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

Abstract

The invention discloses a neural network film reverse design method based on feature enhancement. The method comprises the following steps: forming a feature-enhanced thin film structure data set by using the combination of a feature film thickness and a random film thickness of a classic film system; a convolutional neural network (CNN) and a long short-term memory network (LSTM) are combined to form a reverse hybrid neural network, and the reverse hybrid neural network is connected in series with a forward single full-connection neural network to form a thin film reverse design neural network model; training the film reverse design neural network model to obtain optimized model parameters; and inputting the target spectrum into the trained reverse hybrid neural network, and finally designing a film system structure matched with the target spectrum. According to the method, the dependence of the neural network on the size of the data set is reduced based on the feature-enhanced thin film structure data set, the loss value of the target spectrum test set is remarkably reduced, and the accuracy of target spectrum prediction is improved. Meanwhile, the multi-dimensional features of the spectrum are captured by using the hybrid neural network, so that the training speed of a thin film reverse design neural network model is improved.

Description

Technical Field

[0001] The present invention belongs to the field of optical film design technology, specifically a feature-enhanced neural network film inverse design method and optical film thereof, which is suitable for the design of various optical films such as high-reflective films and narrow-band filters, and can be widely used in communication technology, display devices, laser technology, optical instruments and other fields. Background Art

[0002] Optical thin films, composed of multiple layers of dielectric materials and exhibiting diverse optical properties, are widely used in numerous fields, including communications technology, display devices, laser technology, optical instruments, and scientific research. They are indispensable components of modern optical systems. However, the rapid development of technology has also raised higher standards and expectations for the performance of optical thin films.

[0003] In recent years, with the rapid development of deep learning, its powerful nonlinear fitting capabilities and efficient optimization characteristics have made it a very effective tool in optical thin film design. For example, neural networks are used to predict the spectral response of thin films, and the composition and thickness of film systems are predicted based on spectra.

[0004] A neural network typically consists of three components: an input layer, one or more hidden layers, and an output layer. Each layer contains numerous neurons, each with its own unique parameters. The process of training a neural network using a dataset involves learning the characteristics of the dataset and adjusting the model's internal parameters. When designing thin film structures, different thin film structures may produce similar spectra. This makes the reverse design process, where the input is the desired spectrum and the output is the thin film structure, susceptible to multiple solutions, making it difficult to obtain good reverse design results using a neural network directly. To address this issue, researchers have proposed using two single neural networks in series to train the neural network model using a dataset of thin film structures with random film thicknesses. Examples include CN202310480403.X, CN202310335182.7, and ACS Photonics, 2018, 5:1365-1369. This single neural network cascade approach can achieve good thin film structure design results for target spectra with a small number of film layers.

[0005] However, when the target spectrum requires a large number of film layers, the dataset size required for neural network model training increases exponentially, leading to an exponential increase in the difficulty of model training. Existing single neural network cascade models based on random data suffer from long training times due to the large amount of training data, and poor fitting results due to the single neural network's poor ability to extract features from the dataset. This makes it difficult to meet the reverse design requirements for thick film layers and complex film systems. Summary of the Invention

[0006] The technical problem addressed by the present invention is to overcome the shortcomings of the aforementioned prior art by proposing a feature-enhanced neural network-based thin film reverse design method. Based on a feature-enhanced thin film structure dataset, the present invention reduces the neural network's dependence on dataset size, significantly lowers the loss value of the target spectral test set, and improves the accuracy of target spectrum prediction. Furthermore, the present invention utilizes a hybrid neural network to capture the multidimensional features of the spectrum, thereby increasing the training speed of the thin film reverse design neural network model.

[0007] The technical solutions of the present invention are as follows:

[0008] A neural network thin film inverse design method based on feature enhancement is characterized in that it includes the following steps:

[0009] Step 1: Dataset creation

[0010] (1) Establish a program to calculate spectra using the characteristic matrix method;

[0011] (2) Determine the thin film system structure used for training, including 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 n; X groups of ideal film thickness data of the thin film at different working wavelengths; Y groups of random variation data of the thickness of each film layer; a total of N groups of feature-enhanced thin film structure data sets, N = X + Y;

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

[0013] (4) Random data and feature data are mixed in a preset ratio to form a feature-enhanced training data set.

[0014] Step 2: Model Construction

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

[0016] (2) Convolutional neural networks and long short-term memory neural networks are combined to construct an inverse hybrid neural network, starting from the input layer, and sequentially connecting the convolution layer, normalization layer, activation layer, pooling layer, 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 reduction factor and period, activation function, mini-batch data size, L2 regularization factor, and maximum number of training cycles.

[0018] Step 3: Model training

[0019] (1) using the feature-enhanced training data set in step 1 to train a forward single fully connected neural network to form a trained forward single fully connected neural network;

[0020] (2) Use the feature-enhanced training data set in step 1 to train the thin film reverse design neural network model of the series structure, and construct the loss function as where R ij is the jth data value of the i-th group of reflectance spectrum in the training data set, r ij It is the jth data value of the reflectance spectrum output by the forward single fully connected neural network module after the film thickness parameters of the i-th group of reflectance spectra in the training data set are output by the reverse hybrid neural network module; i = 1, 2, 3...N, j = 1, 2, 3...n; N is the number of 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] The target spectrum of the film system to be designed is input, and the designed film thickness data is obtained using the inverse hybrid neural network module in the trained thin film inverse design neural network model.

[0024] In the step 1, the classical film system includes a high-reflection film, a narrow-band filter, an anti-reflection film, and the like.

[0025] The forward 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 reflection spectrum; and an output layer for outputting the predicted reflection spectrum.

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

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

[0028] An optical film designed using the method is characterized in that the 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 is characterized by comprising: determining film thickness parameters using the method; and sequentially depositing layers of thin film material on a substrate according to the parameters.

[0030] Compared with the existing technology, the above 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 capturing the physical laws of the target spectrum. The method of the present invention uses a combination of characteristic film thicknesses of the classical film system and random film thicknesses to form a feature-enhanced thin film structure dataset, and the corresponding reflectance spectra to form a feature-enhanced training dataset for training the thin film reverse design neural network model. This enhances the physical significance of the training data, significantly reduces the neural network's dependence on dataset size, and significantly improves the accuracy of the neural network's prediction of the target spectrum.

[0032] 2. The method of the present invention combines convolutional neural networks with long short-term memory networks, which not only achieves better fitting effects but also reduces the time required for training.

[0033] 3. A closed-loop optimization loop is formed by cascading 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 reverse output film thickness matches the target spectrum and adjusts the parameters inversely using a loss function. This model is suitable for the optimization design of optical coatings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 . Schematic diagram of the neural network model for reverse design of series thin films proposed in this invention.

[0035] Figure 2 .Schematic diagram of the forward single fully connected neural network in the present invention.

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

[0037] Figure 4 . Spectral prediction effect diagram of the neural network trained based on the high-reflective film feature enhancement dataset in Example 1 of the present invention.

[0038] Figure 5 . Spectral prediction effect diagram of the neural network trained based on three film feature enhancement data sets in Example 2 of the present invention. DETAILED DESCRIPTION

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

[0040] Example 1

[0041] This embodiment takes a highly reflective film at 0° incidence as an example to illustrate the neural network film inverse design method based on feature enhancement of the present invention.

[0042] Step 1: Dataset creation

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

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

[0045] (3) The thickness of each film layer was randomly set in the range of 10-270 nm, and the reflectance spectrum was calculated using the characteristic matrix method to generate 200,000 sets of random data sets1.

[0046] The wavelength range of 400-1200nm is sampled at 2nm intervals, and 400 wavelength data are obtained. The thickness of each layer of the high reflective film at each wavelength λ is n′ is the refractive index of each film at wavelength λ. The reflectance spectrum is calculated using the characteristic matrix method to generate 400 sets of characteristic data sets.

[0047] (4) Repeatedly replace the above 400 sets of feature data into the random data set 1 to form 200,000 sets of feature enhanced data sets 2, 3, and 4 with 10%, 20%, and 30% high-reflection film feature data.

[0048] Step 2: Model Construction

[0049] (1) Construct a forward single fully connected neural network with the following model structure: Figure 2 As shown in Figure 2, the forward single fully connected neural network sequentially connects the input layer, several fully connected neural network layers, and the output layer. The input layer transmits the thickness of each layer of the film to the fully connected neural network layer. 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) Using a hybrid neural network that combines convolutional neural networks with long short-term memory neural networks as the reverse hybrid neural network, starting from the input layer, sequentially connect the convolution layer, normalization layer, activation layer, pooling layer, long short-term memory network module layer, several fully connected neural network layers and the output layer, such as Figure 3 As shown in the figure, the input layer transmits the reflectance spectrum data of the optical film to the convolution module, extracting the spatial local features of the spectral data. The long-term short-term memory module then captures the long-term dependence between the wavelength and reflectivity of the spectral sequence. Finally, it passes through several fully connected neural network layers, utilizing their powerful nonlinear fitting capabilities to optimize the model, and the output layer outputs the film thickness parameters of the optical film. The convolutional layer has a convolution kernel size of 3×1 and a total of 8 convolution kernels; the long-term short-term memory neural network module has 400 hidden units; and the fully connected neural network layer has a neural unit ratio of 512-1024-512-256-128.

[0051] (3) The reverse hybrid neural network and the forward single fully connected neural network are constructed into a series-structured thin film reverse design neural network model, such as Figure 1 Set the optimization algorithm, initial learning rate, learning rate reduction factor and period, activation function, mini-batch size, L2 regularization factor, and maximum number of training cycles.

[0052] Both neural network modules use the ReLu layer 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) Using the feature enhancement data set in step 1, a forward single fully connected neural network is trained to form a trained forward single fully connected neural network;

[0055] (2) Use the feature enhancement data set in step 1 to train the thin film reverse design neural network model of the series structure, and construct the loss function as where R ij is the jth data value of the i-th group of reflectance spectrum in the data set, r ij It is the jth data value of the reflectance spectrum output by the forward single fully connected neural network module after the film thickness parameters of the i-th group of reflectance spectrum in the data set are output by the reverse hybrid neural network module; i = 1, 2, 3...N, j = 1, 2, 3...n; N is the number of 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: Verification and film reverse design

[0058] (1) Verification

[0059] The thickness of each film layer was randomly set in the range of 10-270 nm, and the reflectance spectrum was calculated using the characteristic matrix method to generate 500 sets of random data test sets, which were not repeated with the random data set 1 in step 1.

[0060] The wavelength range of 401-1199 nm is sampled at 2 nm intervals, and 400 wavelength data are obtained. The thickness of each layer of the high reflective film at each wavelength λ is n′ is the refractive index of each film at wavelength λ. The reflectance spectrum is calculated using the characteristic matrix method to generate 400 sets of characteristic data test sets.

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

[0062] Table 1. Neural network loss values for training different datasets

[0063]

[0064] It can be seen that for the neural network obtained by using the feature enhancement training method, the loss value of the feature data test set can be greatly reduced without increasing the loss value of the random data test set, thereby improving the accuracy of the target spectrum prediction.

[0065] (2) Thin film reverse design

[0066] The reflectance spectrum of the highly reflective 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 designed film thickness data, as shown in Table 2.

[0067] Table 2 Thickness design results of high reflective film at 651nm wavelength

[0068]

[0069]

[0070] The prediction results of four neural networks for the same high reflective film spectrum are as follows: Figure 4 As shown in the figure, it can be seen that the neural network trained with random data has a poor fitting effect on the target spectrum, while the fitting effect of the neural network trained with feature data has been greatly improved, and the target spectrum can be accurately predicted.

[0071] Example 2

[0072] This embodiment takes a high-reflection film and two narrow-band filters at 0° incidence as examples to illustrate the neural network film inverse design method based on feature enhancement of the present invention.

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

[0074] The method of step 1 in Example 1 was used to generate 480,000 random data sets 5.

[0075] The method of step 1 in Example 1 was used to generate characteristic data sets for a high-reflection film and two narrow-band filters, with 400 sets of data for each film system, for a total of 1200 sets of data.

[0076] The above 1,200 sets of feature data are repeatedly replaced into the random dataset 5 to form a feature-enhanced dataset 6 of 480,000 sets of feature data, which accounts for 5% of the total feature data.

[0077] According to steps 2 to 4 in Example 1, the reverse design neural network model of the series structure thin film is trained and verified, and the reverse design of the thin film structure is performed. The design results are as follows: Figure 5 As shown in the figure, although each feature data accounts for only 1.67% of the training set, it can still show a better fitting effect in the design of the target spectrum compared to the neural network 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 their understanding of the inventive concept, may readily make equivalent substitutions, conventional improvements, or adaptive adjustments to the technical solutions, and such derivatives shall still be deemed to be within the scope of protection defined by the claims of this invention.

Claims

1. A neural network thin film inverse design method based on feature enhancement, characterized in that: The following steps are involved: Step 1: construct a training data set containing characteristic data of classic film systems, wherein the characteristic data is obtained by calculating the characteristic matrix method; Step 2: Establish a neural network model for reverse design of tandem thin films, including: A forward 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 uses a hybrid structure of convolutional neural network and long short-term memory network (CNN-LSTM) to predict film thickness parameters based on reflectance spectra; Step 3: jointly training the tandem thin film reverse design neural network model using the training data set, adopting a closed-loop optimization strategy during the training process, and verifying the output of the reverse hybrid neural network through a forward 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.

2. The neural network thin film inverse design method based on feature enhancement according to claim 1 is characterized in that: The construction of the training data set includes: Obtain X sets of ideal film thickness data generated based on classical film system rules and their corresponding reflection spectra as characteristic data; Obtain Y groups of randomly generated film thickness data and their corresponding reflection spectra as random data; The feature data is mixed with random data in a preset ratio to form a feature-enhanced training data set.

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

4. The neural network thin film inverse design method based on feature enhancement according to claim 1 is characterized in that: The forward single fully connected neural network module includes: Input layer: receives film thickness parameters; Multiple fully connected layers: Fit the functional relationship between film thickness parameters and reflectance spectrum; Output layer: Outputs the predicted reflectance spectrum.

5. The neural network thin film inverse design method based on feature enhancement according to claim 1 is characterized in that: The inverse hybrid neural network module includes: Input layer: receives reflectance spectrum data; CNN feature extraction module: extracts the spatial local features of spectral data; LSTM feature extraction module: captures the long-term dependency between wavelength and reflectivity of the spectral sequence; Multiple fully connected layers: enhance the nonlinear fitting capability of the model; Output layer: outputs the predicted film thickness parameters.

6. The neural network thin film inverse design method based on feature enhancement according to claim 1, characterized in that: The thin film system structure used for training includes the substrate material, high-refractive index material, low-refractive index material, number of thin film layers, the 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 n; X sets of ideal film thickness data for the thin film at different operating wavelengths; Y sets of data on randomly varying thickness of each film layer; a total of N sets of feature-enhanced thin film structure data sets, N = X + Y.

7. The neural network thin film inverse design method based on feature enhancement according to claim 1, characterized in that: The loss function used in the training of the neural network model for the reverse design of the series thin film is: Among them, R ij is the jth data value of the i-th group of reflectance spectrum in the training data set, r ij It is the jth data value of the reflectance spectrum output by the forward single fully connected neural network module after the film thickness parameters of the i-th group of reflectance spectra in the training data set are output by the reverse hybrid neural network module; i = 1, 2, 3...N, j = 1, 2, 3...n; N is the number of film structure groups, and n is the number of sampling points for each spectrum.

8. A film reverse design system, characterized in that: include: A data preprocessing module, configured to construct a training data set as claimed in claim 2; A model building module, used to build the tandem thin film reverse design neural network model according to claim 1; A model training module, configured to execute the training process as claimed in claim 1; The reverse design module is used to input the target spectrum and output the film thickness data.

9. An optical film designed by the method according to any one of claims 1 to 7, characterized in that: The film thickness parameters are determined by the output of the trained inverse hybrid neural network module.

10. A method for preparing an optical film, characterized in that: include: Determine the film thickness parameter using the method described in any one of claims 1 to 7; Layers of thin film materials are sequentially deposited on the substrate according to the parameters.

Images