CN-121997727-A - Measuring and calculating method for micro-nano structure and spectrum characteristic relation based on deep learning model

CN121997727ACN 121997727 ACN121997727 ACN 121997727ACN-121997727-A

Abstract

The invention discloses a measuring and calculating method of a micro-nano structure and spectrum characteristic relation based on a deep learning model, which comprises the steps of firstly preparing micro-nano structure arrays containing various unit types and different sizes and periods, measuring and collecting reflection spectrum data of each array in a high flux manner, and constructing a high-quality experimental spectrum data set; and then training a deep learning model by utilizing the experimental data set to realize the experimental reflection spectrum corresponding to the forward predicted micro-nano structure geometric parameter and the micro-nano structure geometric parameter and shape corresponding to the reverse predicted target reflection spectrum. The forward prediction adopts a multi-layer perceptron neural network model, and the reverse prediction adopts a convolution neural network model. The invention effectively makes up the difference between the simulated spectrum and the actual spectrum, and obviously improves the efficiency and the precision of the spectrum design of the micro-nano structure.

Inventors

LI TIANXIN
Xia Pengzhe
LU WEI
YU LI
Ma Qinxiao
XIN RUI

Assignees

中国科学院上海技术物理研究所

Dates

Publication Date: 20260508
Application Date: 20260113

Claims (9)

1. A measuring and calculating method of micro-nano structure and spectrum characteristic relation based on deep learning model is characterized in that the method comprises the following steps, S1, acquiring and dividing a data set; preparing a high-flux micro-column array comprising one or more unit patterns according to the designed micro-column array geometric parameters, collecting reflection spectrum data corresponding to each group of micro-column array geometric parameters one by one, and constructing an experimental data set comprising the micro-column array geometric parameters and reflection spectrums corresponding to the micro-column array geometric parameters; Dividing the experimental data set into a training set and a testing set; s2, model training; Training a deep learning model by utilizing a training set to establish a bidirectional mapping relation between geometric parameters of the micro-column array and corresponding experimental reflection spectrums, wherein the deep learning model comprises a forward prediction model and a reverse prediction model; The forward prediction model of the deep learning model is used for predicting the corresponding reflection spectrum based on the input micro-column array geometric parameters, and the reverse prediction model of the deep learning model is used for predicting the corresponding micro-column array geometric parameters based on the input reflection spectrum; S3, forward spectrum prediction; inputting the geometric parameters of the microcolumn array of the test set into a trained forward prediction model, and outputting a predicted reflection spectrum; S4, reverse structure prediction, namely inputting the reflection spectrum of the test set into a trained reverse prediction model, outputting predicted micro-column array geometric parameters, and comparing the predicted micro-column array geometric parameters with corresponding micro-column array geometric parameters of the test set.
2. The method of measuring according to claim 1, wherein the experimental data set is derived from high throughput preparation and spectroscopic measurement of micro-nanostructure arrays under consistent process conditions, and wherein the experimental data set comprises process non-ideal characteristics resulting from the preparation process.
3. The method of measuring according to claim 1, wherein the array of micropillars in S1 is an array of micropillars comprising ten cell types, different sizes and different cycle parameters.
4. The method for measuring and calculating according to claim 3, wherein the specific steps of data acquisition in S1 are as follows: S101, designing a photoetching plate; the method comprises the steps of designing a photoetching mask plate by drawing two-dimensional pattern arrays with different shapes, different sizes and different period parameters, integrating pattern arrays with various shapes on the photoetching mask plate, and processing micro-column arrays corresponding to the various shapes; S102, processing the micro-column array, namely transferring a two-dimensional pattern to the surface of a semiconductor material by adopting the photoetching mask plate designed in the above way through a photoetching process, and etching the semiconductor material outside a pattern area by utilizing an inductive coupling plasma etching process to form a corresponding micro-column array structure; S103, measuring a reflection spectrum, namely placing the microcolumn array sample prepared in the step S102 into a microscopic Fourier transform infrared spectrum test system, sequentially measuring the reflection spectrum of all microcolumn arrays, and recording the obtained spectrum curve data; S104, dividing and preprocessing the data set, and dividing the experimental data set obtained in S103 into a training set and a testing set by adopting a random seed method, wherein 80% of the training set is used as the training set, and 20% of the testing set is used as the testing set for model training and final performance evaluation.
5. The method of measuring and calculating according to claim 4, wherein the two-dimensional pattern unit shape included in the photolithography mask in S101 comprises an asymmetric parallelogram, an asymmetric rounded parallelogram, an asymmetric ellipse, a parallelogram, a rounded parallelogram, a rectangle, an ellipse, a square with holes and a circle, the position of an optical formant of the finally processed micro-column array can be regulated and controlled by changing the size and the period parameter of the pattern, the position of the optical formant can be adjusted in a wider spectral range by changing the size and the period parameter in a larger range, thereby expanding the coverage of optical response in a spectral database, and the semiconductor material of the processed micro-column array in S102 is gallium arsenide.
6. The measuring and calculating method according to claim 4, wherein the step S104 is characterized in that all data are subjected to standardized preprocessing, the data mean value of each characteristic dimension is normalized to 0, the standard deviation is normalized to 1, and the data distribution shape is ensured to be unchanged, and the calculation formula of the standardized processing is as follows: Where x is the raw data value, μ is the mean of the feature in all samples, σ is the standard deviation of the feature, and z is the normalized output value.
7. The method of measuring according to claim 4, wherein the model training of S2 is performed by the steps of, S201, training a forward prediction model and predicting a spectrum, training a forward prediction deep learning model by using training set data processed in the step S104, wherein the forward prediction model adopts a multi-layer perceptron neural network structure, is input into geometric parameters of a micro-column array and is output into corresponding reflection spectrums, and the input geometric parameters comprise the shape type of the micro-column array, the existence of connecting lines, the number of units in each period, and continuous parameters in the aspects of short side length, long side length, unit area, included angle, rotation angle, x-direction period, y-direction period and duty ratio; S202, training a reverse prediction model and inverting geometric parameters, training a reverse prediction depth learning model by using training set data processed in the step S104, wherein the reverse prediction model adopts a convolution neural network structure, the input of the reverse prediction model is a reflection spectrum, the output of the reverse prediction model is a corresponding micro-column array geometric parameter, the convolution neural network extracts a characteristic mode in the spectrum by carrying out convolution operation on input data layer by layer, a reverse prediction task needs to simultaneously predict discrete category parameters and continuous geometric parameters, the classification task comprises three parameters of the shape type, the existence of connecting lines and the number of units in each period of the prediction micro-column array, and a regression task comprises the parameters of the short side length, the long side length, the unit area, the included angle, the corner, the x-direction period, the y-direction period and the duty ratio of the prediction micro-column array.
8. The method of measuring and calculating according to claim 7, wherein the multi-layer perceptron model of S201 is composed of a plurality of fully connected layers, all neurons of each layer are fully connected with all neurons of the next layer, and the model training adopts mean square error as a loss function, and the formula is as follows: Where n is the number of data points, Y i is the true value of the ith data point, Is the predicted value of the ith data point, and optimizing the model parameters by minimizing the loss function described above.
9. The method according to claim 7, wherein in S202, the inverse prediction model selects different loss functions for different tasks, the classification task uses cross entropy loss functions, the regression task uses mean square error loss functions, and the total loss functions of the inverse prediction model can be defined by the following formula: Where L total is the weighted sum of all task loss functions; And L cont is a continuous value loss function, which is used for measuring the difference between the predicted continuous parameter and the actual value. L shape is a cross entropy loss function of the shape unit type, used for measuring the difference between the predicted unit type category and the real category; L conn is a cross entropy loss function with or without connecting lines and is used for measuring the difference between the predicted category with or without connecting lines and the real category; Weights for a loss function for the number of units per cycle; L units is a cross entropy loss function of the number of units per cycle, and is used for measuring the difference between the predicted number of units per cycle and the real type, and the learning effect of the model on each task can be balanced by adjusting the weight.

Description

Measuring and calculating method for micro-nano structure and spectrum characteristic relation based on deep learning model Technical Field The invention relates to the technical field of deep learning and micro-nano optics in artificial intelligence, in particular to a measuring and calculating method of a micro-nano structure and spectrum characteristic relation based on a deep learning model. Background For the micro-nano optical field, the designed ideal micro-nano structure cannot be completely and perfectly manufactured, and errors are always brought in the actual processing process. For example, due to the limitation of photolithography process, the right-angle structure in the design is rounded in actual preparation, and in the etching process, the etching depth of the dense array and the etching depth of the sparse array may be different, and the angle of the etching sidewall is difficult to be completely vertical. These process errors are all factors that are difficult to fully account for by conventional simulation software. Previous micro-nano optical designs often use a structure, then perform parameter scanning in simulation software to obtain a large number of results, and then select the geometry corresponding to the desired electromagnetic response. This approach takes a lot of time and the resulting parameters are not necessarily optimal parameters. In recent years, deep learning plays an increasingly important role in the micro-nano optical field. Deep learning has strong nonlinear fitting and generalization capability, and useful key information can be mined in a proper amount of data. Once the deep learning model is fully trained, the geometric parameter space of the micro-nano structure and the electromagnetic response space thereof can be mapped in a bidirectional manner, and the prediction speed is extremely high. However, the data set used for training the neural network is almost all simulation data of an ideal model, the influence of actual machining errors is not considered, and the spectrum predicted by the model corresponds to the response of a perfect structure and deviates from the spectrum of a real device. Therefore, development of a new method capable of training and accurately predicting micro-nano structure spectrum including all non-ideal characteristics of the process by using practical experimental data is urgently needed to meet the increasing precision requirement of micro-nano optical field on micro-nano structure spectrum design. Disclosure of Invention The invention aims to provide a measuring and calculating method of a micro-nano structure and spectrum characteristic relation based on a deep learning model, which solves the technical problem that an error exists between a simulated spectrum and an actual spectrum in the prior art. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: A measuring and calculating method of micro-nano structure and spectrum characteristic relation based on deep learning model comprises the following steps, S1, acquiring and dividing a data set; preparing a high-flux micro-column array comprising one or more unit patterns according to the designed micro-column array geometric parameters, collecting reflection spectrum data corresponding to each group of micro-column array geometric parameters one by one, and constructing an experimental data set comprising the micro-column array geometric parameters and reflection spectrums corresponding to the micro-column array geometric parameters; Dividing the experimental data set into a training set and a testing set; s2, model training; Training a deep learning model by utilizing a training set to establish a bidirectional mapping relation between geometric parameters of the micro-column array and corresponding experimental reflection spectrums, wherein the deep learning model comprises a forward prediction model and a reverse prediction model; The forward prediction model of the deep learning model is used for predicting the corresponding reflection spectrum based on the input micro-column array geometric parameters, and the reverse prediction model of the deep learning model is used for predicting the corresponding micro-column array geometric parameters based on the input reflection spectrum; S3, forward spectrum prediction; inputting the geometric parameters of the microcolumn array of the test set into a trained forward prediction model, and outputting a predicted reflection spectrum; S4, reverse structure prediction, namely inputting the reflection spectrum of the test set into a trained reverse prediction model, outputting predicted micro-column array geometric parameters, and comparing the predicted micro-column array geometric parameters with corresponding micro-column array geometric parameters of the test set. The experimental data set is derived from high-throughput preparation and spectroscopic measurement of micro-nanostructure arra