A method for designing the structure of an LED phosphor layer based on deep learning

Abstract

The present invention provides a method for designing the structure of an LED phosphor layer based on deep learning, which includes two parts: a forward network and a reverse network. The forward network is used to establish the mapping relationship between the distribution characteristics of fluorescent particles (concentration, thickness, particle size) and the light output effect (color coordinates, luminous efficiency, contrast ratio); based on the above mapping relationship, the reverse network reversely outputs the corresponding distribution characteristics according to the given optical properties, which is used to guide the preparation process. The steps include S1 collecting and constructing a data set, which is divided into two parts: training and testing; S2 establishing a deep neural network model, including a forward network and a reverse network; S3 training and testing the network model, optimizing the network model according to the results to obtain a deep learning network model; S4 testing the training model to obtain the prediction results. The present invention can solve the problems of low design efficiency and high cost of the LED phosphor layer structure.

Description

Technical Field

[0001] The present invention belongs to the fields of deep learning and LED phosphor layer structure design, and relates to a method for designing an LED phosphor layer structure based on deep learning. Background Art

[0002] The luminescence process of fluorescent particles is complex, and the performance of the phosphor layer is jointly affected by multiple factors (concentration, thickness, particle size). At present, most high-quality light sources are achieved by experiments, with low efficiency and high production costs, and there is a lack of a suitable theoretical prediction model. Traditional optimization methods, such as parameter scanning methods and adjoint optimization methods, have problems of slow convergence speed and being easily trapped in local minima.

[0003] Deep learning is a classification under machine learning and has developed rapidly in recent years. Especially in the fields of machine vision and image processing, it has been widely applied. Up to now, various deep learning network models have been proposed in the academic community. The differences between different neural network models are manifested in the activation rules of neurons, the algorithms for parameter learning, and the topological structures of neural network models. Common deep learning network models include convolutional neural networks, generative adversarial networks, deep neural networks, recurrent neural networks, and deep belief networks, etc. Deep learning learns the input and output rules from data, which helps people establish non-analytical models in various applications. Deep Neural Networks (DNN) is a technology in the field of deep learning that can handle complex problems that require a high degree of scientific rigor and precision. DNN can extract the features hidden in the statistical laws of data and store them as network parameters through the learning process.

[0004] Aiming at the above problems, a DNN deep learning prediction model is established to predict the light-emitting effect of the phosphor layer under the influence of complex factors, shorten the cycle of model simulation, and improve the efficiency of model simulation. In addition, the reverse prediction of the deep learning model can also guide the preparation process of high-quality light sources. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a method for designing an LED phosphor layer structure based on deep learning, which can effectively solve the problems of low efficiency and large error of the prediction model in the existing technology. The deep learning network model is divided into a forward network model and a reverse network model. The forward network is used to establish the mapping relationship between the distribution characteristics of fluorescent particles (concentration, thickness, particle size) and the light-emitting effect (color coordinates, contrast, luminous efficiency); based on the above mapping relationship, the reverse network can inversely output the corresponding distribution characteristics according to the given optical properties, which is used to guide the preparation process. To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for designing the structure of an LED phosphor layer based on deep learning, comprising the following steps:

[0007] S1 Establish a simulation structure model of the LED phosphor layer and construct a data set of structural parameters and light output effects.

[0008] Step S1-1: Establish a simulation structure model in an optical system simulation software according to the geometric parameters of the LED phosphor layer structure.

[0009] Establish a simulation structure model of the LED phosphor layer and conduct numerical simulation. The overall structure model of the LED phosphor layer includes four parts: a substrate, a copper heat sink, an LED chip, and a phosphor layer. Import relevant material parameters into the software and build the structure, including the size and thickness of the substrate and the heat sink.

[0010] Step S1-2: Import relevant parameters of the phosphor in the phosphor layer, set the materials and corresponding parameters of each part in the LED phosphor layer structure, and import relevant refractive indices.

[0011] The fluorescence characteristics of the phosphor include fluorescence spectra (absorption spectrum, excitation spectrum, and emission spectrum), phosphor concentration, and quantum conversion efficiency, etc. The absorption spectrum, excitation spectrum, and emission spectrum of the phosphor are all preprocessed and then imported into LightTools. During the simulation, the phosphor concentration (unit: / mm 3 ) represents the mixing ratio of the phosphor and the encapsulation silica gel. The real part of the refractive index of the phosphor is set to 1.8, and the imaginary part is set to 0.0006.

[0012] Step S1-3: Set the surface of the substrate as specular reflection, set the radiation power of the LED chip, the central wavelength, and the maximum line width of the radiation source. The quasi-parallel light emitted by the blue light chip is incident on the phosphor layer placed parallel to the chip. Create a new material silica in the user materials, set the material type to homogeneous, and set the refractive index of the silica to 1.41.

[0013] The simulation uses a simplified form of the optical model. The chip size is set to 0.875×0.55×0.2mm, the upper surface of the chip emits light, and the quasi-parallel light emitted by the LED chip is incident on the phosphor layer. The size of the phosphor layer is large enough relative to the chip size to prevent light from exiting from the side of the phosphor layer. A spherical far-field receiver is used to record the light intensity of the forward emission and backward scattering of the phosphor layer. Place the spherical far-field receiver above the LED phosphor layer model, and record the light intensity of the forward emission and backward scattering of the phosphor layer through the receiver.

[0014] Step S1-4: Adjust the number of ray tracings and start the simulation to obtain the light output effect of the phosphor layer.

[0015] Set the number of ray tracing to 5,000,000. The accuracy of the calculation results is affected by the total number of rays. To ensure the accuracy of the statistical results during the ray tracing process, the number of rays for each calculation during the simulation is controlled at around 5,000,000.

[0016] Step S1-5: Set different structural parameters, simulate the model multiple times, record the structure of the model and the corresponding light output effects, and complete the construction of the data set.

[0017] The thickness t, concentration d, and particle size r of the phosphor are important factors affecting the light output effect of the model. By changing the structural parameters of the model and simulating multiple times, the effects of changes in the thickness of the phosphor, the concentration of the phosphor layer, and the distribution of the phosphor particle size on the light output effect of the LED phosphor layer model are tested. The light output effect parameters include color coordinates (x, y), contrast c, and luminous efficacy LE. The color coordinates include two sets of data, x and y. The average value of x and y is selected to represent the overall light output color of the model. The contrast c is used to measure the uniformity of the light output of the model. The calculation method is to divide the difference between the maximum and minimum gray values in the illuminance map by the sum of the maximum and minimum values. The smaller the contrast, the more uniform the light output effect of the model. The luminous efficacy LE is selected to measure the luminous efficiency of the model.

[0018] S2 Establish a deep learning forward network model.

[0019] First, establish a deep learning forward network model based on the fully connected layer. The fully connected layer is a typical structure of deep learning. By extracting the internal laws and representation levels of the learning sample data, the rapid calculation of different structural parameters of the phosphor layer and the corresponding light output effects is realized.

[0020] Step S2-1: Preprocess the data set.

[0021] The preprocessing is a normalization process. The normalization process uses the Z-score normalization algorithm. The mathematical description of Z-Score is that for the sequence t1, t2, t3, …, t n Perform the following transformation:

[0022]

[0023] In the formula, The newly generated sequence z1, z2, z3, …, z n Conforms to a distribution with a mean of 0 and a variance of 1, and has no dimension.

[0024] Step S2-2: Establish a forward network. Read the dataset established in S1, and divide the dataset into two parts, a training set and a test set, according to a certain proportion. Among them, the training set accounts for 70%, and the test set accounts for 30%. The training set is used to train the neural network, and the test set is used to detect the training effect of the neural network. The input data x of the network is the phosphor layer structure in the dataset, and the output data y is the light output effect of the phosphor layer.

[0025] Step S2-3: Based on the fully connected layer, establish the main structure of the forward network. It includes an input layer, several hidden layers, and an output layer, and each layer includes several nodes. Data is transmitted from the nodes of the input layer into the neural network and is output to the next node in sequence. Determine the number of nodes in the input layer and the output layer, the number of hidden layers, and the threshold of the mean square error based on the structural parameters of the phosphor layer in the dataset and the number of data of the light output effect.

[0026] Train the forward network model obtained in Step 2 and test the forward network model.

[0027] Step S3-1: Use the training set to train the forward neural network and calculate the mean square error E of the network on the training set train as the training Loss function

[0028]

[0029] where n is the number of samples for each training, is the result predicted for the i-th sample point, and y i is the true value corresponding to the i-th sample point. Multiple iterations are performed during training, and the mean square error of each iteration is calculated.

[0030] Step S3-2: Use the test set to test the network during each iterative training and calculate the mean square error E of the network on the test set test as the test Loss function

[0031]

[0032] where m is the number of samples for each training, is the result predicted for the j-th sample point, and j is the true value corresponding to the j-th sample point. Multiple iterations are performed during testing along with training, and the mean square error of each iteration is calculated.

[0033] Collect the E train and E te,t calculated for each iteration, and draw an image with the number of iterations as the horizontal axis and the Loss function value as the vertical axis, as the training result of the network.

[0034] S4 Optimize the structure of the above network model according to the training results to obtain an optimal deep learning network model.

[0035] Optimize the deep learning network model according to the convergence of the Loss function, and balance the relationship between the network volume and the data fitting ability of the network. When the Loss function converges and is lower than 0.1, reduce the number of hidden layers or the number of nodes in each layer, so as to speed up the calculation speed of the deep learning model and reduce the time cost; when the Loss function fluctuates above 0.1 and the convergence effect is poor, increase the hidden layer or the number of nodes in each layer, so as to enhance the fitting ability of the network.

[0036] S5 Establish a deep learning reverse network model.

[0037] The establishment of the reverse network is similar to that of the forward network. Repeat steps S2 to S4, change the input data x in step S2 to the light output effect of the phosphor layer, and the output y to the structural parameters of the phosphor layer. After training, optimization and testing, a reverse network is obtained.

[0038] So far, the establishment of the forward network model and the reverse network model in the present invention is completed. By inputting the structural parameters (concentration, thickness, particle size) of the phosphor layer in the forward network, the corresponding light output effect (color coordinates, luminous efficiency, contrast) can be obtained; by inputting the light output effect of the phosphor layer in the reverse network, the corresponding structural parameters of the phosphor layer can be obtained. The forward network establishes the mapping relationship between the fluorescence particle distribution characteristics and the light output effect; based on the above mapping relationship, the reverse network can reversely output the corresponding distribution characteristics according to the given optical properties, which is used to guide the preparation process of the phosphor layer and realize the structural design of the LED phosphor layer.

[0039] Compared with the traditional optimization algorithm, the beneficial effects of the present invention are:

[0040] (1) Based on the LED phosphor layer simulation model of the optical system modeling software LightTools, the present invention provides a data set containing different structural parameters of the phosphor layer and the light output effect of the phosphor layer.

[0041] (2) The present invention provides a prediction method for the light output effect and structural parameters of the LED phosphor layer model based on deep learning, which can quickly calculate the corresponding light output effect according to the structure of the phosphor layer, and can also perform high-speed optimization design on the structural parameters of the model according to the light output requirements of the LED phosphor layer.

[0042] (3) The present invention solves the problem of difficult establishment of the fluorescence particle prediction model, shortens the simulation period of the model, reduces the time cost, and improves the model simulation efficiency. Description of the Drawings

[0043] Figure 1This is the flowchart of the present invention.

[0044] Figure 2 This is a schematic diagram of the forward network and the reverse network of the present invention. Among them, (a) is the forward network and (b) is the reverse network.

[0045] Figure 3 This is a schematic diagram of the structure of the deep neural network of the present invention.

[0046] Figure 4 This is a schematic diagram of the training results of the deep learning network of the present invention. Among them, (a) is the change of the Loss value with the number of iterations during the training process, and (b) is the change of the Loss value with the number of iterations during the test process. Detailed implementation manners

[0047] The following further describes the present invention in detail in conjunction with embodiments and the accompanying drawings, but the implementation manners of the present invention are not limited thereto. The process of the invention is as Figure 1 shown.

[0048] A method for designing the structure of an LED phosphor layer based on deep learning includes the following steps:

[0049] S1 Establish a simulation structure model of the LED phosphor layer and construct a data set of structure parameters and light output effects.

[0050] The present invention selects LightTools software for simulation. For the actual extended light source, a collimation and homogenization optimization design is carried out, and the final achieved design effect is combined. When establishing the model, the chip size is set to 0.875×0.55×0.2 mm, the upper surface of the chip emits light, and the radiation power of the blue LED is set to 0.375 W; the central wavelength of the light source is set to 456 nm, and the maximum full width at half maximum is set to 19 nm; a new material silicone is created in the user materials, the material type is set to uniform, and the refractive index of the silicone is set to 1.41. The number of ray traces is set to 5,000,000. The accuracy of the calculation result is affected by the total number of rays. To ensure the accuracy of the statistical results during the ray tracing process, the number of rays calculated each time during the simulation is controlled at about 5,000,000.

[0051] When obtaining the data set, record the fluorescence particle distribution characteristics of the LED phosphor layer simulation model, including concentration, thickness, and the particle size of the phosphor particles; and record the corresponding light output effects of the phosphor, including color coordinates, contrast, and luminous efficiency.

[0052] The thickness of the phosphor layer affects the luminous flux. An increase in thickness will result in a lower luminous flux. This is mainly because as the thickness of the phosphor layer increases, the propagation distance of light in the phosphor layer increases, and the probability of light being reflected or scattered by the phosphor becomes larger, causing the light to be confined in the phosphor layer or emitted backward, reducing the forward emission of light energy. To achieve a better light output effect for the LED phosphor layer model, the thickness of the phosphor layer selected in this invention is from 0.5 mm to 1.2 mm.

[0053] When the thickness is fixed, the concentration of phosphor particles also affects the light output effect to a certain extent. At a low concentration, the luminous flux of the LED chip is the highest. Since the phosphor does not completely cover the LED chip, more blue light penetrates, resulting in an increase in the luminous flux of the LED chip. As the concentration gradually increases to the saturation state, the blue light energy is completely consumed. Continuing to increase the phosphor concentration cannot increase its excitation intensity. Instead, it causes a large amount of light emitted by the phosphor to be scattered and absorbed, increasing the probability of absorption loss and reducing the light efficiency. The particle concentration selected in this invention is from 5000 to 100000 per cubic millimeter.

[0054] The particle radius of the phosphor is also one of the important factors affecting the light output effect. When other factors remain unchanged, the light efficiency will change with the change of the particle radius. The range of the particle radius selected in this invention is from 1000 to 10000 nanometers.

[0055] In terms of the model simulation results, the color coordinates, contrast, and light efficiency are selected to measure the light output effect of the model. The color coordinates include two sets of data, x and y. The average value of x and y is selected to represent the overall light output color of the model. The contrast is used to measure the uniformity of the light output of the model. The calculation method is to divide the difference between the maximum and minimum gray values in the illuminance map by the sum of the maximum and minimum values. The smaller the contrast, the more uniform the light output effect of the model. In addition, the light efficiency is also selected to measure the luminous efficiency of the model.

[0056] Since only one set of data can be obtained in a single simulation after fixing the parameters in LightTools, the construction of the dataset takes a long time. To improve the efficiency, we use Matlab to call LightTools to record the simulation data and obtain the dataset. The calling program can also be further developed to achieve functions of large-scale data simulation, storage, and reading.

[0057] S2 Establish a deep learning forward network model.

[0058] Since each feature of the input has its own dimension, it is difficult to interpret when physical quantities with different dimensions are involved in the prediction calculation of the same deep learning model. Moreover, larger or smaller data can easily affect the calculation results of the deep learning model prediction, affect the generalization ability of the model, and make training difficult and not easy to converge. Therefore, the Z-Score standardization algorithm is used to convert the data into a distribution with a mean of 0 and a standard deviation of 1 according to the mean and standard deviation. The converted data is dimensionless. Read the data set established in S1, and divide the data set into two parts: a training set and a test set according to a certain proportion, where the training set is 70% and the test set is 30%. The training set is used to train the neural network, and the test set is used to detect the training effect of the neural network. The input data x of the forward network is the phosphor layer structure, and the output data y is the light output effect of the phosphor layer.

[0059] First, establish a forward network. The forward network is used to establish the mapping relationship between the fluorescence particle distribution characteristics and the light output effect. The input features of the forward network are the concentration, thickness, and particle size of the phosphor layer, and the output features are the light output effect of the phosphor layer, namely color coordinates, contrast, and light efficiency. According to the multiple regression of deep learning, the prediction of the light emission effect based on the phosphor layer structure is realized, which shortens the software simulation time and improves the simulation efficiency. The forward network is mainly composed of fully connected layers in structure. The neurons in each fully connected layer are connected to all the neurons in the previous layer, integrating the features of the input data, and using ReLU as the activation function. There are 7 fully connected layers in the forward network, as Figure 3 shown, used to fit the multiple regression mapping relationship between the phosphor layer structure and the light output effect.

[0060] Train the forward network model obtained in step 2 of S3 and test the forward network model.

[0061] First, train the forward network. During training, 2640 groups of data are used as the training data set. The Batch Size of the forward network is 256, the learning rate is 0.001, the number of iterations Epochs is 2000, and the Adam optimizer is used to optimize the mean square error Loss function to evaluate the training effect. During the training process, the mean square error Loss function is calculated according to formula (2) for each iteration.

[0062] The training results are as Figure 4 shown, where (a) shows the change of the Loss value with the number of iterations during training, and (b) shows the change of the Loss value with the number of iterations during testing. The mean square error of the data results decreases with the increase of the number of iterations and tends to be stable.

[0063] Optimize the network structure according to the training results to obtain the optimal deep learning network model.

[0064] Optimize the deep learning network model according to the training results and the convergence of the Loss function, and balance the relationship between the network volume and the data fitting ability of the network. Adjust the network structure according to the convergence of the Loss function. When the Loss function converges, reduce the number of hidden layers or the number of nodes in each layer, so as to speed up the calculation speed of the deep learning model and reduce the time cost. When the fluctuation of the Loss function is higher than 0.1 and the convergence effect is poor, increase the number of hidden layers or the number of nodes in each layer to enhance the fitting ability of the network.

[0065] S5 Establish a deep learning inverse network model.

[0066] Repeat steps S2 to S4 to establish an inverse network. Based on the above mapping relationship, the inverse network can inversely output the corresponding phosphor layer structure and distribution characteristics according to the given optical properties, and is used to guide the preparation process. The input of the inverse network is the light output effect of the phosphor layer, that is, the color coordinates, contrast, and luminous efficiency, and the output is the structure of the phosphor layer, that is, the thickness of the phosphor layer, the concentration of the phosphor, and the particle size of the particles.

[0067] The structure of the inverse network is similar to that of the forward network and consists of 12 fully connected layers. The purpose of the inverse network is to inversely predict the structure of the phosphor layer according to the desired ideal light output effect, and to guide the preparation process of the phosphor layer with the result as the guide. The dataset, Batch Size, and learning rate of the inverse network are the same as those of the forward network, and the number of iterations Epochs is 3000. The Adam optimizer is used, and the mean square error is used as the Loss function during the training process to evaluate the training effect.

[0068] In the present invention, the forward network provides a method for predicting the light output effect and structural parameters of an LED phosphor layer model based on deep learning, and can quickly calculate the corresponding light output effect according to the phosphor layer structure. The inverse network can perform high-speed optimization design on the model structure parameters according to the light output requirements of the LED phosphor layer.

[0069] The above embodiments only represent the implementation manners of the present invention, but should not be construed as limiting the scope of the present invention patent. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for designing the structure of an LED phosphor layer based on deep learning, characterized in that It includes the following steps: S1: Establish a simulation structural model of the LED phosphor layer and construct a data set of structural parameters and light output effects; S2: Establish a deep learning forward network model; First, establish a deep learning forward network model based on the fully connected layer. The fully connected layer is a typical structure of deep learning. By extracting the internal laws and representation levels of the learning sample data, it realizes the rapid calculation of different structural parameters of the phosphor layer and the corresponding light output effects; Step S2-1: Preprocess the data set; Step S2-2: Establish a forward network; read the data set and divide the data set into two parts, a training set and a test set, according to a certain proportion; the training set is used to train the neural network, and the test set is used to detect the training effect of the neural network; the input data x of the network is the phosphor layer structure in the data set, and the output data y is the light output effect of the phosphor layer; Step S2-3: Based on the fully connected layer, establish the main structure of the forward network; it includes an input layer, several hidden layers, and an output layer, and each layer includes several nodes; the data is transmitted from the nodes of the input layer into the neural network and is output to the next node in turn; determine the number of nodes in the input layer and the output layer, the number of hidden layers, and the threshold of the mean square error based on the number of data of the structural parameters and light output effects of the phosphor layer in the data set; [[ID=I7]]S3: Train the forward network model obtained in step 2 and test the forward network model; Step S3-1: Train the forward neural network using the training set and calculate the mean square error E of the network on the training set train as the training Loss function; where n is the number of samples for each training, is the result predicted for the i-th sample point, y i is the true value corresponding to the i-th sample point; multiple iterations are performed during training, and the mean squared error of each iteration is calculated; Step S3-2: Use the test set to test the network in each iterative training, and calculate the mean square error E of the network on the test set t*st as the test Loss function; where m is the number of samples for each training, is the result predicted for the j-th sample point, and j is the true value corresponding to the j-th sample point; during testing, it is iterated multiple times with the training, and the mean squared error of each iteration is calculated; Collect the E calculated in each iteration train and E test , and plot an image with the number of iterations as the horizontal axis and the Loss function value as the vertical axis as the training result of the network; S4: Optimize the structure of the above network model according to the training results to obtain an optimal deep learning network model; S5: Establish a deep learning inverse network model; The establishment of the inverse network is similar to that of the forward network; repeat steps S2 to S4, change the input data x in step S2 to the light output effect of the phosphor layer, and change the output y to the structural parameters of the phosphor layer. After training, optimizing, and testing, an inverse network is obtained; So far, the establishment of the forward network model and the inverse network model is completed; by inputting the structural parameters of the phosphor layer into the forward network, the corresponding light output effect can be obtained; by inputting the light output effect of the phosphor layer into the inverse network, the corresponding structural parameters of the phosphor layer can be obtained, where the structural parameters include the thickness t, concentration d, and particle size r of the phosphor, and the light output effects include color coordinates, luminous efficiency, and contrast; the forward network establishes a mapping relationship between the fluorescence particle distribution characteristics and the light output effect; based on the above mapping relationship, the inverse network inversely outputs the corresponding distribution characteristics according to the given optical properties, which is used to guide the preparation process of the phosphor layer to realize the structural design of the LED phosphor layer.

2. The method for designing the LED phosphor layer structure based on deep learning according to claim 1, wherein Step S1 is specifically as follows: Step S1-1: In the optical system simulation software, establish a simulation structural model according to the geometric parameters of the LED phosphor layer structure; Establish a simulation structural model of the LED phosphor layer and conduct numerical simulation; The overall structure model of the LED phosphor layer includes four parts: a substrate, a copper heat sink, an LED chip, and a phosphor layer. Import relevant material parameters and conduct structure construction, including the size and thickness of the substrate and the heat sink; Step S1-2: Import the relevant parameters of the phosphor in the phosphor layer, set the materials and corresponding parameters of each part in the LED phosphor layer structure, and import the relevant refractive index; The fluorescence characteristics of the phosphor include the fluorescence spectrum diagram, the phosphor concentration, and the quantum conversion efficiency. The absorption spectrum, excitation spectrum, and emission spectrum of the phosphor are all preprocessed and then imported into LightTools; during the simulation, the phosphor concentration represents the mixing ratio of the phosphor and the encapsulation silica gel. Step S1-3: Set the substrate surface to specular reflection, set the radiation power of the LED chip, the central wavelength of the radiation source, and the maximum line width; the quasi-parallel light emitted by the blue chip is incident on the phosphor layer placed parallel to the chip; create a new material silica in the user materials, and set the material type to homogeneous. Use a simplified optical model for simulation. Place a spherical far-field receiver above the LED phosphor layer model, and record the light intensity of the forward emission and backward scattering of the phosphor layer through the receiver. Step S1-4: Adjust the number of ray tracings and start the simulation to obtain the light output effect of the phosphor layer. Step S1-5: Set different structural parameters, where the structural parameters include the thickness t, concentration d, and particle size r of the phosphor. Conduct multiple simulations on the model, record the structure of the model and the corresponding light output effect, and complete the construction of the dataset.

3. The method for designing the LED phosphor layer structure based on deep learning according to claim 2, wherein, The specific content of the above step S1-5 is as follows: By changing the structural parameters of the model and conducting multiple simulations to test the influence of the distribution changes of the phosphor thickness, the concentration of the phosphor layer, and the particle size of the phosphor on the light output effect of the LED phosphor layer model. The light output effect parameters include the color coordinates (x, y), the contrast ratio c, and the luminous efficiency LE; the color coordinates include two sets of data x and y, and the average value of x and y is selected to represent the overall light output color of the model; the contrast ratio c is used to measure the uniformity of the light output of the model, and the calculation method is to divide the difference between the maximum and minimum gray values in the illuminance map by the sum of the maximum and minimum gray values. The smaller the contrast ratio, the more uniform the light output effect of the model; the luminous efficiency LE is selected to measure the luminous efficiency of the model.

4. A method for designing the structure of an LED phosphor layer based on deep learning according to claim 1, characterized in that, In the step S2-1 described above, the preprocessing is normalization processing; the normalization processing uses the Z-score normalization algorithm; the mathematical description of Z-Score is that for the sequence t1, t2, t3, …, t n perform the following transformation: wherein, the newly generated sequences z1, z2, z3, …, z n conform to a distribution with a mean of 0 and a variance of 1, and have no dimension.

5. A method for designing the structure of an LED phosphor layer based on deep learning according to claim 1, characterized in that, The specific content of the above step S4 is as follows: Optimize the deep learning network model according to the convergence of the Loss function, and balance the relationship between the network volume and the data fitting ability of the network; when the Loss function converges and is lower than 0.1, reduce the number of hidden layers or the number of nodes in each layer to accelerate the calculation speed of the deep learning model and reduce the time cost. When the Loss function fluctuates above 0.1 and the convergence effect is poor, increase the number of hidden layers or the number of nodes in each layer to enhance the fitting ability of the network.

Images