topology type parameters.

[0023] Step S3: Input the structural parameter combination and the generative adversarial network in the bidirectional generative model into the simulation-generative closed-loop optimization platform. The simulation-generative closed-loop optimization platform outputs the candidate parameter set and the gradient of the generative adversarial network. 3D print the real sample based on the candidate parameter set. The candidate parameters include candidate material parameters, candidate geometric parameters, and candidate topology type parameters.

[0024] Step S4: Update the generative adversarial network in step S1 based on the gradient of the generative adversarial network, and return to step S3.

[0025] The candidate parameter set is refined using cellular automata to obtain refined parameters, and the real sample is obtained by 3D printing based on the refined parameters.

[0026] In the bidirectional generative model, a diffusion model serves as the core generator, and a generative adversarial network (GAN) acts as the discriminator. Based on performance requirements (such as reflectivity curves and mechanical properties), parameter combinations (material parameters, geometric dimensions, and topological configuration) are generated inversely. A GAN is used to construct the discriminator, which includes electromagnetic and mechanical discriminant branches, used to determine the electromagnetic and mechanical feasibility of the generated parameters, respectively. The diffusion model employs a U-Net backbone network, using a time-step encoder and a performance-conditional encoder to conditionally generate multi-dimensional parameters. The discriminator, through an adversarial training mechanism, forces the generator to output parameter combinations that are closer to real-world feasibility.

[0027] The multimodal conditional input mechanism supports multimodal requirement descriptions such as text input (e.g., "Design X-band lightweight stealth structure"), curve input (target reflectivity curve), and numerical input (mechanical performance indicators). The CLIP model maps multimodal requirements to a unified feature space to obtain different modal features. For example, textual requirements are converted into semantic feature vectors, and reflectivity curves are converted into frequency domain feature vectors. These different modal features are then concatenated and used as conditional inputs to the diffusion model, achieving unified processing and parameter generation guidance across modal requirements.

[0028] The simulation-generation closed-loop optimization platform uses reflectivity data and mechanical performance indicators as feedback signals to update the loss function of the bidirectional generative model. The simulation-generation closed-loop optimization platform introduces a reinforcement learning mechanism and designs a reward function that includes electromagnetic performance compliance, mechanical performance compliance and weight constraints to guide the bidirectional generative model to prioritize the generation of highly feasible parameters.

[0029] A simulation-generation closed-loop optimization framework is adopted, which performs electromagnetic-mechanical co-simulation on the generated parameter combinations. The reflectivity data obtained from CST simulation and the mechanical performance indicators calculated by ANSYS are used as feedback signals to update the loss function of the generated model. A reinforcement learning (RL) mechanism is introduced, and a reward function that includes electromagnetic performance compliance, mechanical performance compliance, and weight constraints is designed to guide the generated model to prioritize the generation of highly feasible parameters. Through an iterative process of "generation-simulation-feedback-update", the engineering applicability of the generated parameters is gradually improved until the reward values ​​of the parameters generated in consecutive iterations tend to stabilize.

[0030] Multimodal requirements include electromagnetic performance requirements and mechanical performance requirements; among them, electromagnetic performance includes reflectivity curve, effective bandwidth, and reflectivity threshold of specific frequency bands; mechanical performance includes bending strength, tensile strength, and weight constraints.

[0031] The cellular automata algorithm generates innovative topological structures with fractal characteristics and multi-level porosity by customizing seed parameters and iteration counts.

[0032] Two-way generative model construction: Using a diffusion model as the core generator and a generative adversarial network as the discriminator, an inverse generative mapping of "performance conditions-parameter combinations" is constructed; Multimodal input requirements: Supports multimodal performance requirement inputs such as text, curves, and numerical values, which are mapped to a unified feature space through the CLIP model; Simulation-generation closed-loop optimization: Electromagnetic-mechanical joint simulation is performed on the generation parameters, and the generation model is updated through reinforcement learning to form an iterative closed loop of "generation-simulation-feedback"; Innovative topology design: Through 3-10 iterations of cellular automata, a novel lattice topology containing fractal features is generated, breaking through the limitations of traditional honeycomb / wooden stack configurations.

[0033] Multimodal requirements include: electromagnetic performance requirements: reflectivity curve, effective bandwidth, reflectivity threshold for specific frequency bands; mechanical performance requirements: bending strength, tensile strength, weight constraints.

[0034] The simulation-generated closed-loop optimization reward function comprehensively considers electromagnetic performance compliance, mechanical performance compliance, and weight constraints, and the weight coefficients are determined through cross-validation.

[0035] The process of generating a topology using a cellular automaton includes three steps: initializing the seed, iteratively updating the grid, and generating the topology. The number of iterations is 3-10.

[0036] By combining cellular automata algorithms to generate novel lattice topologies, the limitations of traditional honeycomb and woodpile configurations are overcome. By customizing seed parameters and the number of iterations, innovative topological structures with fractal characteristics and multi-level porosity can be generated. For example, a fractal honeycomb topology generated through 5-fold cellular iteration can increase porosity by 20% compared to traditional structures while maintaining mechanical strength, providing a new configurational basis for the synergistic optimization of broadband stealth and lightweight design.

[0037] Bidirectional generative model construction and training:

[0038] (1) Diffusion model architecture design: (11) Model structure: U-Net is used as the backbone network, which includes 6 layers of downsampling and 6 layers of upsampling, and embeds time step encoder and performance condition encoder; (12) Parameter space definition: Continuous parameters include dielectric constant ε∈[2, 15], magnetic permeability μ∈[1, 5], and thickness d∈[5, 20] mm. Discrete parameters include lattice type (cell / woodpile / custom topology), encoded as one-hot vectors; (13) Training data: 100,000 parameter-performance pairs are generated by CST and ANSYS, where parameters include three categories: material, geometry, and topology, and performance includes reflectivity curves and mechanical indices.

[0039] (2) Multimodal conditional mapping: (21) Text to feature: The CLIP model is used to map the text requirements (such as "Ku band stealth and weight <0.3kg") into a 512-dimensional feature vector; (22) Curve to feature: The target reflectivity curve is converted into a frequency domain feature through Fourier transform, and then concatenated with the text features and input into the diffusion model.

[0040] (3) Adversarial training strategy: The discriminator D includes an electromagnetic discriminant branch D_E (judging reflectivity feasibility) and a mechanical discriminant branch D_M (judging strength feasibility). The loss function is as follows:

[0041] L = L_diffusion + λ1 L_adv_E + λ2 L_adv_M;

[0042] L_adv_E = -E[log(D_E(G(z|c)))] - E[log(1-D_E(x_real))];

[0043] Where G is the diffusion model generator, z is random noise, c is the performance condition, x_real is the true parameter sample, L is the total loss, L_diffusion is the diffusion model loss, L_adv_E is the electromagnetic loss, L_adv_M is the mechanical loss, λ1 is the electromagnetic loss weight, λ2 is the mechanical loss weight, E is the expectation, D_E(G(z|c)) is the electromagnetic probability value of the generated sample, D_E(x_real) is the electromagnetic probability value of the true sample, G(z|c) is the generated sample, and z|c is the input z under the condition c.

[0044] Multimodal performance requirement input and parameter generation: (1) Requirement analysis example: Input case: The user inputs "Design a lightweight structure with 2-18GHz reflectivity ≤-10dB and bending strength ≥120MPa"; Requirement analysis result: Electromagnetic requirements: 2-18GHz bandwidth ≥16GHz, reflectivity ≤-10dB; Mechanical requirements: bending strength ≥120MPa, weight ≤0.4kg. (2) Conditional parameter generation: Divided into two sub-parts: diffusion sampling and topology innovation. The former generates the parameter combination x_hat = G (z|c) through the reverse diffusion process under the constraint of performance condition c. The latter generates a new lattice configuration for the topology type in the discrete parameters through cellular automata.

[0045] Simulation-Generation Closed-Loop Optimization: (1) Co-simulation Verification: For electromagnetic simulation, the generated parameter x_hat is imported into CST, the reflectivity of 2-40GHz is simulated, and the actual bandwidth B_actual is calculated. For mechanical simulation, the bending strength S_actual and weight W_actual are imported into ANSYS to calculate. (2) Reinforcement Learning Feedback: The reward function is R=w1·f(B_actual≥16GHz)+w2·min(S_actual / 120,1)- w3·max(W_actual / 0.4, 1). Where w1=0.5, w2=0.3, w3=0.2, f is an indicator function that returns 1 when the condition is met, and 0 otherwise. The strategy is updated by using the PPO algorithm to update the generation strategy of the diffusion model, so that the reward value of the subsequent generation parameters gradually increases. (3) Iterative optimization process: Generate parameter x_1 → Simulate to obtain (B1, S1, W1) → Calculate R1 → Update model; Generate parameter x_2 → Simulate to obtain (B2, S2, W2) → Calculate R2 → Update model; Iterate until the parameter reward fluctuation of 5 consecutive generation is ≤5% and then terminate.

[0046] Innovative configuration output and verification: (1) Optimal parameter generation: The final output parameter combination is ε=7.2, μ=2.8, the topology is fractal honeycomb (cell iteration 5 times), d=9mm. Then the performance is predicted: B_pred=16.5GHz, S_pred=125MPa, W_pred=0.38kg. (2) Engineering verification: 3D printed samples are prepared and their performance is measured.

[0047] This embodiment also provides a stealth structural parameter inverse design system based on a diffusion model and multimodal generation. The system includes: a generative model module for constructing a bidirectional generative model, comprising a diffusion model and a generative adversarial network (GAN); a requirement parsing module for mapping and unifying textual descriptions into a CLIP model to obtain multimodal requirement vectors, concatenating the multimodal requirement vectors to obtain concatenated modal requirement features, and inputting the concatenated modal requirement features into the diffusion model to obtain a structural parameter combination; a co-simulation module for inputting the structural parameter combination and the GAN into a simulation-generation closed-loop optimization platform, which outputs a candidate parameter set and the gradient of the GAN; and an update module for generating the GAN based on the gradient update steps of the GAN.

[0048] This embodiment also provides an electronic device, including: a memory for storing computer-readable instructions; and a processor for running the computer-readable instructions to execute a reverse design method for stealth structure parameters based on diffusion models and multimodal generation.

[0049] This embodiment shortens the generation cycle from performance requirements to feasible parameters to within 24 hours, an improvement of over 30% compared to traditional methods. This embodiment can generate novel topologies that traditional methods cannot design, such as hybrid configurations of fractal honeycomb and woodpile structures, increasing the -10dB bandwidth to 18GHz. This embodiment achieves a multi-objective balance between reflectivity bandwidth, mechanical strength, and weight through the global search capability of the generated model.

[0050] This embodiment realizes the reverse design of stealth structure parameters based on generative AI, directly generating innovative parameter combinations from performance requirements, breaking through the efficiency and innovation bottlenecks of traditional design methods, and providing a new paradigm for the rapid iteration of stealth structures.

[0051] This embodiment utilizes Generative Adversarial Networks (GANs) and Diffusion Models to implement a method for reverse design of stealth structure parameters. By constructing a reverse design process of "performance requirements - parameter generation - feasibility verification", it solves the problems of low efficiency and insufficient innovative design capabilities of traditional forward optimization methods.

[0052] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A method for reverse design of stealth structure parameters based on diffusion models and multimodal generation, characterized in that... include: Step S1: Construct a bidirectional generative model; wherein, the bidirectional generative model includes a diffusion model and a generative adversarial network; Step S2: Input the text description into the CLIP model for mapping and unification to obtain a multimodal demand vector. Concatenate the multimodal demand vectors to obtain concatenated modal demand features. Input the concatenated modal demand features into the diffusion model to obtain the structural parameter combination. The multimodal requirements include electromagnetic performance requirements and mechanical performance requirements. Among them, electromagnetic performance includes reflectivity curve, effective bandwidth, and reflectivity threshold of a specific frequency band; mechanical performance includes bending strength, tensile strength, and weight constraints. Step S3: Input the structural parameter combination and generative adversarial network into the simulation-generated closed-loop optimization platform. The simulation-generated closed-loop optimization platform outputs a set of candidate parameters and the gradient of the generative adversarial network. The simulation-generated closed-loop optimization platform uses reflectivity data and mechanical performance indicators as feedback signals to update the loss function of the bidirectional generative model. The simulation-generated closed-loop optimization platform introduces a reinforcement learning mechanism and designs a reward function that includes electromagnetic performance compliance, mechanical performance compliance, and weight constraints to guide the bidirectional generative model to prioritize the generation of highly feasible parameters. Step S4: Update the generative adversarial network in step S1 based on the gradient of the generative adversarial network, and return to step S3.

2. The reverse design method for stealth structure parameters based on diffusion model and multimodal generation according to claim 1, characterized in that: The candidate parameter set is refined using a cellular automata algorithm to obtain refined parameters, and then 3D printing is performed based on the refined parameters to obtain a real sample.

3. The reverse design method for stealth structure parameters based on diffusion model and multimodal generation according to claim 1, characterized in that: In the bidirectional generative model, the diffusion model is used as the core generator, and the generative adversarial network is used as the discriminator.

4. The reverse design method for stealth structure parameters based on diffusion model and multimodal generation according to claim 1, characterized in that: Structural parameters include material parameters, geometric parameters, and topology parameters.

5. The reverse design method for stealth structure parameters based on diffusion model and multimodal generation according to claim 1, characterized in that: Candidate parameters include candidate material parameters, candidate geometric parameters, and candidate topology type parameters.

6. The reverse design method for stealth structure parameters based on diffusion model and multimodal generation according to claim 2, characterized in that: The cellular automata algorithm generates innovative topological structures with fractal characteristics and multi-level porosity by customizing seed parameters and iteration counts.

7. A reverse design system for stealth structure parameters based on diffusion models and multimodal generation, characterized in that... include: The generative model module is used to construct bidirectional generative models, which include diffusion models and generative adversarial networks. The requirement parsing module is used to input text descriptions into the CLIP model for mapping and unification to obtain multimodal requirement vectors. These multimodal requirement vectors are then concatenated to obtain concatenated modal requirement features. Finally, these features are input into the diffusion model to obtain structural parameter combinations. Multimodal requirements include electromagnetic performance requirements and mechanical performance requirements. Electromagnetic performance includes reflectivity curves, effective bandwidth, and reflectivity thresholds for specific frequency bands. Mechanical performance includes bending strength, tensile strength, and weight constraints. The co-simulation module is used to input structural parameter combinations and generative adversarial networks into the simulation-generation closed-loop optimization platform. The simulation-generation closed-loop optimization platform outputs a set of candidate parameters and the gradient of the generative adversarial network. The simulation-generation closed-loop optimization platform uses reflectivity data and mechanical performance indicators as feedback signals to update the loss function of the bidirectional generative model. The simulation-generation closed-loop optimization platform introduces a reinforcement learning mechanism and designs a reward function that includes electromagnetic performance compliance, mechanical performance compliance, and weight constraints to guide the bidirectional generative model to prioritize the generation of highly feasible parameters. The update module is used to generate the adversarial network based on the gradient update steps of the generative adversarial network.

8. An electronic device, characterized in that, include: Memory: Used to store computer-readable instructions; as well as Processor: for executing the computer-readable instructions to perform the method as described in any one of claims 1 to 6.

Images