How to predict LED structure performance

Abstract

The present invention relates to the technical field of semiconductor electronic devices, and provides a method for predicting the performance of an LED structure, which mainly involves collecting and extracting the input characteristic parameters and output characteristic parameters of an LED structure, constructing a corresponding data set, preprocessing the data in the data set according to a known standard, constructing a model using a machine learning algorithm, and performing structure parameter setting and initialization training for the model, and using the preprocessed data set to perform training optimization for the model after the structure parameter initialization training, thereby obtaining a prediction model, inputting the test data of the input characteristic parameters of the LED structure to be predicted into the prediction model, and obtaining the predicted value of the output characteristic parameters of the LED structure to be predicted, thereby quickly predicting the performance of the LED structure, with a fast prediction time and high prediction accuracy.

Description

[Technical Field]

[0001] The present invention relates to the technical field of semiconductor electronic devices, and more particularly to a method for predicting LED (Light Emitting Diode) structure performance using machine learning algorithm models. [Background technology]

[0002] Light-emitting diodes (LEDs) are characterized by high efficiency, energy saving, environmental protection, and long life, and are widely used in many fields such as traffic signs, architectural decoration, display lighting, etc. Among them, semiconductor materials such as InGaN and GaN have been rapidly developed and are being rapidly commercialized.

[0003] Since the establishment of the "Research on GaN-based Materials and Blue-Green Light Devices" project in 1994, InGaN and GaN-based LEDs have been widely applied in fields such as general lighting, LCD backlighting, outdoor displays, landscape lighting, and automotive lighting. GaN-based LED products are considered promising in the field of solid-state lighting, and are a highly energy-efficient alternative to incandescent lamps and fluorescent lamps.

[0004] The structural design of high-performance LEDs is usually carried out using a trial-and-error approach, in which the results of optimizing LED performance are compared with previous simulations and experimental results from the perspectives of material synthesis development, new structural designs, and new manufacturing technologies. Device performance optimization generally takes a long time and consumes a large amount of resources, such as time, materials, equipment, and manpower.

[0005] Machine learning encompasses probabilistic, statistical, approximation theory, and complex algorithmic knowledge. It is a cross-disciplinary approach that uses computers to simulate real-time human learning patterns and breaks down existing content into knowledge structures to effectively improve learning efficiency. Machine learning is the science that studies how computers can be used to simulate or realize human learning activities. It possesses the most intelligent characteristics within artificial intelligence and is one of the earliest research fields. Machine learning is a collaborative research hotspot in the fields of artificial intelligence and pattern recognition, and its theories and methods are widely used to solve complex problems in engineering and science. In today's era of rapid development of internet technology, applying machine learning techniques in artificial intelligence to LED structural design could potentially achieve geometrically doubled efficiency improvements.

[0006] In the design of high-performance LED structures, one of the challenges that those skilled in the art must proactively address is how to use machine learning techniques to predict the performance of the designed LED structure and use the prediction results to timely adjust the design proposal for the LED structure to obtain more efficient electronic devices. Summary of the Invention [Problem to be solved by the invention]

[0007] In order to solve the above-mentioned shortcomings of the prior art in predicting the structural performance of high-performance LEDs in structural design, the present invention provides a method for predicting the structural performance of LEDs. The method uses different algorithm models in machine learning (such as neural network models, decision trees, MLPs, etc.) to realize the prediction of the structural performance of high-performance LEDs, and timely adjusts the design proposal of the LED structure based on the indications of the prediction results, so that the structural design of the high-performance LED has better overall light-emitting performance. [Means for solving the problem]

[0008] In one embodiment, a method for predicting performance of an LED structure includes the steps of: (S1) collecting and extracting data of input feature parameters and corresponding output feature parameters of an LED structure, and dividing the data into a raw data set and a predicted data set; (S2) preprocessing the raw data set and the predicted data set to obtain a preprocessed raw data set and a preprocessed predicted data set; (S3) constructing an initial model using a machine learning algorithm; (S4) setting structural parameters for the initial model and performing initialization training on the structural parameters to obtain an initialized model; (S5) optimizing the initialized model and training the initialized model using the preprocessed raw data set to obtain corresponding network weights and biases, thereby obtaining a prediction model; and (S6) inputting the preprocessed test data set of input feature parameters of the LED structure to be predicted into the prediction model, and obtaining predicted values of the output feature parameters of the LED structure to be predicted.

[0009] In one embodiment, the input characteristic parameters of the LED structure include the structure, composition, and content of the barrier layers and potential well layers of the quantum well region in the LED structure, and the structure, composition, and content of the electron blocking layer, and the predicted values of the output characteristic parameters of the LED structure include the internal quantum efficiency (IQE) of the LED structure, the optical output power and its corresponding current density, IQE Droop, peak current density, etc.

[0010] In one embodiment, the machine learning algorithm may include, but is not limited to, one of a deep learning algorithm, a multi-layer perceptron (MLP), a decision tree, a linear regression, and a gradient boosting regression (GBR), where the deep learning algorithm may include, but is not limited to, one of a convolutional neural network (CNN), a recurrent neural network (KNN), an autoencoder, and a deep belief network (DBN).

[0011] In one embodiment, in the process of predicting the performance of an LED structure using the prediction method, the LED structure includes, but is not limited to, an InGaN-based visible LED, an AlGaN-based deep ultraviolet LED, a GaAs-based LED, a GaAlAs-based LED, and a GaP-based LED, and the LED structure described in the present invention is an LED structure including a PN junction and a quantum well layer.

[0012] Since the LED structure is more complex than other photovoltaic device structures, the selection of the structure characteristic parameters should be based on the importance of the structure characteristics and actual requirements. That is, the data of the input characteristic parameters and corresponding output characteristic parameters of the LED structure can be selected and adjusted according to different types of the LED structure. In other words, the input characteristic parameters and output characteristic parameters of the selected LED structure can be deleted or amplified as needed.

[0013] In one embodiment, the method for preprocessing the raw data set and the predicted data set may include a feature selection step (1) of selecting input feature parameters for the LED structure based on known physical knowledge and relationships between data, a data processing step (2) of performing a normalization process on the selected feature data, and a data recombination step (3) of recombining the size of the processed feature data.

[0014] In one embodiment, after the selected feature data is normalized, the data average value of the feature parameters is 0 and the standard deviation is 1.

[0015] In one embodiment, when optimizing the initialized model, the mean square error is adopted to judge the training result of the initialized model optimization, such as the goodness or prediction accuracy of the neural network model. The mean square error formula is:

number

[0016] In one embodiment, the neural network model in the deep learning algorithm is a convolutional neural network model, and the convolutional neural network model includes: an input layer for inputting test data of input feature parameters of the LED structure; a plurality of convolutional layers connected to the input layer for performing feature extraction on the test data input to the input layer, and connected to a plurality of fully connected layers after processing data output from the convolutional layer, with neurons installed in the plurality of fully connected layers for making predictions; and an output layer connected to the fully connected layer for outputting predicted values of output feature parameters of the LED structure.

[0017] In one embodiment, when performing parameter setting and parameter initialization training for the convolutional neural network model, the convolutional neural network model includes a first convolutional layer, a second convolutional layer, a first fully connected layer, and a second fully connected layer in that order, the first convolutional layer and the second convolutional layer have the same arrangement but a different total number of kernels, and the first fully connected layer employs a dropout strategy to reduce overfitting; initializing weights in the first convolutional layer and the second convolutional layer to block Gaussian distribution noise and initialize biases in the network to be constant; setting a learning rate within a certain numerical interval based on characteristics of training samples and determining a batch size of the training samples; and iteratively training the convolutional neural network model based on the setting of the training samples, determining the total number of rounds of the iterative training, and further optimizing the convolutional neural network model.

[0018] In one embodiment, when initializing the weights in the first convolutional layer and the second convolutional layer, normally distributed noise is blocked with an average value of 0 and a standard deviation of 0.1, the constant for bias initialization in the network is 1, the numerical range of the learning rate is 0.00001 to 0.1, and the total number of rounds of the iterative training is 100 to 500 rounds. [Effects of the Invention]

[0019] From the above, the LED structure performance prediction method provided by the present invention has the following advantages compared to conventional simulation software such as APSYS. 1. The method of predicting the performance of LED structures using different algorithm models in machine learning can quickly predict the performance of LED devices with different structures, regardless of whether the network structure fitting converges within this model, and can better guide the optimization of LED structure design proposals based on this prediction result. 2. The neural network model used in the machine learning algorithm of the present invention can use strategies such as Dropout to effectively prevent or reduce overfitting of the constructed neural network model, and further improve the accuracy of the constructed neural network model in predicting the performance of high-performance LED structures. 3. In the present invention, after performing machine learning on big data, a corresponding neural network model is constructed, and the neural network model is used to predict the performance of the overall structural material devices of high-performance LEDs with different structures, which allows the relatively complex physical rules of the overall structure of high-performance LEDs to be explored from the data perspective, and is easy to operate.

[0020] Other features and advantages of the present invention will be set forth in the description that follows, and in part will be obvious from the description, or may be learned by the practice of the invention. The objectives and other advantageous advantages of the present invention may be attained by the structure particularly pointed out in the description, claims, and drawings. [Brief explanation of the drawings]

[0021] In order to more clearly explain the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces drawings necessary for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without any creative effort. In the following description, the positional relationships described in the drawings shall be based on the directions indicated by the components shown in the drawings, unless otherwise specified.

[0022] [Figure 1] 1 is a structural schematic diagram of an embodiment of an LED according to the present invention. [Figure 2] FIG. 1 is a schematic diagram of a classical neural network model in the present invention. [Figure 3] 1 is a flowchart of an embodiment of a method for predicting LED structure performance in the present invention. [Figure 4]FIG. 1 is a structural schematic diagram of a convolutional neural network used for predicting LED structure performance in the present invention. [Figure 5] FIG. 1 is a structural schematic diagram of an embodiment of a convolutional neural network model according to the present invention. [Figure 6] FIG. 1 is a comparison of the internal quantum efficiency predicted by the convolutional neural network model of the present invention and the APSYS simulation internal quantum efficiency. [Figure 7] FIG. 1 is a comparison diagram of the internal quantum efficiency predicted by the multilayer perceptron model of the present invention and the APSYS simulation internal quantum efficiency. [Figure 8] FIG. 2 is a schematic diagram of a partial structure of a decision tree used for predicting LED structure performance in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0023] In order to clarify the objectives, technical solutions and advantages of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, but not all of the embodiments. The technical features designed in the various embodiments of the present invention described below can be combined with each other as long as they are not contradictory to each other. All other embodiments that a person skilled in the art can think of based on the embodiments of the present invention without making inventive efforts fall within the protection scope of the present invention.

[0024] In describing the present invention, it should be noted that all terms (including technical and scientific terms) used in the present invention have the same meaning as commonly understood by a person of ordinary skill in the art to which the present invention belongs, and should not be understood as limiting the present invention. It should be further understood that the terms used in the present invention should be understood to have a meaning consistent with their meaning in the context of this specification and in the related art, and should not be understood as an ideal or overly formal meaning unless clearly defined in the present invention.

[0025] GaN-based LEDs have a 30-year history of growth in the design of high-performance LED structures, with ample theoretical and data support. For ease of explanation and understanding, the present invention will be described in detail with reference to an example of a GaN-based LED structure, illustrating a method for predicting LED structure performance using different algorithmic models in machine learning. However, the present invention is not limited thereto, and the LED structure performance prediction method provided by the present invention can be applied to predict the performance of various LED structures, such as InGaN-based visible LEDs, AlGaN-based deep-ultraviolet LEDs, GaAs-based LEDs, GaAlAs-based LEDs, and GaP-based LEDs. The LED structure described in the present invention is an LED structure including a PN junction and a quantum well layer.

[0026] Referring to FIG. 1, FIG. 1 is a structural schematic diagram of one embodiment of an LED according to the present invention. As shown in the figure, the LED structure is a GaN-based LED, and the GaN-based LED structure type is an LED including a pn junction (or pn junction). The overall structure of such a GaN-based LED mainly comprises a substrate 10, an undoped GaN layer 11 (undropped GaN, u-GaN), an n-type layer 12, a multiple quantum well layer 13 (MQWs), an electron blocking layer 14 (Electron Blocking Layer, EBL), a p-type layer 15, a p-electrode 16 (P-contact), and an n-electrode 17 (N-contact), which are stacked in this order on the substrate. As shown in FIG. 1, in this example, the substrate 10 is a sapphire substrate, the n-type layer 12 is an n-type doped GaN layer (n-GaN), and the p-type layer 15 is a p-type doped GaN layer (p-GaN). The quantum well layer 13 (MQW) includes alternatingly grown InGaN and GaN layers. In other words, the quantum well layer 13 is a composite light-emitting region consisting of multiple periods of InGaN / GaN multiple quantum wells. The undoped GaN layer 11 is interposed between the substrate 10 and the n-type doped GaN layer and functions as a buffer layer. A p-electrode 16 (P-contact) and an n-electrode 17 (N-contact) are formed on both ends of the p-type layer 15 and the n-type layer 12, respectively.

[0027] As shown in Figure 1, the GaN-based LED in this example is a GaN-based blue LED that employs an InGaN / GaN multiple quantum well structure. This GaN-based blue LED structure can improve the luminous efficiency of the chip, and electronic devices manufactured using this structure have the advantages of high brightness and high optical efficiency. In this GaN-based blue LED structure, factors that significantly affect the luminous efficiency of the electronic device include, but are not limited to, the multiple quantum well layer 13 (MQWs) and the electron blocking layer 14 (EBL). The structures of the multiple quantum well layer 13 (MQWs) and the electron blocking layer 14 (EBL) are diverse, and their combinations are also diverse. To improve the luminous efficiency of the electronic device, it is necessary to pay attention to the structures of the multiple quantum well layer 13 (MQWs) and the electron blocking layer 14 (EBL) during the structural design of the GaN-based blue LED and to predict the structural performance of the multiple quantum well layer 13 (MQWs) and the electron blocking layer 14 with greater efficiency and accuracy.

[0028] As research and development progresses, a wide variety of machine learning techniques have been published, allowing for various classification methods based on the emphasis. Based on the classification of learning strategies, machine learning can be divided into machine learning that mimics the human brain and machine learning that directly uses mathematical techniques. Machine learning that mimics the human brain can be further divided into symbolic learning and neural network learning (or federated learning). Machine learning that directly uses mathematical techniques mainly includes statistical machine learning. Based on the classification of learning methods, machine learning can be divided into inductive learning, deductive learning, analogical learning, and analytical learning. Inductive learning can be further divided into symbolic inductive learning (e.g., learning by example, decision tree learning) and functional inductive learning (also known as discovery learning, including neural network learning, learning by example, discovery learning, and statistical learning). Based on the classification of learning methods, machine learning can be divided into supervised learning, unsupervised learning, and reinforcement learning (boosting learning). Machine learning can be divided into structured learning and unstructured learning based on the classification of data format, and into concept learning, rule learning, function learning, class learning, and Bayesian network learning based on the classification of the learning target.

[0029] Relatively commonly used algorithms in machine learning include, but are not limited to, decision tree algorithms, naive Bayesian algorithms, support vector machine algorithms, random forest algorithms, artificial neural network algorithms, boosting and bagging algorithms, association rule algorithms, EM (maximum likelihood estimation) algorithms, and deep learning. Deep learning (DL) is a new research direction in the field of machine learning (ML) that can learn the inherent regularities and representation hierarchies of sample data. The ultimate goal of deep learning is to enable machines to have human-like analytical learning capabilities and recognize data such as text, images, and voice.

[0030] Different deep learning models are primarily based on neural networks. A neural network is an algorithmic mathematical model that simulates the behavioral characteristics of biological neural networks and can receive multiple inputs and generate outputs. As neural networks continue to develop and deep learning algorithms are repeatedly updated, the structure of network models is also continuously adjusted and optimized, with greater room for improvement in feature extraction and feature selection techniques in particular. It can map any complex nonlinear relationship and has very strong robustness, memory capacity, self-learning capabilities, etc., making it widely used in classification, prediction, pattern recognition, etc.

[0031] Based on the above, in the examples of the present invention, the adopted machine learning algorithm may be a deep learning algorithm, a multi-layer perceptron, a decision tree, a linear regression, or a gradient boosting regression algorithm, where the deep learning algorithm may be a convolutional neural network, a recurrent neural network, an autoencoding, or a deep belief network.

[0032] Referring to Figure 2, Figure 2 is a schematic diagram of a classical neural network model in the present invention. A neural network (NN) is an algorithmic mathematical model that processes information by simulating an actual human neural network. It is a complex network system formed by connecting a large number of simple processing units (called neurons), and is a highly complex nonlinear dynamic learning system. A neuron is a multi-input, single-output information processing unit that processes information nonlinearly. A neural network is one of the models in machine learning.

[0033] Similar to actual human neural networks, neural networks are composed of neurons and nodes with connections (synapses) between them. Each neural network unit, also known as a perceptron, receives multiple inputs and generates an output. Actual neural network decision-making models are often multilayer networks composed of multiple perceptrons. As shown in Figure 2, a classic neural network model mainly consists of an input layer, an implicit layer, and an output layer. In the example shown in the figure, Layer L1 represents the input layer, Layer L2 represents the implicit or hidden layer, and Layer L3 represents the output layer. Taking advantage of the advantages of neural network models in machine learning, a neural network model can be used in the design process of the GaN-based blue LED structure shown in Figure 1 to relatively efficiently and accurately predict the characteristic factors of the LED structure that affect the luminous efficiency of the electronic device. Based on the prediction results, the design proposal for the GaN-based blue LED structure can be timely adjusted to achieve the desired overall functional effect. The characteristic factors include, but are not limited to, the structural performance of the multiple quantum well layer 13 (MQWs) and the electron blocking layer 14 (EBL).

[0034] As shown in Figure 2, in a feedforward neural network, information flows in only one direction: forward from the input node through hidden nodes (if any) to the output node. There are no cycles or loops within the network. In the field of machine learning, representative deep learning models mainly include three types: (1) neural network systems based on convolutional operations, i.e., convolutional neural networks (CNNs); (2) autoencoding neural networks based on multi-layer neurons, including autoencoders and sparse coding; and (3) deep belief networks (DBNs), which use pre-training in the form of multi-layer autoencoding neural networks and further optimize the neural network weights in conjunction with discriminative information.

[0035] Deep learning models include not only the three representative deep learning models mentioned above, but also recurrent neural networks, recursive neural networks, etc.

[0036] A convolutional neural network (CNN) is a deep feedforward neural network with properties such as local connectivity and weight sharing. It consists of three parts: the first part is an input layer; the second part is a combination of n convolutional layers and a pooling layer (also known as a hidden layer or implicit layer); and the third part is a fully connected multilayer perceptron classifier (also known as a fully connected layer). A convolutional neural network includes a feature extractor consisting of a convolutional layer and a subsampling layer. In a convolutional layer of a convolutional neural network, a neuron is only connected to a subset of neurons in adjacent layers. A convolutional layer of a CNN typically contains several feature planes (FeatureMaps), each consisting of several rectangularly arranged neurons. Neurons in the same feature plane share weights, which are called convolutional kernels. The convolutional kernels are typically initialized as random fractional matrices and learn to obtain reasonable weights during network training. Shared weights (convolution kernels) have the direct benefit of reducing the risk of overfitting while reducing the connections between layers of the network. Subsampling, also known as pooling, typically comes in two forms: mean pooling and max pooling. Subsampling layers, also known as pooling layers, perform feature selection and reduce the number of features, thereby reducing the number of parameters. Subsampling can be considered a specific convolution process. Convolution and subsampling significantly simplify the model complexity and reduce the model parameters.

[0037] 3 in conjunction with FIG. 1, which is a flowchart of an embodiment of a method for predicting LED structure performance in the present invention. Taking the GaN-based LED structure shown in FIG. 1 as an example, the procedure for predicting LED structure performance using a machine learning algorithm model is as follows: S01: Collect data of input characteristic parameters and corresponding output characteristic parameters of a GaN-based blue LED multiple quantum well structure to construct corresponding raw data sets and predicted data sets, pre-process the raw data sets and predicted data sets, and obtain pre-processed raw data sets and pre-processed test data sets; S02: Build an initial machine learning model, S03: Setting network structure parameters for the constructed initial model, and performing initialization training for the set network structure parameters to obtain an initialized model; S04: Train and optimize the initialization model using the preprocessed raw dataset to obtain a predictive model; S05: The preprocessed test data set is input into the prediction model, and the predicted values of the performance parameters of the GaN-based blue LED multiple quantum well structure are output.

[0038] In one embodiment, the training and prediction process of the machine learning predictive model includes the steps of: selecting feature variables as model input data for a raw dataset (S11); preprocessing the raw dataset and performing data normalization (S12); extracting sample data from the preprocessed raw dataset and dividing the data into batches (S13); and performing multiple rounds of training on the batch-divided preprocessed raw dataset to output prediction results (S14). In one embodiment, when configuring the predictive model, the parameters of the predictive model include, but are not limited to, the number of convolution kernels, the convolution kernel length, and an activation function.

[0039] Specifically, the following will explain how to use machine learning models to predict LED structure performance through the procedure of predicting GaN-based LED structure performance using different machine learning algorithm models.

[0040] Example 1: Using deep learning algorithm models in machine learning to predict the performance of LED structures.

[0041] 1 to 3, and also FIG. 4 and FIG. 5. FIG. 4 is a structural schematic diagram of a convolutional neural network used to predict the performance of an LED structure in the present invention, and FIG. 5 is a structural schematic diagram of an embodiment of a convolutional neural network model in the present invention. The example in FIG. 5 further explains and interprets a neural network model constructed using a convolutional neural network (CNN) as an example. The procedure for predicting the performance of a GaN-based blue LED multiple quantum well structure using a deep learning convolutional neural network algorithm model is as follows.

[0042] Step S1: Collect and extract data of input characteristic parameters and corresponding output characteristic parameters of a GaN-based blue LED multiple quantum well structure, and construct a corresponding data set for the collected data.

[0043] In the process of collecting and extracting input characteristic parameters for a GaN-based blue LED multiple quantum well structure, it is necessary to select characteristic parameters for the GaN-based blue LED multiple quantum well structure. This selection process mainly involves data collection and extraction or selection of input characteristic parameters that have a significant impact on the predicted values of the output characteristic parameters of the GaN-based blue LED multiple quantum well structure. The selected input characteristic parameters for the GaN-based blue LED multiple quantum well structure include, but are not limited to, the structure, composition, and content of the barrier layer and potential well layer in the quantum well region, as well as the structure, composition, and content of the electron blocking layer. The predicted values of the output characteristic parameters for the selected GaN-based blue LED multiple quantum well structure include, but are not limited to, the internal quantum efficiency (IQE) of the LED structure, the optical output power and its corresponding current density, IQE Droop, peak current density, etc.

[0044] Next, a corresponding dataset is constructed for these selected input feature parameters, and corresponding dataset parameters are designed. The dataset is divided into a raw dataset and a test dataset, and the dataset is preprocessed. The dataset parameters can indicate that the quantum well region or the electron blocking layer region adopts a complex structure such as a superlattice structure or a composition gradient, thereby allowing a large amount of data on LED structures to be collected and recorded. This allows data per LED structure to be one sample, and data on multiple LED structures to be one sample set. Each sample or each sample set can be used as an input layer in a neural network.

[0045] Step S2: Pre-process the data in the dataset constructed in step S1 to obtain a pre-processed raw dataset and a pre-processed test dataset. The pre-processing method includes the following steps: (1) In the constructed dataset, input feature parameters of the LED structure are selected and ranked according to their influence on the physical quantity that needs to be predicted, such as the internal quantum phase rate, based on known physical knowledge and the correlation between each data. (2) The selected feature data is subjected to normalization processing. The specific calculation formula is x' = (x - μ) / σ, where μ is the mean value of the sample and σ is the standard deviation of the sample. By performing normalization processing on the data, the mean value of the input data becomes 0 and the standard deviation becomes 1 in each dimension, and the data follows the standard normal distribution. (3) Data recombination: recombining the size of the processed data and dividing it into multiple batches.

[0046] Note that the input dimensions required for a 2D convolutional neural network are 4D (samples, rows, cols, channels), so in this example, the raw data is an array read in as a txt file. Therefore, the arrangement of the raw data needs to be adjusted to match the input size of the 2D convolutional neural network.

[0047] Step S3: Build a convolutional neural network model based on a deep learning algorithm in machine learning.

[0048] 4, the convolutional neural network model structure constructed using the convolutional neural network in this example mainly includes an input layer, multiple convolutional layers, multiple fully connected layers, and an output layer, with each layer being connected in sequence. The input layer can be used to input data or a dataset of input feature parameters of the sample or sample set of the above-mentioned LED structure, such as a preprocessed raw dataset or a preprocessed test dataset.

[0049] In the example shown in this figure, the convolutional neural network model structure has an input layer followed by two convolutional layers, each of which contains an activation function.

[0050] The convolution layer extracts the feature map by convolution calculation of the correspondence between the convolution kernel and the feature map. The specific process of the convolution of the first convolution layer is as follows: (l) = Σx (l-1) * ω (l) +b (l) where * represents the matrix convolution calculation and ω (l) represents the neuron weight of the lth layer, and b (l) represents the bias of the lth layer. Generally, if the input matrix size is ω, the convolution kernel size is k, the step size is k, and the number of zero-padding layers is p, the formula for calculating the size of the feature map generated after convolution is:

number

[0051] The activation function in the convolutional layer is a linear rectifier function (Relu). The mathematical formula for the linear rectifier function is f(x)=max(0,x), which allows for nonlinear transformation of feature maps.

[0052] The second convolutional layer performs further feature extraction of the image through the activation function, and the image output from the second convolutional layer is activated by the linear rectification function (Relu) in the activation layer and then transmitted to the next part, such as the fully connected layer.

[0053] Referring to FIG. 5 in conjunction with FIG. 4, in the example of FIG. 5, the convolutional neural network model structure has two fully connected layers. The images activated by the activation layer undergo a flattening process (960) and are connected to the first fully connected layer, which has 128 neurons. Considering the relatively small number of training samples used in this example, a dropout strategy is adopted to randomly deactivate 20% of the neurons per training round to suppress or reduce overfitting. The final prediction is completed using the second fully connected layer for the images that have passed through the activation function.

[0054] Step S4: Set network structure parameters for the constructed convolutional neural network model, and perform initialization training on the set network structure parameters to obtain an initialized convolutional neural network model.

[0055] The method for performing initialization training on the network structure parameters set in the convolutional neural network model involves setting the step size of the first convolutional layer to 1, the number of output channels to 16, and the same padding mode. The step size of the second convolutional layer is set to 1, the number of output channels to 32, and the same padding mode. The weights in the first and second convolutional layers are initialized to truncated normally distributed noise with a mean of 0 and a standard deviation of 0.1, and all biases in the network are initialized to a constant of 1. The initialization training of the convolutional neural network model is completed by setting a learning rate within a certain numerical range based on the characteristics of the training samples, determining a batch size of the training samples, repeatedly training the convolutional neural network model based on the training sample settings, and determining the total number of turns of the repeated training. Here, the learning rate is set to a numerical range of 0.00001 to 0.1, and the total number of rounds of the repeated training is 100 to 500.

[0056] To further explain, in the example shown in the figure, the learning rate of the training samples is set to 0.0001. The batch size of the training samples is set to 16, that is, 16 images are input to the convolutional neural network for each training, and the average loss of the samples across all batches is calculated. The total number of training rounds is 300, and the constructed convolutional neural network model is initially optimized using the stochastic gradient descent (SGD) algorithm.

[0057] Step S5: Using the dataset of input feature parameters of the LED structure preprocessed in step S2, train and optimize the initialized convolutional neural network model, obtain and store the network weights and biases of the convolutional neural network model, and further obtain a convolutional neural network prediction model, where the dataset is the preprocessed raw dataset.

[0058] In machine learning, a loss function is used to measure the loss (gap) between the model output value and the target value. Based on this, in step S5, the loss function during training of the convolutional neural network model is expressed using mean squared error to clarify the quality of the training results of this convolutional neural network model. This mean squared error formula is

number

[0059] Step S6: The preprocessed test data set of input feature parameters of the GaN-based LED structure to be predicted is input to the convolutional neural network prediction model as an input layer, thereby outputting predicted values of output feature parameters of the GaN-based LED structure to be predicted, including but not limited to the internal quantum efficiency (IQE) of the LED structure, the optical output power, and the corresponding current density.

[0060] The prediction model was built and trained using the Python platform. The convolutional neural network prediction model was trained and predicted using Python, and a comparison was obtained between the internal quantum efficiency predicted by the convolutional neural network model and the internal quantum efficiency simulated by APSYS, as shown in Figure 6. As can be seen from Figure 6, the predicted IQE (internal quantum efficiency of the LED structure) was 0.2422%, which was nearly identical to the actual value and within a small error range.

[0061] Example 2: Using a multi-layer perceptron model in machine learning to predict the performance of LED structures.

[0062] A multilayer perceptron (MLP) is a feedforward artificial neural network model that maps multiple input data sets to a single output data set.

[0063] In conjunction with Figure 2, Figure 2 may also be a partial structural schematic diagram of a multilayer perceptron used in predicting the performance of a GaN-based LED structure in the present invention. In fact, in the example of Figure 2, using 30 neurons and 10 hidden layers, the procedure for predicting the performance of a GaN-based blue LED multiple quantum well structure using a multilayer perceptron algorithm in machine learning is as follows:

[0064] Step S1: Collect and extract data of input characteristic parameters and corresponding output characteristic parameters of a GaN-based blue LED multiple quantum well structure, and construct a corresponding data set for the collected data.

[0065] In the process of collecting and extracting input characteristic parameters for a GaN-based blue LED multiple quantum well structure, it is necessary to select characteristic parameters for the GaN-based blue LED multiple quantum well structure. This selection process mainly involves data collection and extraction or selection of input characteristic parameters that have a significant impact on the predicted values of the output characteristic parameters of the GaN-based blue LED multiple quantum well structure. The selected input characteristic parameters for the GaN-based blue LED multiple quantum well structure include, but are not limited to, the structure, composition, and content of the barrier layer and potential well layer in the quantum well region, as well as the structure, composition, and content of the electron blocking layer. The predicted values of the output characteristic parameters for the selected GaN-based blue LED multiple quantum well structure include, but are not limited to, the internal quantum efficiency (IQE) of the LED structure, the optical output power and its corresponding current density, IQE Droop, peak current density, etc.

[0066] Next, a corresponding dataset is constructed for these selected input feature parameters, and corresponding dataset parameters are designed. The dataset is divided into a raw dataset and a test dataset, and the dataset is preprocessed. The dataset parameters can indicate that the quantum well region or the electron blocking layer region adopts a complex structure such as a superlattice structure or a composition gradient, thereby allowing a large amount of data on LED structures to be collected and recorded. This allows data per LED structure to be one sample, and data on multiple LED structures to be one sample set. Each sample or each sample set can be an input layer in a multi-layer neuron.

[0067] Step S2: Pre-process the data in the dataset constructed in step S1 to obtain a pre-processed raw dataset and a pre-processed test dataset. The pre-processing method includes the following steps: (1) In the constructed dataset, input feature parameters of the LED structure are selected based on known physical knowledge and correlation coefficients between each data. (2) The selected feature data is subjected to normalization processing. The specific calculation formula is x' = (x - μ) / σ, where μ is the mean value of the sample and σ is the standard deviation of the sample. By performing normalization processing on the data, the mean value of the input data becomes 0 and the standard deviation becomes 1 in each dimension, and the data follows the standard normal distribution. (3) Data recombination: recombining the size of the processed data and dividing it into multiple batches.

[0068] In this example, the raw data is an array read in as a txt file. Therefore, the arrangement of the raw data needs to be adjusted to match the input size of the multi-layer neuron.

[0069] Step S3: Construct a multi-layer neuron model based on the machine learning algorithm.

[0070] 2, in this example, the multi-layer neuron model structure mainly includes an input layer, multiple hidden layers, and an output layer, and each layer is connected in sequence. The input layer can be used to input data or data sets of input feature parameters of the sample or sample set of the LED structure.

[0071] In the example of Figure 2, in the multi-layer neuron model structure, one hidden layer is connected after the input layer and contains an activation function. The activation function in the hidden layer is a linear rectifier function (Relu). The number of nodes in the hidden layer is 10, and it is connected to the output layer.

[0072] Step S4: Set network structure parameters for the constructed multi-layer neuron model, and perform initialization training for the set network structure parameters to obtain an initialized multi-layer neuron model.

[0073] In a method for performing initialization training on network structural parameters set in a multi-layer neuron model, the weights in the hidden layer are initialized to truncated normally distributed noise with a mean value of 0 and a standard deviation of 0.1, and all biases in the network are initialized to a constant value of 1. The learning rate is set within a certain numerical range based on the characteristics of the training samples, a batch size of the training samples is determined, and the convolutional neural network model is repeatedly trained based on the training sample settings. The total number of rounds of the repeated training is determined, thereby completing the initialization training of the convolutional neural network model. Here, the learning rate is set within a numerical range of 0.00001 to 0.1, and the total number of rounds of the repeated training is 100 to 500.

[0074] To further explain, in the example shown in the figure, the learning rate of the training samples is set to 0.0001. The batch size of the training samples is set to 16, that is, 16 images are input to the convolutional neural network for each training, and the average loss of the samples across all batches is calculated. The total number of training rounds is 300, and the constructed convolutional neural network model is initially optimized using the stochastic gradient descent (SGD) algorithm.

[0075] Step S5: Use the dataset of input feature parameters of the LED structure preprocessed in step S2 to train and optimize the initialized multi-layer neuron model, obtain and store the network weights and biases of the multi-layer neuron model, and further obtain a multi-layer neuron model, where the dataset is the preprocessed raw dataset.

[0076] The forward propagation process is X (l) =Y (l-1) W (l) +B (l) where W (l) denotes the weight matrix when the l-1th layer is mapped to the lth layer, and B (l) denotes the bias vector of the l-th layer. The activation function is Y (l) =max(0,x (l) The output feature parameters are obtained by forward propagation, and then the loss function is calculated, while the partial derivatives of each parameter are further optimized using the backward propagation algorithm.

[0077] In machine learning, a loss function is used to measure the loss (gap) between the model output value and the target value. Based on this, in step S5, the loss function during training of the multi-layer neuron model is expressed using mean squared error to clarify the quality of the training results of this convolutional neural network model. This mean squared error formula is

number

[0078] Step S6: The preprocessed test data set of input feature parameters of the GaN-based LED structure to be predicted is input to the multi-layer neuron model as an input layer, thereby outputting predicted values of output feature parameters of the GaN-based LED structure to be predicted, including but not limited to the internal quantum efficiency (IQE) of the LED structure, the optical output power, and the corresponding current density.

[0079] The prediction model was built and trained using the Python platform. The multi-layer neuron prediction model was trained and predicted using Python, and a comparison of the multi-layer neuron predicted IQE and the APSYS simulated IQE was obtained, as shown in Figure 7. As can be seen from Figure 7, the predicted IQE (internal quantum efficiency of the LED structure) was 0.7056%, which was nearly identical to the actual value and within a small error range.

[0080] Example 3: Using decision tree models in machine learning to predict the performance of LED structures.

[0081] A decision tree is a tree structure where each internal node represents a split of an attribute, each branch represents a classification output, and each leaf node represents a category.

[0082] Referring to Figure 8, Figure 8 is a partial structure schematic diagram of a decision tree used to predict the LED structure performance in the present invention. In the example of Figure 8, only two layers are actually shown, and a depth of 10 is actually adopted. The procedure for predicting the performance of a GaN-based blue LED multiple quantum well structure using a decision tree algorithm in machine learning is as follows.

[0083] Step S1: Collect and extract data of input characteristic parameters and corresponding output characteristic parameters of a GaN-based blue LED multiple quantum well structure, and construct a corresponding data set for the collected data.

[0084] In the process of collecting and extracting input characteristic parameters for a GaN-based blue LED multiple quantum well structure, it is necessary to select characteristic parameters for the GaN-based blue LED multiple quantum well structure. This selection process mainly involves data collection and extraction or selection of input characteristic parameters that have a significant impact on the predicted values of the output characteristic parameters of the GaN-based blue LED multiple quantum well structure. The selected input characteristic parameters for the GaN-based blue LED multiple quantum well structure include, but are not limited to, the structure, composition, and content of the barrier layer and potential well layer in the quantum well region, as well as the structure, composition, and content of the electron blocking layer. The predicted values of the output characteristic parameters for the selected GaN-based blue LED multiple quantum well structure include, but are not limited to, the internal quantum efficiency (IQE) of the LED structure, the optical output power and its corresponding current density, IQE Droop, peak current density, etc.

[0085] Next, a corresponding dataset is constructed for these selected input feature parameters, and corresponding dataset parameters are designed. The dataset is divided into a raw dataset and a test dataset, and the dataset is preprocessed. The dataset parameters can indicate that the quantum well region or the electron blocking layer region adopts a complex structure such as a superlattice structure or a composition gradient, thereby allowing a large amount of data on LED structures to be collected and recorded. This allows data per LED structure to be one sample, and data on multiple LED structures to be one sample set. Each sample or each sample set can be an input layer in a multi-layer neuron.

[0086] Step S2: Pre-process the data in the dataset constructed in step S1 to obtain a pre-processed raw dataset and a pre-processed test dataset. The pre-processing method includes the following steps: (1) In the constructed dataset, input feature parameters of the LED structure are selected based on known physical knowledge and correlation coefficients between each data. (2) By performing normalization data processing on the selected feature data, the average value of these data becomes 0 and the standard deviation becomes 1. (3) Data recombination: recombining the size of the processed data and dividing it into multiple batches.

[0087] In this example, the raw data is an array read in as a txt file. Therefore, the arrangement of the raw data needs to be adjusted to match the input size of the decision tree.

[0088] Step S3: Build a decision tree model based on a machine learning algorithm.

[0089] The maximum depth in a decision tree model is 10, the minimum number of samples required to split an internal node is 2, and the non-purity function is

number

[0090] Step S4: Hyperparameters are set for the constructed decision tree model to obtain an initialized decision tree model.

[0091] The maximum depth in a decision tree model is 10, the minimum number of samples required to split an internal node is 2, and the non-purity function is

number

[0092] Step S5: Using the dataset of input feature parameters of the LED structure preprocessed in step S2, train and optimize the initialized decision tree model, obtain and store the network weights and biases of the decision tree model, and further obtain a decision tree model, where the dataset is the preprocessed raw dataset.

[0093] In machine learning, a loss function is used to measure the loss (gap) between the model output value and the target value. Based on this, in step S5, the loss function during training of the decision tree model is expressed using mean squared error to clarify the quality of the training results of this convolutional neural network model. This mean squared error formula is:

number

[0094] Step S6: The preprocessed test data set of input feature parameters of the GaN-based LED structure to be predicted is input to the decision tree model as an input layer, thereby outputting predicted values of output feature parameters of the GaN-based LED structure to be predicted, including but not limited to the internal quantum efficiency (IQE) of the LED structure, the optical output power, and the corresponding current density.

[0095] As described above, the convolutional neural network model, multilayer perceptron prediction model, decision tree prediction model, etc. provided by the present invention can more accurately predict luminous efficiency parameters such as the internal quantum efficiency (IQE), optical output power, and corresponding current density of a GaN-based blue LED multiple quantum well structure during the overall structure design process of the LED structure compared to conventional techniques, and based on these prediction results, can better guide the optimization of the design proposal for a new GaN-based LED structure and further design a new GaN-based LED overall structure with desired luminous efficiency.In addition, the convolutional neural network prediction model provided by the present invention can also predict the predicted values of output parameters such as the laser and probe.

[0096] Furthermore, those skilled in the art will understand that although there are many problems in the prior art, each embodiment or technical solution of the present invention can be improved only in one or more aspects without simultaneously solving all of the technical problems listed in the prior art or background art. Those skilled in the art will understand that anything not recited in the claims should not be construed as limiting the scope of the claims.

[0097] Terms such as LED, GaN-based LED, machine learning, and neural network are frequently used in this application, but this does not exclude the possibility of using other terms. These terms are used only to more conveniently explain and interpret the essence of the present invention, and any additional restrictions should not be construed as being contrary to the spirit of the present invention. The terms "first," "second," and the like (if any) in the specification and claims of the embodiments of the present invention and the drawings are used to distinguish similar items and are not necessarily used to describe a specific order or chronological order.

[0098] Finally, the above-mentioned embodiments are only used to explain the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the above-mentioned embodiments, those skilled in the art may still modify the technical solutions described in the above-mentioned embodiments or replace part or all of the technical features with equivalents, and it will be understood that these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention. [Explanation of symbols]

[0099] 10-substrate, 11 - Undoped GaN layer, 12-n type layer, 13-Multiple quantum well layer, 14-electron blocking layer, 15-p type layer, 16-p electrode, 17-n electrode.

Claims

1. A method for predicting LED structure performance executed by a computer system, comprising: The computer collects and extracts data of input characteristic parameters and corresponding output characteristic parameters of the LED structure through a convolutional neural network model, and divides the data into a raw data set and a predicted data set; The computer pre-processes the raw data set and the predicted data set using a pre-stored data pre-processing program to obtain a pre-processed raw data set and a pre-processed predicted data set; The computer constructs an initial model using a machine learning algorithm; The computer performs structural parameter setting for the initial model and initialization training for the structural parameters to obtain an initialized model to support model construction and analysis; The computer optimizes the initialized model using an optimization algorithm, and trains the initialized model using the pre-processed raw data set using a model training module to obtain corresponding network weights and biases, thereby obtaining a predictive model; the computer inputs the pre-processed prediction data set of input characteristic parameters of the LED structure to be predicted into the prediction model, and obtains predicted values of output characteristic parameters of the LED structure to be predicted, thereby quickly predicting the performance of the LED structure; the input characteristic parameters of the LED structure include the structure, composition, and content of a barrier layer and a potential well layer in a quantum well region in the LED structure, and the structure, composition, and content of an electron blocking layer; and the corresponding output characteristic parameters include the internal quantum efficiency, optical output power, and corresponding current density of the LED structure.

2. 2. The method of claim 1, wherein the machine learning algorithm is at least one of a deep learning algorithm, a multi-layer perceptron, a decision tree, a linear regression, and a gradient boosting regression.

3. 3. The method of claim 2, wherein the deep learning algorithm is at least one of a convolutional neural network, a recurrent neural network, an autoencoding, and a deep belief network.

4. 2. The method for predicting LED structure performance according to claim 1, wherein the LED structure includes an InGaN-based visible LED, an AlGaN-based deep ultraviolet LED, a GaAs-based LED, a GaAlAs-based LED, and a GaP-based LED.

5. 2. The method for predicting LED structure performance according to claim 1, wherein the data of the input characteristic parameters and corresponding output characteristic parameters of the LED structure can be selected and adjusted according to the type of the LED structure.

6. The method for pre-processing the raw dataset and the predicted dataset includes: a feature selection step of selecting input feature parameters of the LED structure based on known physical knowledge and relationships between data; a data processing step of performing a normalization process on the selected feature data; 2. The method for predicting LED structure performance according to claim 1, further comprising: a data recombination step of recombining the size of the processed feature data.

7. 7. The method for predicting LED structure performance according to claim 6, wherein after normalization processing is performed on the selected feature data, the data average value of the feature parameters is 0 and the standard deviation is 1.

8. In the step of optimizing the initialized model, a mean square error is adopted to determine the training result of the initialized model, and the mean square error formula is: [Equation 1] 2. The method of claim 1, wherein Predicti and Actuali are the predicted value and the true value of the i-th sample, respectively.

Images