Neural network film reverse design method based on feature enhancement and optical film

Abstract

The application discloses a neural network film reverse design method based on feature enhancement. The application uses the feature film thickness and random film thickness of a classical film system to form a feature-enhanced film structure dataset; a convolutional neural network (CNN) and a long short-term memory network (LSTM) are combined to form a reverse hybrid neural network, which is connected in series with a forward single fully connected neural network to form a film reverse design neural network model; the film reverse design neural network model is trained to obtain optimized model parameters; a target spectrum is input into the trained reverse hybrid neural network, and finally a film system structure that matches the target spectrum is designed. The application is based on the feature-enhanced film structure dataset, reduces the dependence of the neural network on the dataset size, significantly reduces the loss value of the target spectrum test set, and improves the accuracy of the target spectrum prediction. Meanwhile, the application uses the hybrid neural network to capture the multi-dimensional features of the spectrum, and improves the training speed of the film reverse design neural network model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of optical film design, and specifically relates to a neural network film reverse design method based on feature enhancement and an optical film thereof, which is suitable for the design of various optical films such as high-reflection films and narrow-band optical filters, and can be widely applied in the fields of communication technology, display devices, laser technology and optical instruments. BACKGROUND

[0002] As a material composed of multiple layers of media and having various optical properties, optical films have been widely used in many fields such as communication technology, display devices, laser technology, optical instruments and scientific research, and are indispensable components of modern optical systems. However, the rapid development of technology also puts forward higher standards and expectations for the performance of optical films.

[0003] In recent years, with the rapid development of deep learning, its powerful nonlinear fitting capability and efficient optimization characteristics make it a very efficient tool in the design of optical films, such as using neural networks to predict the spectral response of thin films, and predicting the composition of film systems and film thickness according to the spectrum.

[0004] A neural network is usually composed of an input layer, one or more hidden layers and an output layer, each layer containing a number of neurons, each neuron having its own unique parameters. The process of training a neural network using a data set is also the process of learning the characteristics of the data set and adjusting the parameters inside the model. When designing a thin film structure, different thin film structures may produce similar spectra, which causes the inverse design process of inputting the expected spectrum and outputting the thin film structure to be prone to multiple solutions, making it difficult to obtain good inverse design results directly using a neural network. To solve the above problems, researchers propose to use two single neural networks in series to train the neural network model with a data set of thin film structures formed by random film thickness, such as CN202310480403.X, CN202310335182.7 and ACS Photonics, 2018, 5: 1365-1369. This kind of single neural network in series method can achieve good thin film structure design effect for target spectra with fewer layers.

[0005] However, when the number of layers required by the target spectrum is large, the training of the neural network model requires an exponential increase in the size of the data set, resulting in an exponential increase in the difficulty of model training. The existing single neural network in series model based on random data has the problems of long training time due to large training data volume, and poor fitting effect due to the poor feature extraction capability of the single neural network, which makes it difficult to meet the inverse design requirements of thick film layers and complex film system structures. SUMMARY

[0006] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art, and a neural network thin film reverse design method based on feature enhancement is proposed. The present application is based on a feature-enhanced thin film structure dataset, which reduces the dependence of the neural network on the size of the dataset, significantly reduces the loss value of the target spectrum test set, and improves the accuracy of target spectrum prediction. At the same time, the present application uses a hybrid neural network to capture the multi-dimensional features of the spectrum, and improves the training speed of the thin film reverse design neural network model.

[0007] The technical solution of the present application is as follows:

[0008] A neural network thin film reverse design method based on feature enhancement, characterized in that it comprises the following steps:

[0009] Step 1, dataset creation

[0010] (1) Establish a program for calculating the spectrum using the feature matrix method;

[0011] (2) Determine the thin film system structure for training, including the base material, high refractive index material, low refractive index material, number of thin film layers, material of each thin film layer, angle of incident light, polarization state and wavelength range, number of wavelength sampling points n; X sets of ideal film thickness data of thin films under different working wavelengths; Y sets of random film thickness data of each layer of thin film; N sets of feature-enhanced thin film structure data sets, N = X + Y;

[0012] (3) The X sets of ideal film thickness data and the corresponding reflection spectrum are used as feature data; the Y sets of random film thickness data and the corresponding reflection spectrum are used as random data; each spectrum contains n data points;

[0013] (4) Mix the random data and the feature data according to a predetermined ratio to form a feature-enhanced training dataset.

[0014] Step 2, model building

[0015] (1) Build a forward single fully connected neural network, starting from the input layer, sequentially connecting several fully connected neural network layers, and an output layer;

[0016] (2) Combine the convolutional neural network and the long short-term memory neural network to build a reverse hybrid neural network, starting from the input layer, sequentially connecting the convolutional layer, the normalization layer, the activation layer, the pooling layer, the long short-term memory network module layer, several fully connected neural network layers, and the output layer;

[0017] (3) Connect the reverse hybrid neural network and the forward single fully connected neural network in series to form a thin film reverse design neural network model. Set the optimization algorithm, initial learning rate, learning rate drop factor and period, activation function, small batch data size, L2 regularization factor and maximum training period number.

[0018] Step 3: Model Training

[0019] (1) Use the training dataset with feature enhancement in step one to train the feedforward single fully connected neural network to form a trained feedforward single fully connected neural network.

[0020] (2) Train the neural network model for the reverse design of the tandem thin film using the training dataset for feature enhancement in step one, and construct the loss function as follows: Where R ij It is the j-th data value of the i-th group of reflectance spectra in the training dataset, r ij It is the j-th data value of the reflection spectrum output by the inverse hybrid neural network module after the reflection spectrum of the i-th group in the training dataset is processed by the reflection spectrum output by the forward single fully connected neural network module, i = 1, 2, 3...N, j = 1, 2, 3...n; N is the number of thin film structure groups, and n is the number of sampling points for each spectrum.

[0021] (3) Obtain the trained thin-film reverse design neural network model.

[0022] Step 4: Reverse Engineering

[0023] Input the target spectrum of the membrane system to be designed, and use the inverse hybrid neural network module in the trained thin film inverse design neural network model to obtain the thickness data of the designed membrane layer.

[0024] In step one, the classical film system includes high-reflection films, narrow-band filters, and anti-reflection films, etc.

[0025] The feedforward single fully connected neural network module includes: an input layer for receiving film thickness parameters; multiple fully connected layers for fitting the functional relationship between the film thickness parameters and the reflectance spectrum; and an output layer for outputting the predicted reflectance spectrum.

[0026] The inverse hybrid neural network module includes: an input layer for receiving reflectance spectral data; a CNN feature extraction module for extracting spatial local features of the spectral data; an LSTM feature extraction module for capturing the long-term dependence of wavelength and reflectance in the spectral sequence; multiple fully connected layers for enhancing the nonlinear fitting capability of the model; and an output layer for outputting the predicted film thickness parameters.

[0027] A thin film reverse design system is characterized by comprising: a data preprocessing module for generating the training dataset; a model building module for constructing the tandem thin film reverse design neural network; a training module for training the tandem thin film reverse design neural network; and a reverse design module for inputting the target spectrum and outputting film thickness data.

[0028] An optical thin film designed using the method described above is characterized in that its film thickness parameter is determined by the output of a trained inverse hybrid neural network module.

[0029] A method for preparing an optical thin film, characterized by comprising: determining film thickness parameters using the method; and sequentially depositing thin film materials on a substrate according to the parameters.

[0030] Compared with existing technologies, the technical solution of the present invention has the following advantages:

[0031] 1. Traditional methods rely solely on randomly generated film thickness data to train neural networks, resulting in extremely high data requirements and difficulty in capturing the physical properties of the target spectrum. This invention uses characteristic film thicknesses of the relevant classical film system and random film thickness combinations to form a feature-enhanced thin film structure dataset. Its corresponding reflectance spectra form a feature-enhanced training dataset, used for training neural network models for thin film inverse design. This enhances the physical meaning of the training data, significantly reduces the neural network's dependence on dataset size, and greatly improves the accuracy of the neural network's target spectrum prediction.

[0032] 2. The method of this invention combines convolutional neural networks and long short-term memory networks, which reduces the training time while achieving better fitting results.

[0033] 3. A closed-loop optimization is formed by concatenating a reverse hybrid neural network (CNN-LSTM) with a forward single fully connected neural network. The forward single fully connected neural network verifies whether the film thickness output by the reverse is consistent with the target spectrum, and adjusts the parameters inversely using a loss function. The model is suitable for the optimization design of optical coatings. Attached Figure Description

[0034] Figure 1 A schematic diagram of the tandem thin-film reverse design neural network model proposed in this invention.

[0035] Figure 2 A schematic diagram of the feedforward single fully connected neural network in this invention.

[0036] Figure 3 A schematic diagram of the inverse hybrid neural network proposed in this invention.

[0037] Figure 4 The spectral prediction effect of the neural network trained on the high-reflectivity thin film feature enhancement dataset in Embodiment 1 of the present invention.

[0038] Figure 5 The spectral prediction effect of the neural network trained on three types of membrane feature enhancement datasets in Embodiment 2 of this invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the embodiments and accompanying drawings, but the listed embodiments should not limit the scope of protection of the present invention.

[0040] Example 1

[0041] This embodiment uses a high-reflectivity thin film with 0° incident light as an example to illustrate the feature-enhanced neural network thin film reverse design method of the present invention.

[0042] Step 1: Dataset Creation

[0043] (1) Establish a program to calculate the spectrum using the characteristic matrix method;

[0044] (2) In this embodiment, the film structure of the high-reflectivity film used for training is sub / H (LH). 5 / air, the substrate sub is made of SiO2, the high refractive index film H is made of HfO2 material, and the low refractive index film L is made of SiO2 material. The incident light polarization state is S polarization, and the incident angle is 0°.

[0045] (3) Randomly set the thickness of each thin film in the range of 10-270nm, calculate the reflection spectrum using the feature matrix method, and generate 200,000 sets of random datasets1.

[0046] Sampling was performed at 2nm intervals within the wavelength range of 400-1200nm, resulting in 400 wavelength data points. The thickness of each layer of the high-reflectivity thin film at each wavelength λ is... n′ is the refractive index of each thin film at wavelength λ. The reflection spectrum is calculated using the characteristic matrix method, generating 400 sets of characteristic datasets.

[0047] (4) Repeatedly replace the above 400 sets of feature data into random dataset 1 to form 200,000 sets of feature enhancement datasets 2, 3, and 4 of 10%, 20%, and 30% high reflectivity thin film feature data.

[0048] Step 2: Model Building

[0049] (1) Construct a feedforward single fully connected neural network, the model structure is as follows: Figure 2 As shown, the feedforward single fully connected neural network sequentially connects to an input layer, several fully connected neural network layers, and an output layer. The input layer transmits the film thickness of each layer to the fully connected neural network layers. The several fully connected neural network layers fit the functional relationship between the film thickness parameters and the reflectance spectrum. The output of the output layer is the reflectance spectrum data of the film. The neural units of the fully connected neural network layer are 64-128-256-512-1024-512.

[0050] (2) A hybrid neural network combining convolutional neural networks and long short-term memory neural networks is used as the inverse hybrid neural network. Starting from the input layer, convolutional layers, normalization layers, activation layers, pooling layers, long short-term memory network module layers, several fully connected neural network layers, and the output layer are connected in sequence, such as... Figure 3 As shown. The input layer transmits the reflectance spectral data of the optical thin film to the convolutional module, extracts the spatial local features of the spectral data, and then captures the long-term dependence of the wavelength and reflectance of the spectral sequence through the long short-term memory module. Finally, it passes through several fully connected neural network layers, utilizing its powerful nonlinear fitting ability to optimize the model, and outputs the film thickness parameters of the optical thin film through the output layer. The convolutional layer has a kernel size of 3×1 and a kernel count of 8; the long short-term memory neural network module has 400 hidden units; and the fully connected neural network layer has 512-1024-512-256-128 neural units.

[0051] (3) Construct a thin-film reverse design neural network model with a cascaded structure of inverse hybrid neural network and forward single fully connected neural network, such as... Figure 1 As shown, configure the optimization algorithm, initial learning rate, learning rate reduction factor and cycle, activation function, mini-batch size, L2 regularization factor, and maximum number of training cycles.

[0052] Both neural network modules use ReLU layers as the activation function of the neural network, and the optimizer used is SGDM (Stochastic Gradient Descent with Momentum).

[0053] Step 3: Model Training

[0054] (1) Use the feature enhancement dataset in step one to train the feedforward single fully connected neural network to form a trained feedforward single fully connected neural network.

[0055] (2) Train the neural network model for the reverse design of the tandem structure thin film using the feature enhancement dataset from step one, and construct the loss function as follows: Where R ij It is the j-th data value of the i-th group of reflectance spectra in the dataset, r ij It is the j-th data value of the reflection spectrum output by the inverse hybrid neural network module after the reflection spectrum of the i-th group in the dataset is processed by the reflection spectrum output by the forward single fully connected neural network module, i = 1, 2, 3...N, j = 1, 2, 3...n; N is the number of thin film structure groups, and n is the number of sampling points for each spectrum.

[0056] (3) Obtain the trained thin-film reverse design neural network model.

[0057] Step 4: Validation and Thin Film Reverse Design

[0058] (1) Verification

[0059] The thickness of each thin film was randomly set within the range of 10-270 nm. The reflection spectrum was calculated using the characteristic matrix method to generate 500 sets of random data test sets, which were not duplicates of the random dataset 1 in step one.

[0060] Sampling was performed at 2nm intervals within the wavelength range of 401-1199nm, resulting in 400 wavelength data points. The thickness of each layer of the high-reflectivity thin film at each wavelength λ is... n′ is the refractive index of each thin film at wavelength λ. The reflection spectrum is calculated using the characteristic matrix method, generating a test set of 400 characteristic data sets.

[0061] The trained inverse hybrid neural network was tested using the test set described above, and the results of the loss values ​​are shown in Table 1.

[0062] Table 1. Loss values ​​of neural networks trained on datasets with different compositions.

[0063]

[0064] As can be seen, for neural networks obtained using feature enhancement training methods, the loss value of the feature data test set can be significantly reduced without increasing the loss value of the random data test set, thereby improving the accuracy of target spectrum prediction.

[0065] (2) Thin film reverse design

[0066] The reflection spectrum of the high-reflectivity thin film at a wavelength of 651 nm was used as the target spectrum and input into the trained inverse hybrid neural network to obtain the film thickness data after design, as shown in Table 2.

[0067] Table 2. Thickness design results of high-reflectivity thin films at 651nm wavelength

[0068]

[0069]

[0070] The prediction results of four neural networks for the same high-reflectance thin film spectrum are as follows: Figure 4 As shown, the neural network trained with random data has a poor fitting effect on the target spectrum, while the neural network trained with feature data has a significantly improved fitting effect and can accurately predict the target spectrum.

[0071] Example 2

[0072] This embodiment uses a high-reflectivity thin film with 0° incident light and two narrowband filters as examples to illustrate the feature-enhanced neural network thin film reverse design method of the present invention.

[0073] In this embodiment, the filter used for training has film structures of sub / (LH)^3L 2H(LH)^3 / air and sub / LHL(LHLHLHLHL)(LH)^2 / air, and the high-reflectivity film has a film structure of sub / (LH)^7 / air. The substrate sub is a SiO2 substrate, the high-refractive-index film H is made of HfO2 material, and the low-refractive-index film L is made of SiO2 material. The incident light polarization state is S-polarization, and the incident angle is 0°.

[0074] 480,000 sets of random datasets were generated using the method in step one of Example 1.

[0075] The method in step one of Example 1 was used to generate a feature dataset of a high-reflectivity thin film and two narrowband filters, with 400 sets of data for each film system, for a total of 1200 sets of data.

[0076] The above 1200 sets of feature data were repeatedly replaced in the random dataset 5 to form a feature-enhanced dataset 6 with 480,000 sets of feature data accounting for 5% of the total feature data.

[0077] The neural network model for reverse design of the tandem thin film was trained and verified according to steps two to four in Example 1, and the reverse design of the thin film structure was performed. The design results are as follows: Figure 5 As shown, although each feature data accounts for only 1.67% of the training set, it still demonstrates a better fitting effect in the design of the target spectrum compared to neural networks trained with random data.

[0078] The specific embodiments described herein are for illustrative purposes only and do not constitute any limitation on the scope of protection of the claims. Those skilled in the art, based on an understanding of the inventive concept, can easily make equivalent substitutions, conventional improvements, or adaptive adjustments to the technical solutions; such derivatives should still be considered within the scope of protection defined by the claims of this invention.

Claims

1. A reverse design method for neural network thin films based on feature enhancement, characterized in that, Includes the following steps: Step 1: Construct a training dataset containing classic membrane system feature data, which is obtained by calculating the feature matrix method; Step 2: Establish a neural network model for the reverse design of tandem thin films, including: A forward-feeding single fully connected neural network module is used to predict the reflectance spectrum based on the film thickness parameter; The inverse hybrid neural network module employs a hybrid structure of convolutional neural networks and long short-term memory networks to predict film thickness parameters based on reflectance spectra. Step 3: Use the training dataset to jointly train the tandem thin film reverse design neural network model. During the training process, a closed-loop optimization strategy is adopted, and the output of the reverse hybrid neural network is verified through a feedforward single fully connected neural network. Step 4: Input the target spectrum into the trained inverse hybrid neural network module, and output the designed film thickness parameters; The construction of the training dataset includes: Obtain X sets of ideal film thickness data and their corresponding reflectance spectra generated based on classical film system rules as feature data; Obtain the randomly generated film thickness data and its corresponding reflectance spectrum from group Y as random data; Feature data and random data are mixed in a preset ratio to form a training dataset with enhanced features; The inverse hybrid neural network module includes: Input layer: Receives reflectance spectral data; CNN feature extraction module: Extracts local spatial features from spectral data; LSTM feature extraction module: captures the long-term dependence of wavelength and reflectance in spectral sequences; Multiple fully connected layers: enhance the model's non-linear fitting capability; Output layer: Outputs the predicted film thickness parameters; The thin film structure used for training includes the substrate material, high refractive index material, low refractive index material, number of thin film layers, material of each thin film layer, as well as the angle, polarization state and wavelength range of the incident light, and the number of wavelength sampling points; X sets of ideal film thickness data for different working wavelengths; Y sets of data on random variations in the thickness of each thin film layer; and a total of N sets of feature-enhanced thin film structure datasets, N = X + Y. The loss function used in training the tandem thin-film reverse design neural network model is: Among them, R ij It is the j-th data value of the i-th group of reflectance spectra in the training dataset, r ij It is the j-th data value of the reflection spectrum output by the inverse hybrid neural network module after the reflection spectrum of the i-th group in the training dataset is processed by the reflection spectrum output by the forward single fully connected neural network module, i = 1, 2, 3...N, j = 1, 2, 3...n; N is the number of thin film structure groups, and n is the number of wavelength sampling points for each spectrum.

2. The reverse design method for neural network thin films based on feature enhancement according to claim 1, characterized in that, The classic film system includes high-reflection films, narrow-band filters, and anti-reflection films.

3. The reverse design method for neural network thin films based on feature enhancement according to claim 1, characterized in that, The feedforward single fully connected neural network module includes: Input layer: receives film thickness parameters; Multiple fully connected layers: Fitting the functional relationship between film thickness parameters and reflectance spectrum; Output layer: Outputs the predicted reflectance spectrum.

4. A thin film reverse engineering system, characterized in that, For implementing the method as described in any one of claims 1-3; comprising: The data preprocessing module is used to build the training dataset; The model building module is used to build neural network models for tandem thin film reverse design. The model training module is used to perform the joint training process; The reverse engineering module is used to input the target spectrum and output film thickness data.

5. An optical thin film designed using the method described in any one of claims 1-3, characterized in that, Its film thickness parameter is determined by the output of the trained inverse hybrid neural network module.

Images