Luneberg lens 3D printing material generation method based on diffusion model and reinforcement learning
By combining a diffusion model and a closed-loop framework of reinforcement learning with a property predictor, the problem of slow material design iteration speed in existing technologies is solved, and efficient, directional generation of polymer materials that meet the requirements of Luneburg lenses is achieved, thereby improving material quality and generation efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUBEI CHUCK TECH CO LTD
- Filing Date
- 2026-01-19
- Publication Date
- 2026-05-12
AI Technical Summary
Existing generative models focus on structural diversity and effective sampling capabilities in material design, but fail to establish an optimization mechanism that is tightly coupled with the target properties. This results in slow iteration speed, high cost, and lack of directionality in traditional material design.
A closed-loop framework based on diffusion model and reinforcement learning is constructed. Combined with an attribute predictor, molecular structures are generated through diffusion model, and the generation process is optimized by reinforcement learning. Multi-head attention mechanism and group relative advantage strategy are introduced to achieve targeted material generation.
This study achieved efficient and directional generation of polymer materials that meet the electromagnetic and mechanical performance requirements of Luneburg lenses, significantly improving generation efficiency and material quality while reducing experimental costs and resource consumption.
Smart Images
Figure CN122024952A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Luneburg lens printing and manufacturing technology, specifically to a method for generating Luneburg lens 3D printing materials based on diffusion models and reinforcement learning. Background Technology
[0002] Luneburg lens It is an electromagnetic lens Its radially continuously varying refractive index distribution enables it to achieve near-ideal aberration-free focusing of incident electromagnetic waves, thus finding wide application in radar systems, wireless communication antennas, and imaging equipment. With the development of additive manufacturing (3D printing) technology, to meet practical engineering needs, lens materials not only need to possess a specific dielectric constant range, but also must meet requirements for mechanical strength, density, and glass transition temperature. To meet multiple performance requirements, traditional material development methods typically rely on manual formulation design, experimental synthesis, and high-throughput screening. This process is time-consuming, costly, and the uncertainty of material properties is relatively large, making it difficult to shorten the research and development cycle. In recent years, generative models and reinforcement learning have been used in the generation of printed materials. Generative models, especially diffusion models, have shown strong capabilities in the field of molecular and material generation. Reinforcement learning is an optimization framework that can utilize states, actions, and rewards to form feedback loops.
[0003] Currently, most generative models focus on structural diversity and effective sampling capabilities, but have not established an optimization mechanism that is tightly coupled with the target properties. Their high dependence on reward signals limits their application in large-scale generative tasks, resulting in slow iteration speed in traditional materials design.
[0004] Therefore, there is an urgent need for a generation-evaluation-optimization closed-loop framework that can deeply integrate diffusion models, reinforcement learning, and high-precision property predictors to achieve efficient and directional generation of electromagnetic lens materials. To address the above technical pain points, a method for generating Luneburg lens 3D printing materials based on diffusion models and reinforcement learning is proposed. Summary of the Invention
[0005] This invention provides a method for generating Luneburg lens 3D printing materials based on diffusion models and reinforcement learning. It can effectively solve the problem that most of the generative models proposed in the background technology focus on structural diversity and effective sampling capabilities, but do not establish an optimization mechanism that is closely coupled with the target properties. Their high dependence on reward signals limits their application in large-scale generation tasks and leads to the slow iteration speed of traditional material design.
[0006] To achieve the above objectives, this invention provides the following technical solution: a method for generating Luneburg lens 3D printing materials based on diffusion models and reinforcement learning, demonstrating a complete closed loop from diffusion model generation to reinforcement learning optimization, including the following steps:
[0007] S1. Construction and initialization of the diffusion model;
[0008] S2, Training and integration of attribute predictors;
[0009] S3, Reinforcement Learning Framework Design;
[0010] S4. Materials generation and evaluation;
[0011] S5. Model optimization iteration;
[0012] S6. Verification and Application.
[0013] According to the above technical solution, S1 further includes: constructing a generator based on a diffusion model to generate a molecular structure representation of polymer materials. The model includes a forward diffusion process and a reverse denoising process. In order to handle the discrete properties of molecular structures and perform efficient diffusion processes in continuous space, a variational autoencoder is introduced to encode the molecular structure into a continuous latent space representation.
[0014] According to the above technical solution, step S2 further includes: the attribute predictor adopts a method based on... neural networks, The principle stems from its attention mechanisms, particularly self-attention and multi-head attention. The architecture includes an encoder and a decoder, each consisting of multiple stacked layers. Each encoder layer contains a multi-head self-attention sublayer and a feedforward network sublayer, supplemented by layer normalization and residual connections. The attention mechanism is calculated using the following formula:
[0015] ;
[0016] in These are query, key, and value matrices, respectively. For dimensions;
[0017] The model supports multimodal inputs: combined with In addition to 3D coordinates, multi-head attention uses multiple attention heads in parallel, obtaining different vector representations through linear transformations, enabling the model to simultaneously focus on multiple different types of dependencies in the molecular sequence. Position encoding adds positional information through sine and cosine functions, allowing the self-attention mechanism to perceive the sequence. The sequence information.
[0018] According to the above technical solution, S3 further includes: to achieve directional optimization of the generated polymer molecular structure, introducing a method based on the diffusion model. ,Right now Strengthen the learning framework It is a tradition ,Right now The algorithm extension abandons the value function estimator in the policy optimization process;
[0019] During reinforcement learning training, the diffusion model generates multiple candidate output trajectories for the same input noise and latent variables. The sampled trajectories are divided into several groups, and the relative advantage of each trajectory is calculated within each group:
[0020] ;
[0021] in, It is the first one generated corresponding to the same input. The reward value for each output result. It is all of this group The reward value of each output. and These are the mean and standard deviation of the group reward, respectively;
[0022] The strategy update still adopts the same approach as... akin However, it introduces the concept of relative group advantage:
[0023] ;
[0024] in, yes The importance ratio of levels yes Hyperparameters yes The weight of the penalty is used to constrain the new policy. Do not deviate from the initial strategy ;
[0025] To adapt to diffusion-based molecule generation tasks, the reinforcement learning problem is modeled as follows: the state is defined as the current generated molecule noise vector representation and the intermediate feature representation of the denoising network; the action is the sampling decision that controls the generation process; and the reward function is based on the constructed... The multidimensional output of the property predictor is used to calculate the key performance indicators required for Luneburg lens materials.
[0026] ;
[0027] in, The corresponding properties are density, mechanical strength, melting point, relative permittivity, and dielectric loss. and These are the predicted value and the target value, respectively. The corresponding normalization scale / tolerance range, These are the weights of each item. It is a penalty item.
[0028] According to the above technical solution, S4 further includes: using the initialized diffusion model to generate a batch of candidate molecular structures, with a batch size of 100-500. The generated molecules are first input into the attribute predictor for rapid evaluation, and the comprehensive reward value of each molecule is calculated. In order to avoid mode collapse, a diversity measure is introduced. Molecules with reward values higher than the threshold are used as positive samples, and low reward molecules are recorded as negative samples for subsequent reinforcement learning optimization. This step is parallelized, and the generation time is controlled in the minute range.
[0029] According to the above technical solution, S5 further includes: based on the accumulated reward value, using The algorithm iteratively optimizes the generation strategy parameters of the diffusion model to achieve directional material generation. During the iteration process, it works by... The strategy balances exploration and utilization. In the initial stage, new molecular spaces are explored, and the number of iterations is set to 50–200 rounds until the average reward of the generated materials exceeds a preset threshold.
[0030] According to the above technical solution, S6 further includes: performing multi-level verification on the optimized material, including computational simulation, laboratory synthesis, 3D printing testing and electromagnetic performance measurement. If the verification results do not meet the requirements, the experimental data is fed back to the model for fine-tuning, forming a closed-loop system for continuous learning. Through transfer learning, the method can be extended to the design of other 3D printing materials.
[0031] According to the above technical solution, in S1, the number of time steps in the specific construction process. Noise scheduling adopts linear Function, from arrive The input molecule embedding dimension is set to 256, using Auxiliary molecular diagrams are used to capture interatomic bonding relationships, through Optimizer Training Learning rate The batch size is 64. During training, the variational lower bound loss is monitored to ensure model convergence.
[0032] During model initialization, a pre-training dataset is used for preliminary training to ensure that the model has good generalization ability to polymer chemical space. In combination with the specific relative permittivity requirements of Luneburg lenses, the model incorporates a conditional generation mechanism.
[0033] According to the above technical solution, the key properties in S2 include density, mechanical strength, melting point, relative permittivity, and dielectric loss.
[0034] Among them, the density prediction value is The predicted mechanical strength value is Modulus The predicted melting point is the glass transition temperature. The predicted relative permittivity is 1-2.5, and the predicted dielectric loss is... .
[0035] According to the above technical solution, in step S4, the target range for the evaluated attribute is: density is Mechanical strength Melting point The relative permittivity is 1-2.5, and the dielectric loss is 0.001-0.05.
[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0037] 1. This invention overcomes the limitations of traditional methods, such as low efficiency, high cost, and lack of directionality, by constructing a closed-loop framework of generation-evaluation-optimization to efficiently generate polymer materials that meet the electromagnetic and mechanical performance requirements of Luneburg lenses. This method combines the generative capabilities of diffusion models with the optimization mechanism of reinforcement learning and introduces a system based on... The property predictor acts as a bridge to ensure that the generated molecular structure is not only stable, but also accurately matches the target properties.
[0038] 2. This invention ingeniously uses a diffusion model to generate the molecular structure of polymer materials and achieves dynamic adjustment through a reinforcement learning mechanism, wherein, based on The neural network acts as a property predictor to evaluate the generated material properties in real time and provide accurate reward signals accordingly, thereby enabling the directional generation of polymer materials that highly meet the requirements of Luneburg lenses. Compared with traditional random generation or empirical trial and error methods, this method not only significantly improves generation efficiency and material quality, but also greatly reduces experimental costs and resource consumption.
[0039] In summary, this invention seamlessly integrates generative models with optimization algorithms. First, candidate molecular structures are gradually generated from noise using a diffusion model. Then, a property predictor is used to predict the comprehensive reward value of key properties such as mechanical strength, melting point, density, dielectric constant, and dielectric loss. Finally, advanced reinforcement learning algorithms are used to update and refine the model parameters. This method is particularly suitable for the reverse design of 3D printed materials and has broad application prospects. It is not limited to electromagnetic wave lenses but can also be extended to the development of 3D printed functional materials. By integrating artificial intelligence technology, it provides an efficient and intelligent solution for materials science, promoting rapid iteration from computational simulation to actual manufacturing. Attached Figure Description
[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0041] In the attached diagram:
[0042] Figure 1 This is an overall flowchart of the method of the present invention;
[0043] Figure 2 This is a step diagram in an embodiment of the present invention;
[0044] Figure 3 This is a diagram illustrating the overall process of forward diffusion and reverse denoising in the diffusion model of this invention.
[0045] Figure 4 This is a diagram illustrating the training process of reverse denoising in the diffusion model of this invention;
[0046] Figure 5 This is a diagram illustrating the generation process of reverse denoising in the diffusion model of this invention;
[0047] Figure 6 This is the invention Algorithm framework diagram;
[0048] Figure 7 This is a schematic diagram of the material property prediction algorithm of the present invention. Detailed Implementation
[0049] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0050] Example: Figure 1 and Figure 2 As shown, this invention provides a technical solution for generating Luneburg lens 3D printing materials based on diffusion models and reinforcement learning, demonstrating a complete closed loop from diffusion model generation to reinforcement learning optimization, including the following steps:
[0051] S1. Construction and initialization of the diffusion model:
[0052] Construct a diffusion-based generator to generate polymer materials such as... Molecular structure representations such as strings, molecular diagrams, and three-dimensional coordinates, etc. Figure 3 , Figure 4 and Figure 5 As shown, the model includes a forward diffusion process that gradually adds Gaussian noise to the molecular embedding to achieve the transformation from order to disorder, and a reverse denoising process that gradually removes noise through a neural network to restore a stable molecular structure. To handle the discrete nature of the molecular structure and perform an efficient diffusion process in continuous space, a variational autoencoder is introduced. To visualize the molecular diagram Molecular structure is encoded into a continuous latent space representation, specifically, It consists of an encoder and a decoder: the encoder uses... Molecular diagrams are mapped to mean and variance parameters, and latent variables are sampled. This latent space captures the structural details and continuous distribution of molecules, while the decoder reconstructs the molecular structure from the latent variables by minimizing the reconstruction loss and Divergence regularization ensures the continuity and regularization of the latent space, facilitating subsequent diffusion models to operate on this latent space and avoiding the complexity of diffusion directly in discrete molecular space. This is particularly useful in polymer generation because it allows the capture of variability in long-chain structures and repeating units.
[0053] The principle of the diffusion model is mainly based on the denoising diffusion probability model. ,exist In this process, the forward process is a Markov chain that progressively adds Gaussian noise to the data until it approximates a pure noise distribution. Specifically, given a data sample... The forward process is defined as:
[0054] ;
[0055] in, These are noise scheduling parameters. From 1 to (Number of time steps, typically 1000); this process gradually corrupts the data structure, causing it to tend towards a standard Gaussian distribution. The reverse process involves training a neural network to gradually remove noise and restore the original data distribution.
[0056] ;
[0057] The training objective is to minimize the variational lower bound. This can be simplified to the mean square error loss of the prediction noise:
[0058] ;
[0059] in, For noise predicted by the neural network, Added Gaussian noise,
[0060] Starting with sampling from random noise and generating new target samples through reverse iteration, this mechanism makes diffusion models perform exceptionally well in generating high-dimensional data such as images or molecular structures because it captures fine-grained distribution details and avoids... The mode collapse problem in models, in molecular generation, where diffusion models are used to generate from noise. Strings and 3D conformations ensure the validity and diversity of the generated molecular chemistry;
[0061] In the specific construction process, the number of time steps Noise scheduling adopts linear Function, from arrive The input molecule embedding dimension is set to 256, using Auxiliary molecular diagrams are used to capture interatomic bonding relationships, through Optimizer Training Learning rate The batch size is 64. During training, the variational lower bound loss is monitored to ensure model convergence.
[0062] During model initialization, a pre-training dataset containing tens of thousands of polymer molecules and a custom polymer library is used for preliminary training to ensure that the model has good generalization ability to polymer chemical space. In combination with the specific relative permittivity requirement of Luneburg lenses, the model incorporates a conditional generation mechanism, specifically, using an additional target relative permittivity to guide the generation of specific materials.
[0063] S2. Training and ensemble of attribute predictors:
[0064] The attribute predictor uses a method based on neural networks, The principle stems from its attention mechanism. Especially self-attention and bullish attention , The architecture includes encoders and decoder Each layer consists of multiple stacked layers. Each encoder layer contains a multi-head self-attention sublayer and a feedforward network sublayer, supplemented by layer normalization and residual connections. The attention mechanism calculation formula is as follows:
[0065] ;
[0066] in These are query, key, and value matrices, respectively. For dimensions;
[0067] The model supports multimodal inputs: combined with and 3D coordinates (using) (Generation), multi-head attention is achieved through the parallel processing of multiple attention heads. Different vector representations obtained through linear transformation enable the model to simultaneously focus on multiple different types of dependencies in the molecular sequence, including position encoding. By adding positional information using sine and cosine functions, the self-attention mechanism can perceive the sequence. Sequence information, Its advantage lies in its strong parallel computing capability, compared to or It boasts faster processing speeds and has achieved breakthroughs in fields such as natural language processing and sequence prediction. In this invention, Property prediction for molecular sequences, by encoding molecules The dependence of these properties allows for accurate estimation of physicochemical properties, including mechanical strength, melting point, density, relative permittivity, and dielectric loss.
[0068] During implementation, the encoder layer number was 6, the attention head number was 8, the hidden dimension was 512, and the data was loaded from... And the weights pre-trained on the polymer database, The mean squared error loss function is used to ensure that the prediction error is less than 5%. In addition, during training, the following is introduced: The ratio is set to 0.1 to prevent overfitting, and layer normalization is used to improve training stability;
[0069] The predictor receives the molecular structure generated by the diffusion model as input, and then... After encoding by morphology and position, a multi-layer attention mechanism is used to capture long-term dependencies in the molecular sequence, ultimately outputting predicted values for key attributes, specifically density. Mechanical strength Modulus Melting point and glass transition temperature The relative permittivity range is 1-2.5 and the dielectric loss range is... to In between, the model underwent supervised training by pre-collecting a large amount of molecular structure data and experimentally verified property data of polymer materials, utilizing... Leveraging the advantages of parallel computing, efficient predictions are achieved, providing reliable reward signals for reinforcement learning;
[0070] S3. Reinforcement Learning Framework Design:
[0071] like Figure 6 As shown, to achieve targeted optimization of the generated polymer molecular structure, a method is introduced based on the diffusion model. ,Right now Strengthen the learning framework It is a tradition ,Right now The algorithm extension abandons the value function estimator during policy optimization. Instead, a group relative comparison mechanism is adopted to reduce the variance of the advantage estimation and improve the stability and efficiency of the strategy update, which is suitable for high-dimensional action space tasks of molecular generation.
[0072] During reinforcement learning training, the diffusion model generates multiple candidate output trajectories for the same input noise and latent variables. The sampled trajectories were divided into several groups. Calculate the relative advantage of each trajectory within each group:
[0073] ;
[0074] in, It is the first one generated corresponding to the same input. The reward value for each output result. It is all of this group The reward value of each output. and These are the mean and standard deviation of the group reward, respectively. It is a very small constant to prevent division by zero;
[0075] The strategy update still adopts the same approach as... akin However, it introduces the concept of relative group advantage:
[0076] ;
[0077] in, yes The importance ratio of levels yes Hyperparameters yes The weight of the penalty is used to constrain the new policy. Do not deviate from the initial strategy ;
[0078] To adapt to diffusion-based molecular generation tasks, the reinforcement learning problem is modeled as follows: the state is defined as the noise vector representation of the currently generated molecule and the intermediate feature representation of the denoising network; the action is the sampling decision that controls the generation process, selecting the next molecular fragment to be generated and adjusting the denoising strategy parameters; and the reward function is based on the constructed... The multidimensional output of the property predictor is used to calculate the key performance indicators required for Luneburg lens materials.
[0079] ;
[0080] in, The corresponding properties are density, mechanical strength, melting point, relative permittivity, and dielectric loss. and These are the predicted value and the target value, respectively. The corresponding normalization scale / tolerance range, These are the weights of each item. These are penalty factors, including instability, difficulty of synthesis, and toxicity;
[0081] It is particularly suitable for high-dimensional action spaces, such as parameter optimization tasks in molecular generation, because it significantly reduces training instability and improves sample efficiency in large language models. In the task of material optimization, It has been proven to enhance targeted generation capabilities compared to traditional methods. It has faster convergence speed and stronger generalization performance.
[0082] S4. Material Generation and Evaluation:
[0083] like Figure 7 As shown, a batch of candidate molecular structures is generated using the initialized diffusion model, with a batch size of 100-500. The generated molecules are first input into the attribute predictor for rapid evaluation. The comprehensive reward value of each molecule is calculated, including density, mechanical strength, melting point, relative permittivity, and dielectric loss index. To avoid mode collapse, a diversity measure is introduced, specifically based on molecular fingerprint similarity screening. Molecules with reward values higher than the threshold are used as positive samples, and molecules with low reward values are recorded as negative samples for subsequent reinforcement learning optimization. This step can be parallelized, with the generation time controlled in the minute range, generating 100 candidate polymers such as variant polystyrene and acrylate copolymers.
[0084] The target range for the evaluation attribute is: density. Mechanical strength Melting point The relative permittivity is 1-2.5, and the dielectric loss is 0.001-0.05.
[0085] The evaluation results showed that the initial batch had an average reward of 0.6, with high-reward molecule accounting for 30%. Further analysis of low-reward molecules identified common problems, specifically... Molecular instability was verified and stored as negative samples, generating 200 candidates. Evaluation showed that 3D input reduced the proportion of invalid molecules by 15%. The entire evaluation process used... Accelerates the process, reducing time to no more than 5 minutes, and supports large-scale parallel generation;
[0086] S5. Model optimization iteration:
[0087] Based on the cumulative reward value, use The algorithm iteratively optimizes the generation strategy parameters of the diffusion model to achieve directional material generation. During the iteration process, it works by... The strategy balances exploration and utilization. Initially, new molecular spaces are explored, with 50 iterations per round, until the average reward for generated materials exceeds a preset threshold. During each iteration, the following can be introduced: The divergence regularization constraint strategy is updated to prevent generated molecules from deviating from the initial distribution and to ensure molecular chemical feasibility.
[0088] S6. Verification and Application:
[0089] The optimized materials were validated at multiple levels, including molecular dynamics simulations to verify stability, small-scale synthesis testing to verify feasibility, and application. The printer manufactures prototypes and uses a vector network analyzer and resonant cavity method to measure the relative permittivity and dielectric loss. If the verification results do not meet the requirements, the experimental data are fed back to the model for fine-tuning, forming a closed-loop system for continuous learning. This method also supports transfer learning and can be extended to the design of other 3D printing materials.
[0090] The specific implementation process is achieved through molecular dynamics simulations. Software, time simulation To verify stability, use Software calculates electronic structure and bandgap; 3D printing testing is conducted. The technology involves printing prototypes with a resolution of 0.5mm. Electromagnetic measurements are performed using a vector network analyzer and a resonant cavity method to accurately measure the relative permittivity and dielectric loss. Throughout the feedback loop, if the attribute deviation exceeds 10%, 10% of the training data is added for iterative updates until the accuracy is improved to 98%.
[0091] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning, characterized in that: This demonstrates the complete closed loop from diffusion model generation to reinforcement learning optimization, including the following steps: S1. Construction and initialization of the diffusion model; S2, Training and ensemble of attribute predictors; S3, Reinforcement Learning Framework Design; S4. Materials generation and evaluation; S5. Model optimization iteration; S6. Verification and Application.
2. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 1, characterized in that: The S1 further includes: constructing a generator based on a diffusion model to generate a molecular structure representation of polymer materials. The model includes a forward diffusion process and a reverse denoising process. In order to handle the discrete properties of molecular structures and perform efficient diffusion processes in continuous space, a variational autoencoder is introduced to encode the molecular structure into a continuous latent space representation.
3. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 1, characterized in that: S2 further includes: the attribute predictor adopts a based neural networks, The principle stems from its attention mechanisms, particularly self-attention and multi-head attention. The architecture includes an encoder and a decoder, each consisting of multiple stacked layers. Each encoder layer contains a multi-head self-attention sublayer and a feedforward network sublayer, supplemented by layer normalization and residual connections. The attention mechanism is calculated using the following formula: ； in These are query, key, and value matrices, respectively. For dimensions; The model supports multimodal inputs: combined with With 3D coordinates, multi-head attention uses multiple attention heads in parallel and different vector representations obtained through linear transformation, enabling the model to simultaneously focus on multiple different types of dependencies in the molecular sequence. Position encoding adds positional information through sine and cosine functions, allowing the self-attention mechanism to perceive the order information of tokens in the sequence.
4. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 1, characterized in that: S3 further includes: to achieve directional optimization of the generated polymer molecular structure, introducing a method based on the diffusion model. ,Right now Strengthen the learning framework It is a tradition ,Right now The algorithm extension abandons the value function estimator in the policy optimization process; During reinforcement learning training, the diffusion model generates multiple candidate output trajectories for the same input noise and latent variables. The sampled trajectories are divided into several groups, and the relative advantage of each trajectory is calculated within each group: ； in, It is the first one generated corresponding to the same input. The reward value for each output result. It is all of this group The reward value of each output. and These are the mean and standard deviation of the group reward, respectively; The strategy update still adopts the same approach as... akin However, it introduces the concept of relative group advantage: ； in, yes The importance ratio of the levels yes Hyperparameters yes The weight of the penalty is used to constrain the new policy. Do not deviate from the initial strategy ; To adapt to diffusion-based molecule generation tasks, the reinforcement learning problem is modeled as follows: the state is defined as the current generated molecule noise vector representation and the intermediate feature representation of the denoising network; the action is the sampling decision that controls the generation process; and the reward function is based on the constructed... The multidimensional output of the property predictor is used to calculate the key performance indicators required for Luneburg lens materials. in, The corresponding properties are density, mechanical strength, melting point, relative permittivity, and dielectric loss. and These are the predicted value and the target value, respectively. The corresponding normalization scale / tolerance range These are the weights of each item. It is a penalty item.
5. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 1, characterized in that: S4 further includes: generating a batch of candidate molecular structures using the initialized diffusion model, with a batch size of 100-500. The generated molecules are first input into the attribute predictor for rapid evaluation, and the comprehensive reward value of each molecule is calculated. In order to avoid mode collapse, a diversity measure is introduced. Molecules with reward values higher than the threshold are recorded as positive samples, and low reward molecules are recorded as negative samples for subsequent reinforcement learning optimization. This step is parallelized, and the generation time is controlled in the minute range.
6. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 1, characterized in that: S5 further includes: based on the cumulative reward value, using... The algorithm iteratively optimizes the generation strategy parameters of the diffusion model to achieve directional material generation. During the iteration process, it works by... The strategy balances exploration and utilization. In the initial stage, new molecular spaces are explored, and the number of iterations is set to 50–200 rounds until the average reward of generated materials exceeds a preset threshold.
7. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 1, characterized in that: The S6 further includes: performing multi-level verification on the optimized preferred material, including computational simulation, laboratory synthesis, 3D printing testing and electromagnetic performance measurement. If the verification results do not meet the requirements, the experimental data is fed back to the model for fine-tuning, forming a closed-loop system for continuous learning. Through transfer learning, the method can be extended to the design of other 3D printing materials.
8. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 2, characterized in that: In S1, during the specific construction process, the number of time steps Noise scheduling adopts linear Function, from arrive The input molecule embedding dimension is set to 256, using Auxiliary molecular diagrams are used to capture interatomic bonding relationships, through Optimizer training Learning rate The batch size is 64. During training, the variational lower bound loss is monitored to ensure model convergence. During model initialization, a pre-training dataset is used for preliminary training to ensure that the model has good generalization ability to polymer chemical space. In combination with the specific relative permittivity requirements of Luneburg lenses, the model incorporates a conditional generation mechanism.
9. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 3, characterized in that: In S2, key properties include density, mechanical strength, melting point, relative permittivity, and dielectric loss. Among them, the density prediction value is The predicted mechanical strength value is Modulus The predicted melting point is the glass transition temperature. The predicted relative permittivity is 1-2.5, and the predicted dielectric loss is... .
10. The method for generating Luneburg lens 3D printing materials based on diffusion model and reinforcement learning according to claim 4, characterized in that: In S4, the target range for the evaluated attribute is: density is Mechanical strength Melting point The relative permittivity is 1-2.5, and the dielectric loss is 0.001-0.05.