Thermal radiation material design method and system based on neural network and Bayesian optimization
Through the combination of deep learning neural network and Bayesian optimization, high-dimensional space is mapped into low-dimensional space and efficient search, solving the time-consuming and local optimal problems in the design of thermal radiation material, and achieving efficient thermal radiation material Design on demand shortens R&D cycle and costs.
Patent Information
- Application Number
- CN202210300613.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-25
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-03-25
AI Technical Summary
In the prior art, thermal radiation material design has problems such as iterative calculation, easy to fall into local optimality and difficult to use in high-degree of freedom structural systems. Neural networks have problems such as time-consuming data set preparation and difficulty converging loss functions, resulting in low Effective thermal radiation material design methods increase R&D costs and cycles.
Using a design framework that combines deep learning neural networks and Bayesian optimization, iterates through training deep learning neural network models and combining Bayesian optimization to loop iteration, map high-dimensional candidate spaces into low-dimensional continuous spaces, and then conducts efficient searches to achieve On-demand design of thermal radiation materials.
It effectively solves the problem that high-dimensional thermal radiation materials are difficult to achieve globally optimized design, significantly saves search costs, can quickly find the material structure of the target thermal radiation characteristics, and shortens the material research and development costs and cycles.
Smart Images

Figure CN114664395B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of nanophotonic structure design, and in particular, to a method and system for designing thermal radiation materials based on neural networks and Bayesian optimization. In particular, it preferably relates to an on-demand design method for thermal radiation materials based on deep learning neural networks and Bayesian optimization. Background Art
[0002] Artificially designed photonic structures (including photonic crystals, metamaterials, and plasma nanostructures) have shown great advantages in customizing the interaction between light and matter, and can be widely used in thermal radiation fields such as radiative cooling, thermophotovoltaics, and thermal camouflage. In order to achieve a variety of novel and complex optical properties, efficient search in high-dimensional candidate spaces to achieve on-demand material design is becoming increasingly important, which has brought great difficulties to the further promotion of thermal radiation materials.
[0003] In recent years, various gradient-based or metaheuristic-based algorithms have been proposed for complex "non-intuitive" metamaterial designs. However, these algorithms often fall into local optimality due to different initial population selections, rely on time-consuming electromagnetic simulation iterative calculations, and are difficult to apply to structural systems with high degrees of freedom. Deep learning neural networks, as a data-driven method, can quickly design material structures with target thermal radiation properties by learning and training the complex interaction relationship between light and matter. However, the existence of time-consuming data set preparation and difficulty in convergence of loss functions also limits its design efficiency.
[0004] In the prior art, Atsushi Sakurai et al. disclosed a thermal radiation material design method based on Bayesian optimization on pages 319-326 of the journal "ACS central science" in January 2019. Although this method achieves efficient global optimization search, it cannot be applied to thermal radiation material systems with high-dimensional candidate spaces because Bayesian optimization relies heavily on computing memory. Sunae So et al. proposed a nanophotonic structure design based on conditional deep convolutional generative adversarial networks on pages 1255-1261 of the journal "Nanophotonics" in June 2019, and Zhaxylyk A. Kudyshev et al. disclosed a thermal radiation material design method based on adversarial autoencoders in the journal "Applied Physics Reviews" in May 2020. Although the deep learning model achieves the learning of material design characteristics under specific thermal radiation targets, there is no further efficient search method to assist in determining the optimal material design.
[0005] The Chinese invention patent document with publication number CN104277241A discloses a thermal radiation material, which contains the following fillers: titanium dioxide, 10-45% by weight; zirconium dioxide, 5-25% by weight; magnesium oxide, 2-30% by weight; rare earth metal oxide, 0.01-0.5% by weight. The thermal conductivity of the thermal radiation material is 0.34-1.5 W / m·K, and the infrared emissivity of the thermal radiation material at 40°C and in the infrared spectrum region of 4-14 μm is greater than or equal to 88%.
[0006] Regarding the above-mentioned related technologies, the inventor believes that traditional optimization algorithms have problems such as time-consuming iterative calculations, easy to fall into local optimality, and difficult to use in structural systems with high degrees of freedom. Neural networks have problems such as time-consuming data set preparation and difficulty in convergence of loss functions. As for some of the latest proposed material design methods, they also have limitations. In other words, inefficient thermal radiation material design methods greatly increase R&D costs and cycles. Summary of the invention
[0007] In view of the defects in the prior art, the object of the present invention is to provide a thermal radiation material design method and system based on neural network and Bayesian optimization.
[0008] According to the present invention, a method for designing thermal radiation materials based on neural network and Bayesian optimization comprises the following steps:
[0009] Model acquisition step: obtain and train the deep learning neural network model;
[0010] Iteration step: Combine the trained deep learning neural network model with Bayesian optimization to iterate the thermal radiation material structure design to obtain the thermal radiation material structure design.
[0011] Preferably, the model acquisition step includes the following steps:
[0012] Step S1: Prepare data sets based on material systems and design goals;
[0013] Step S2: Use the dataset to train the adversarial autoencoder and extract the decoder in the adversarial autoencoder as a generative model.
[0014] Preferably, step S1 comprises the following steps:
[0015] Characteristic acquisition step: the material structure is represented by a digital sequence and used as the input of electromagnetic simulation to obtain the thermal radiation characteristics of the material;
[0016] Standard determination steps: Determine the evaluation criteria based on the design goals;
[0017] Dataset acquisition step: Use traditional optimization algorithms or random search to iterate the feature acquisition step and the standard determination step, and obtain the data set after screening.
[0018] Preferably, step S2 comprises the following steps:
[0019] Network determination step: expand the data set and determine the adversarial autoencoder network architecture;
[0020] Training model steps: Use the data set after data expansion to train the adversarial autoencoder and save the model inside the adversarial autoencoder in blocks.
[0021] Preferably, in the iterative step, the output of the generated model is connected to the input of the electromagnetic simulation, the output of the electromagnetic simulation is connected to the evaluation standard, the calculation result under the evaluation standard is connected to the input of the Bayesian optimization, the output of the Bayesian optimization is connected to the input of the generated model, and the cycle is iterated until the design meets the target.
[0022] According to the present invention, a thermal radiation material design system based on neural network and Bayesian optimization includes the following modules:
[0023] Model acquisition module: acquire and train deep learning neural network models;
[0024] Iteration module: The trained deep learning neural network model is combined with Bayesian optimization to iterate the thermal radiation material structure design to obtain the thermal radiation material structure design.
[0025] Preferably, the model acquisition module includes the following modules:
[0026] Module M1: Prepare data sets for material systems and design goals;
[0027] Module M2: Use the dataset to train the adversarial autoencoder and extract the decoder in the adversarial autoencoder as the generative model.
[0028] Preferably, the module M1 includes the following modules:
[0029] Characteristic acquisition module: The material structure is represented by a digital sequence and used as the input of electromagnetic simulation to obtain the thermal radiation characteristics of the material;
[0030] Standard determination module: Determine the evaluation criteria based on the design objectives;
[0031] Dataset acquisition module: Use traditional optimization algorithms or random search to iterate the feature acquisition module and the standard determination module to obtain the data set after screening.
[0032] Preferably, the module M2 includes the following modules:
[0033] Network determination module: expand the data set and determine the adversarial autoencoder network architecture;
[0034] Training model module: Use the data set after data expansion to train the adversarial autoencoder and save the model inside the adversarial autoencoder in blocks.
[0035] Preferably, in the iterative module, the output of the generated model is connected to the input of the electromagnetic simulation, the output of the electromagnetic simulation is connected to the evaluation standard, the calculation result under the evaluation standard is connected to the input of the Bayesian optimization, the output of the Bayesian optimization is connected to the input of the generated model, and the loop is iterated until the design meets the target.
[0036] Compared with the prior art, the present invention has the following beneficial effects:
[0037] 1. The present invention adopts a design framework that combines a deep learning neural network model and Bayesian optimization to map a high-dimensional candidate space into a low-dimensional, continuous and compact space, and then uses Bayesian optimization to perform efficient search, thereby solving the problem that it is difficult to achieve global optimization design of high-dimensional thermal radiation materials;
[0038] 2. The on-demand design method for thermal radiation materials based on deep learning neural networks and Bayesian optimization proposed in the present invention can achieve material structure design under target thermal radiation characteristics by calculating the number of structures far less than 1% of the total candidate space. At the same time, it solves the problems that the Bayesian optimization algorithm is difficult to use for global search in high-dimensional problems and the time-consuming preparation of neural network training sets, which greatly saves the search cost. The results of using deep learning neural networks to effectively compress high-dimensional candidate space into low-dimensional space and then combining Bayesian optimization for efficient search prove the effectiveness, feasibility and accuracy of the design framework. The proposed method can also be extended to other high-dimensional complex thermal radiation material design problems;
[0039] 3. The present invention can be used to reversely design complex thermal radiation materials on demand based on spectral response, which can greatly shorten the research and development cost and cycle of the materials. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:
[0041] Figure 1 A diagram of the training process for a deep learning neural network;
[0042] Figure 2 A schematic diagram of the process of combining Bayesian optimization with generative models in deep learning neural networks;
[0043] Figure 3 This is a schematic diagram of the network architecture of a deep learning neural network;
[0044] Figure 4 This is the spectral response diagram corresponding to the optimal structure design at three target wavelengths of 4.5, 5.5 and 6.5 microns. DETAILED DESCRIPTION
[0045] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
[0046] The embodiment of the present invention discloses a method for designing thermal radiation materials on demand based on deep learning neural network and Bayesian optimization, such as Figure 1 and Figure 2 As shown, the following steps are included: Model acquisition step: obtain and train a deep learning neural network model.
[0047] The model acquisition step includes the following steps: Step S1: Prepare a data set for the material system and design target, and the data set includes a structure-property data set or a structure data set. Specifically, for a specific material system and design target, a small amount of calculation is used to prepare a structure-property data set or a structure data set. The structure-property data set represents a data set composed of structure and property, and the structure and property correspond one to one. The structure represents the structural design of the material, and the property represents the thermal radiation property obtained by electromagnetic simulation calculation based on the structural design of the material.
[0048] Step S1 includes the following steps: Characteristic acquisition step: the structure of the material is represented by a digital sequence, and used as an input of electromagnetic simulation to obtain the thermal radiation characteristics of the material. That is, the structure of the material is represented by a digital sequence in the form of "0" or "1", and used as an input of electromagnetic simulation (including strict coupling analysis, time-domain finite difference, etc.) to obtain accurate thermal radiation characteristics of the material.
[0049] Standard determination step: Determine the evaluation standard according to the design goal. That is, map the calculated high-dimensional thermal radiation characteristic data to a single value to evaluate the matching of the current thermal radiation characteristic with the target thermal radiation characteristic. The maximum value is 1, indicating a complete match with the target thermal radiation characteristic.
[0050] Dataset acquisition step: Use traditional optimization algorithms or random search to iterate the characteristic acquisition step and standard determination step, and obtain the data set after screening. That is, use traditional optimization algorithms or random search to implement the iterative cycle of the above steps to obtain a small number of better structure-characteristic data sets or structure data sets. The structure that needs to be calculated in the data set acquisition process is determined by using traditional optimization algorithms or random search, and the structural design that best meets the target thermal radiation characteristic requirements is screened out based on electromagnetic simulation and evaluation index calculations.
[0051] Specifically, a wavelength-selective narrowband thermal emitter between 4 and 7 microns (the target wavelengths are set at 4.5, 5.5, and 6.5 microns, respectively) consists of 36 layers of film structures, each of which has two candidate materials: germanium (Ge) and silicon dioxide (SiO2). The candidate materials for each layer are selected from commonly used high-refractive-index semiconductors and low-refractive-index dielectric materials, and the substrate material uses opaque metal tungsten (W). The thickness of each layer is fixed at 0.11 microns. In order to make the input of the deep learning neural network more intuitive and universal, we use binary values to represent the materials selected for each layer, such as "0" and "1" to represent the selection of germanium and silicon dioxide, respectively, that is, a 36-dimensional binary vector is used to represent all possible multilayer structures.
[0052] During the data acquisition process, the electromagnetic simulation method (rigorous coupled wave analysis RCWA, Rigorous Coupled Wave Analysis) is used to accurately calculate the spectral response corresponding to the material structure and obtain the evaluation index. The specific definition of the evaluation index FOM is as follows:
[0053]
[0054] Where ε represents the emissivity of each structure; E b Represents the spectral radiation density at a given wavelength λ and temperature T; λ min and λ max They represent the minimum operating wavelength of 4 microns and the maximum operating wavelength of 7 microns respectively; λ1 and λ2 represent the lower and upper boundaries of the narrowband, i.e., the target wavelength minus / plus 4 nanometers; the last item P represents the penalty item added when the spectral curve is not smooth, which is equal to 0.1×the number of peaks outside the target band. According to the current indicators, the spectral response of an ideal thermal emitter is: there is a sharp peak at the target wavelength and the bandwidth does not exceed 4 nanometers, and the spectral response at the remaining wavelengths is expected to be 0.
[0055] The feature acquisition and evaluation criteria calculation were combined with traditional optimization algorithms (such as evolutionary algorithms, Bayesian optimization, and random search) to search for optimal structures in small batches, and the results were used as data sets for subsequent model training. In the end, for three types of targets with wavelengths of 4.5, 5.5, and 6.5 microns, 118, 157, and 136 sets of data were prepared for the training of deep learning neural networks, and the evaluation indicators ranged from 0.28 to 0.81.
[0056] Step S2: Use the dataset to train the adversarial autoencoder, and extract the decoder in the adversarial autoencoder as the generative model. That is, use the structure-feature dataset to simultaneously train the three-part model in the adversarial autoencoder and save them separately, and extract the decoder in the adversarial autoencoder after training as the generative model.
[0057] Step S2 includes the following steps: Data expansion and network determination: The data set is expanded to improve the robustness of model training, and the adversarial autoencoder network architecture suitable for the current problem is determined. That is, the data set is expanded using data expansion methods such as jitter and scaling, and the applicable adversarial autoencoder network architecture is determined through multiple rounds of testing.
[0058] Model training steps: Use the data set after data augmentation to train the finalized adversarial autoencoder, and save the model inside the adversarial autoencoder in blocks. That is, use the data set after data augmentation to train the adversarial autoencoder until the model converges, and then save the encoder, decoder, and discriminator inside the adversarial autoencoder in blocks. Among them, the training of the adversarial autoencoder uses the structure design in the screened structure-feature dataset or the structure dataset.
[0059] Specifically, before training the deep learning neural network, the data is expanded to more than 20,000 groups using the data augmentation method of jittering and scaling in order to comprehensively consider the data specificity and save the preparation time of the data set. Figure 3 This is a deep learning neural network architecture diagram after multiple rounds of adjustments for the current problem, including encoder, decoder and discriminator. The encoder is used to map the 36-dimensional structural parameters to the 2-dimensional space, and the two hidden layers contain 512 and 512 neurons respectively. The decoder is used to restore the 2-dimensional output of the encoder to the 36-dimensional structural parameters, and the number of neurons is the reverse order of the encoder network architecture. The discriminator is used to restrict the 2-dimensional space to obey the preset distribution (Gaussian distribution). The number of neurons in each layer is 2, 512, 256, and 1 respectively, and the last neuron outputs a true or false judgment. Figure 3Each column in the above example represents a layer of the deep learning neural network. For example, the second column in the decoder, 512 indicates that the current layer has 512 neurons. The connection order of the functions is Linear, Batchnorm, and LeakRelu. In addition, the generative model uses mean square error as the loss function, and the discriminative model uses binary cross entropy to update the weights of the hidden layer. The training process is as follows: Figure 1 As shown in the figure, the trained decoder can generate high-quality thermal radiation material designs for tens of thousands of 2D inputs within a few seconds, and the distribution of the 2D space conforms to the preset Gaussian distribution. In order to combine with the electromagnetic simulation and calculation evaluation index steps later, the threshold is set to 0.5 to achieve the binarization of the generated model output.
[0060] Iterative steps: Combine the trained deep learning neural network model with Bayesian optimization to iterate the thermal radiation material structure design to obtain the optimal thermal radiation material structure design. That is, combine the trained generative model with Bayesian optimization to perform iterative cycles of recommendation and evaluation, and repeat until the optimal structural design under the target requirements is obtained.
[0061] The iterative steps include the following steps: Step S3.1: Digital sequence representation, electromagnetic simulation and evaluation index are the same as step S1. Step S3.2: The decoder in the adversarial autoencoder saved in step S2 is used as a generative model, and a threshold is set for the output to adapt to the real material design. Step S3.3: The output of the generative model is connected to the input of the electromagnetic simulation, the output of the electromagnetic simulation is connected to the evaluation index calculation, and the calculation result is connected to the input of the Bayesian optimization, and the output of the Bayesian optimization is connected to the input of the generative model to form a closed loop. Repeat the above steps until the design meets the target or the iteration exceeds the maximum number of times. Figure 2 As shown, Bayesian optimization is used to search in 2D space, and the 2D coordinates are mapped to the actual material structure design according to the generative model. Finally, electromagnetic simulation and evaluation index calculation are combined to feed back to the Bayesian model, and an iterative cycle of recommendation and evaluation is carried out until the design meets the target.
[0062] Specifically, first, 300 2D coordinates are randomly selected as inputs and the corresponding structural designs are obtained according to the generative model as the initial candidate structures. The spectral response and evaluation indicators are calculated and input into the Bayesian model. The Bayesian model will then learn the posterior distribution of the black box objective function based on the prior distribution. Combined with the acquisition function, the Bayesian model will recommend and verify the next 2D coordinate within the set range of (-10, 10). Repeating the above process can quickly find the target material structure design in high-dimensional space. The spectral responses corresponding to the optimal structural design when the final target wavelength is 4.5 (dashed line), 5.5 (dotted line) and 6.5 (solid line) microns are shown below. Figure 4As shown, it can be found that the designs under each type of objectives can meet the target requirements.
[0063] The present invention is based on a data-driven deep learning neural network method. The neural network trained on the basis of the data set can learn the complex correspondence between high-dimensional thermal radiation structure and optical response, realize continuous and compact mapping from high-dimensional space to low-dimensional space, and ensure the design premise of high-quality structure. By using the decoder in the model obtained by training and combining iterative search with Bayesian optimization, the reverse structural design of the target thermal radiation characteristics can be finally realized with extremely high accuracy.
[0064] The present invention discloses a method for designing thermal radiation materials on demand based on deep learning neural networks and Bayesian optimization. The method is aimed at designing thermal radiation materials in high-dimensional candidate spaces, uses an adversarial autoencoder model to learn the correspondence between the structure of thermal radiation materials and the spectral response, and compresses the distribution of the candidate space, and then applies Bayesian optimization to the compressed space, thereby achieving global and efficient search. It solves the problems of time-consuming iterative calculations, easy to fall into local optimality, and difficult to apply to structural systems with high degrees of freedom in traditional optimization algorithms, and the problems of time-consuming data set preparation and difficulty in convergence of loss functions in neural network methods. The present invention can be used to design complex thermal radiation materials on demand based on spectral inverse, which can greatly shorten the cost and cycle of material research and development.
[0065] The present invention is a method for optimizing the design of thermal radiation materials. By adopting a design framework combining adversarial autoencoders and Bayesian optimization, a high-dimensional candidate space is mapped into a low-dimensional, continuous and compact space, and then Bayesian optimization is used to efficiently search in the mapped space to achieve on-demand design of high-dimensional thermal radiation materials. The present invention is aimed at the design of thermal radiation materials in high-dimensional candidate spaces, uses an adversarial autoencoder model to achieve compression of the candidate space, and uses Bayesian optimization for compressed space to achieve global optimization.
[0066] An embodiment of the present invention also discloses a thermal radiation material design system based on neural network and Bayesian optimization, including the following modules: a model acquisition module: acquiring and training a deep learning neural network model.
[0067] The model acquisition module includes the following modules: Module M1: Prepare the data set according to the material system and design goals. Module M1 includes the following modules: Characteristic acquisition module: Represent the material structure with a digital sequence and use it as the input of electromagnetic simulation to obtain the thermal radiation characteristics of the material. Standard determination module: Determine the evaluation criteria according to the design goals. Data set acquisition module: Use traditional optimization algorithms or random search to iterate the characteristic acquisition module and the standard determination module, and obtain the data set after screening.
[0068] Module M2: Use the data set to train the adversarial autoencoder and extract the decoder in the adversarial autoencoder as the generative model. Module M2 includes the following modules: Network determination module: Perform data expansion on the data set and determine the adversarial autoencoder network architecture. Training model module: Use the data set after data expansion to train the adversarial autoencoder and save the model inside the adversarial autoencoder in blocks.
[0069] Iteration module: Combine the trained deep learning neural network model with Bayesian optimization to iterate the thermal radiation material structure design to obtain the thermal radiation material structure design. Connect the output of the generated model to the input of the electromagnetic simulation, the output of the electromagnetic simulation to the evaluation standard, the calculation result under the evaluation standard to the input of the Bayesian optimization, the output of the Bayesian optimization to the input of the generated model, and iterate until the design meets the target.
[0070] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.
[0071] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.
Claims
1. A thermal radiation material design method based on neural network and Bayesian optimization, It is characterized in that The steps include: Model acquisition step: obtain and train the deep learning neural network model; Iteration step: Combine the trained deep learning neural network model with Bayesian optimization to iterate the thermal radiation material structure design to obtain the thermal radiation material structure design; The model acquisition step comprises the following steps: Step S1: Prepare data sets based on material systems and design goals; Step S2: Use the dataset to train the adversarial autoencoder and extract the decoder in the adversarial autoencoder as a generative model; The step S1 comprises the following steps: Characteristic acquisition step: the material structure is represented by a digital sequence and used as the input of electromagnetic simulation to obtain the thermal radiation characteristics of the material; Standard determination steps: Determine the evaluation criteria based on the design goals; Dataset acquisition step: Use traditional optimization algorithms or random search to iterate the feature acquisition step and the standard determination step, and obtain the data set after screening; The step S2 comprises the following steps: Network determination step: expand the data set and determine the adversarial autoencoder network architecture; Training model steps: Use the data set after data expansion to train the adversarial autoencoder and save the model inside the adversarial autoencoder in blocks; In the iterative step, the output of the generated model is connected to the input of the electromagnetic simulation, the output of the electromagnetic simulation is connected to the evaluation standard, the calculation result under the evaluation standard is connected to the input of the Bayesian optimization, the output of the Bayesian optimization is connected to the input of the generated model, and the cycle is iterated until the design meets the target.
2. A thermal radiation material design system based on neural network and Bayesian optimization, It is characterized in that Includes the following modules: Model acquisition module: acquire and train deep learning neural network models; Iteration module: Combine the trained deep learning neural network model with Bayesian optimization to iterate the thermal radiation material structure design to obtain the thermal radiation material structure design; The model acquisition module includes the following modules: Module M1: Prepare data sets for material systems and design goals; Module M2: Use the dataset to train the adversarial autoencoder and extract the decoder in the adversarial autoencoder as the generative model; The module M1 includes the following modules: Characteristic acquisition module: The material structure is represented by a digital sequence and used as the input of electromagnetic simulation to obtain the thermal radiation characteristics of the material; Standard determination module: Determine the evaluation criteria based on the design objectives; Dataset acquisition module: Use traditional optimization algorithms or random search to iterate the feature acquisition module and the standard determination module to obtain the data set after screening; The module M2 includes the following modules: Network determination module: expand the data set and determine the adversarial autoencoder network architecture; Training model module: Use the data set after data expansion to train the adversarial autoencoder and save the model inside the adversarial autoencoder in blocks; In the loop iteration module, the output of the generated model is connected to the input of the electromagnetic simulation, the output of the electromagnetic simulation is connected to the evaluation standard, the calculation result under the evaluation standard is connected to the input of the Bayesian optimization, the output of the Bayesian optimization is connected to the input of the generated model, and the loop iteration is carried out until the design meets the target.
Citation Information
Patent Citations
Heat radiating material
CN104277241A
Neural network architecture evaluation method based on attribute graph optimization
CN110232434A
Heterogeneous software workload estimation method based on deep learning
CN110322222A
Bayesian structure learning method and device of deep neural network
CN110738242A
Variable bit rate generative compression method based on adversarial learning
CN111787323A
