Search

CN-122024952-A - Diffusion model and reinforcement learning-based Longber lens 3D printing material generation method

CN122024952ACN 122024952 ACN122024952 ACN 122024952ACN-122024952-A

Abstract

The invention discloses a method for generating a 3D printing material of a Longber lens based on a diffusion model and reinforcement learning, which relates to the technical field of printing and manufacturing of Longber lenses and comprises the following steps of: s1, constructing and initializing a diffusion model, S2, training and integrating an attribute predictor, S3, designing a reinforcement learning framework, S4, generating and evaluating materials, S5, iterating model optimization, S6, verifying and applying, seamlessly combining a generating model and an optimizing algorithm, firstly gradually generating candidate molecular structures from noise through the diffusion model, and then, predicting comprehensive rewarding values of the key attributes of mechanical strength, melting point, density, dielectric constant and dielectric loss by using an attribute predictor, and updating and refining model parameters by using an advanced reinforcement learning algorithm.

Inventors

  • WANG JIAKUI
  • FAN DING
  • LIU YALAN
  • YANG YUE

Assignees

  • 湖北查克科技有限责任公司

Dates

Publication Date
20260512
Application Date
20260119

Claims (10)

  1. 1. The method for generating the 3D printing material of the Longber lens based on the diffusion model and reinforcement learning is characterized by showing a complete closed loop from the generation of the diffusion model to the optimization of reinforcement learning, and comprises the following steps of: S1, constructing and initializing a diffusion model; s2, training and integrating attribute predictors; s3, strengthening the learning frame design; s4, generating and evaluating materials; s5, model optimization iteration; S6, verification and application.
  2. 2. The method for generating a Longber lens 3D printing material based on diffusion modeling and reinforcement learning of claim 1 wherein S1 further comprises constructing a generator based on diffusion modeling for generating a molecular structure representation of the polymer material, the modeling comprising a forward diffusion process and a backward denoising process, introducing a variational self-encoder for encoding the molecular structure into a continuous latent space representation in order to process discrete properties of the molecular structure and perform an efficient diffusion process in continuous space.
  3. 3. The method for generating a Longber lens 3D printing material based on diffusion model and reinforcement learning as set forth in claim 1, wherein S2 further comprises the attribute predictor employing a diffusion-based method Is used for the neural network of (a), The principle of (c) derives from its attention mechanisms, in particular self-attention and multi-headed attention, The architecture includes encoders and decoders, each stacked by multiple layers, each layer encoder containing a multi-headed self-attention sub-layer and a feed forward network sub-layer, aided by layer normalization and residual connection, the attention mechanism calculation formula is: ; Wherein the method comprises the steps of Respectively a query, a key and a value matrix, Is a dimension; model support multi-modal input-combining And 3D coordinates, the multi-head attention is represented by different vectors obtained by parallel multiple attention heads and linear transformation, so that the model can pay attention to various dependency relations in a molecular sequence at the same time, position codes add position information through sine and cosine functions, and a self-attention mechanism can sense the sequence information of the token in the sequence.
  4. 4. The method for generating a Robert lens 3D printing material based on a diffusion model and reinforcement learning as set forth in claim 1, wherein S3 further comprises introducing on the basis of the diffusion model for achieving directional optimization of the molecular structure of the generated polymer I.e. The frame of the learning is strengthened, Is conventional in that I.e. The expansion of the algorithm abandons a cost function estimator in the strategy optimization process; in the reinforcement learning training process, the diffusion model generates a plurality of candidate output tracks aiming at the same input noise and latent variable, Dividing the sampled trajectories into a plurality of groups, and calculating the relative advantage of each trajectory in each group: ; Wherein, the Is generated corresponding to the same input The prize value of the outcome is output, Is all of the group The value of the prize to be output, And Respectively the average value and standard deviation of group rewards; Policy updates are still employed with Similarly, the But introduces population relative advantages: ; Wherein, the Is that The importance ratio of the level(s), Is that The super-parameters of the parameters are used for processing the data, Is that Penalty weights for constraining new policies Without deviating from the initial strategy ; To adapt to the diffuse molecular generation task, the reinforcement learning problem is modeled by defining states as a currently generated molecular noise vector representation, an intermediate feature representation of the denoising network, actions as sampling decisions for controlling the generation process, and reward functions based on the construction The multidimensional output of the attribute predictor calculates key performance indexes required by the comprehensive primary lens material: Wherein, the Corresponding to 5 items of density, mechanical strength, melting point, relative dielectric constant and dielectric loss, And Respectively a predicted value and a target value, The corresponding normalized scale/tolerance range, Is the weight of each item, Is a penalty term.
  5. 5. The method for generating a 3D printing material based on a diffusion model and reinforcement learning according to claim 1, wherein S4 further comprises generating a batch of candidate molecular structures by using the initialized diffusion model, wherein the batch size is 100-500, the generated molecules are firstly input into an attribute predictor for quick evaluation, a comprehensive rewarding value of each molecule is calculated, in order to avoid mode collapse, molecules with rewarding values higher than a threshold value are introduced as positive samples, low rewarding molecules are recorded as negative samples for subsequent reinforcement learning optimization, the step of parallelization processing is performed, and the generation time is controlled at a minute level.
  6. 6. The method for generating the 3D printing material of the luneberg lens based on the diffusion model and the reinforcement learning according to claim 1, wherein the method comprises the following steps: the S5 further comprises using, based on the jackpot value The algorithm iterates and optimizes the generation strategy parameters of the diffusion model to realize the generation of the directional material, and in the iteration process, the generation strategy parameters are obtained by And (3) strategy balance exploration and utilization, wherein new molecular space is explored in the initial stage, and the iteration times are set to 50-200 rounds until the average rewards of the generated materials exceed a preset threshold value.
  7. 7. The method for generating the 3D printing material of the Robert lens based on the diffusion model and reinforcement learning according to claim 1, wherein S6 is further characterized by further comprising the steps of carrying out multi-level verification on the optimized generated preferable material, including calculation simulation, laboratory synthesis, 3D printing test and electromagnetic performance measurement, if the verification result does not meet the requirement, feeding experimental data back to the model for fine adjustment, forming a continuous learning closed-loop system, and expanding the method to other 3D printing material designs through transfer learning.
  8. 8. The method for generating a 3D printing material for a Robert lens based on a diffusion model and reinforcement learning as set forth in claim 2, wherein in said S1, the number of time steps is determined during the specific construction process Noise scheduling employing linearity Function, slave of To the point of The input molecule intercalation dimension was set to 256, using Auxiliary molecular graph representation to capture bonding relationship among atoms by Optimizer training Learning rate Batch size 64, monitoring variable lower bound loss in the training process, and ensuring model convergence; when the model is initialized, a pre-training data set is used for preliminary training to ensure that the model has good generalization capability on polymer chemical space, and the model is integrated into a condition generating mechanism by combining the specific relative dielectric constant requirement of the luneberg lens.
  9. 9. The method for generating a Tobert lens 3D printing material based on diffusion model and reinforcement learning according to claim 3, wherein in S2, key properties include density, mechanical strength, melting point, relative dielectric constant and dielectric loss; wherein the density prediction value is Predicted value of mechanical strength is Modulus of The predicted melting point is the glass transition temperature A relative dielectric constant of 1-2.5 and a dielectric loss of 1 。
  10. 10. The method for generating a Longber lens 3D printing material based on diffusion model and reinforcement learning as claimed in claim 4, wherein in S4, the evaluation attribute target range is density of Mechanical strength Melting point of The relative dielectric constant is 1-2.5, and the dielectric loss is 0.001-0.05.

Description

Diffusion model and reinforcement learning-based Longber lens 3D printing material generation method Technical Field The invention relates to the technical field of printing and manufacturing of a Robert lens, in particular to a method for generating a 3D printing material of the Robert lens based on a diffusion model and reinforcement learning. Background Dragon's primary lensIs an electromagnetic lensThe refractive index distribution of the radial continuous change of the lens material enables the lens material to realize near ideal aberration-free focusing effect on incident electromagnetic waves, so the lens material has wide application in radar systems, wireless communication antennas and imaging equipment, and along with the development of additive manufacturing (3D printing) technology, the lens material not only needs to have a specific dielectric constant range, but also needs to meet the requirements of mechanical strength, density and glass transition temperatureThe traditional material development mode generally depends on manual formula design, experimental synthesis and high-throughput screening, the process is long in time consumption and high in cost, the uncertainty of the material performance is large, the development period is difficult to shorten, in recent years, a generation model and reinforcement learning are used in the generation of printing materials, the generation model, particularly a diffusion model, shows strong capability in the field of molecular and material generation, and the reinforcement learning is used as an optimization framework capable of forming a feedback closed loop by utilizing states, actions and rewards; At present, most of the generated models pay attention to structural diversity and effective sampling capability, an optimization mechanism tightly coupled with target attributes is not established, and the high dependence on reward signals limits the application of the generated models in large-scale generation tasks, so that the iteration speed of traditional material design is slow; Therefore, a generation-evaluation-optimization closed-loop framework capable of deeply integrating a diffusion model, reinforcement learning and a high-precision attribute predictor is needed to realize efficient and directional generation of electromagnetic lens materials, and a method for generating a Longber lens 3D printing material based on the diffusion model and reinforcement learning is provided for the technical pain point. Disclosure of Invention The invention provides a method for generating a 3D printing material of a Robert lens based on a diffusion model and reinforcement learning, which can effectively solve the problems that in the background art, most of the generated model focuses on structural diversity and effective sampling capacity, an optimization mechanism tightly coupled with target attributes is not established, the high dependence on a reward signal limits the application of the method in large-scale generation tasks, and the iteration speed of traditional material design is slow. In order to achieve the purpose, the invention provides the following technical scheme that the method for generating the 3D printing material of the Robert lens based on a diffusion model and reinforcement learning shows a complete closed loop from the generation of the diffusion model to the optimization of reinforcement learning, and comprises the following steps: S1, constructing and initializing a diffusion model; s2, training and integrating attribute predictors; s3, strengthening the learning frame design; s4, generating and evaluating materials; s5, model optimization iteration; S6, verification and application. According to the technical solution, the S1 further comprises a generator based on a diffusion model, which is used for generating a molecular structure representation of the polymer material, wherein the model comprises a forward diffusion process and a reverse denoising process, and a variation self-encoder is introduced for encoding the molecular structure into a continuous latent space representation in order to process the discrete nature of the molecular structure and perform efficient diffusion process in the continuous space. According to the technical scheme, the S2 further comprises that the attribute predictor adopts the method based onIs used for the neural network of (a),The principle of (c) derives from its attention mechanisms, in particular self-attention and multi-headed attention,The architecture includes encoders and decoders, each stacked by multiple layers, each layer encoder containing a multi-headed self-attention sub-layer and a feed forward network sub-layer, aided by layer normalization and residual connection, the attention mechanism calculation formula is: ; Wherein the method comprises the steps of Respectively a query, a key and a value matrix,Is a dimension; model support multi-modal input-combining And 3D coordinates, the mu