Method for measuring and calculating relation between micro-nano structure and spectral characteristics based on deep learning model
By employing a deep learning-based method for calculating the relationship between micro/nano structures and spectral features, and training the model using a high-throughput experimental dataset, the problem of spectral errors caused by process errors in micro/nano optical design is solved, achieving high-precision spectral prediction and design of micro/nano structures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI INSTITUTE OF TECHNICAL PHYSICS CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-01-13
- Publication Date
- 2026-05-08
AI Technical Summary
In existing technologies, the ideal structure of micro-nano optical design has errors in simulation and actual spectrum due to process errors during actual manufacturing, which cannot meet the requirements of high-precision design.
By constructing a method for measuring the relationship between micro/nano structures and spectral features based on a deep learning model, the deep learning model is trained using a high-throughput experimental dataset to establish a bidirectional mapping relationship between the geometric parameters of micropillar arrays and the reflectance spectrum, including a forward prediction model and a reverse prediction model, thereby achieving accurate prediction of micro/nano structures.
It achieves high-precision spectral prediction of micro-nano structures, improves the reliability and efficiency of design, overcomes the defects of local optimal design in traditional methods, and significantly improves the accuracy of spectral prediction and the speed of data acquisition.
Smart Images
Figure CN121997727A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of deep learning and micro / nano optics in artificial intelligence, specifically to a method for calculating the relationship between micro / nano structures and spectral characteristics based on a deep learning model. Background Technology
[0002] In the field of micro-nano optics, the ideal micro-nano structures designed cannot be perfectly fabricated; the actual fabrication process always introduces errors. For example, due to limitations in photolithography, right-angled structures in the design may have rounded corners during actual fabrication; during etching, the etching depth of dense arrays and sparse arrays may differ, and the angles of the etched sidewalls are difficult to make perfectly perpendicular. These process errors are factors that traditional simulation software cannot fully consider.
[0003] Previous micro-nano optical design often involved using a single structure, then performing parameter scanning in simulation software to obtain a large number of results, and finally selecting the geometry corresponding to the desired electromagnetic response. This method is time-consuming, and the parameters obtained are not necessarily optimal.
[0004] In recent years, deep learning has played an increasingly important role in the field of micro-nano optics. Deep learning possesses powerful nonlinear fitting and generalization capabilities, capable of extracting useful key information from a suitable amount of data. Once a deep learning model is fully trained, a bidirectional mapping can be achieved between the geometric parameter space and the electromagnetic response space of a micro-nano structure, with extremely fast prediction speeds. However, current datasets used for neural network training are almost entirely simulations of ideal models, failing to consider the influence of actual fabrication errors. The spectra predicted by the models correspond to the responses of perfect structures, deviating from the spectra of real devices. Therefore, there is an urgent need to develop a new method capable of training with actual experimental data and accurately predicting the spectra of micro-nano structures, including all non-ideal fabrication features, to meet the ever-increasing precision requirements for the spectral design of micro-nano structures in the field of micro-nano optics. Summary of the Invention
[0005] The purpose of this invention is to provide a method for calculating the relationship between micro / nano structures and spectral characteristics based on a deep learning model, which solves the technical problem of errors between simulated and actual spectra in the prior art.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A method for measuring the relationship between micro / nano structures and spectral features based on a deep learning model, the specific steps of which are as follows:
[0008] S1, Acquisition and partitioning of the dataset;
[0009] Based on the designed micropillar array geometric parameters, a high-throughput micropillar array containing one or more unit patterns was prepared, and reflectance spectral data corresponding to each set of micropillar array geometric parameters were collected one by one to construct an experimental dataset containing the micropillar array geometric parameters and their corresponding reflectance spectra.
[0010] The experimental dataset was then divided into a training set and a test set;
[0011] S2, Model Training;
[0012] A deep learning model is trained using a training set to establish a bidirectional mapping relationship between the geometric parameters of the micropillar array and the corresponding experimental reflectance spectra; the deep learning model includes a forward prediction model and a backward prediction model.
[0013] The forward prediction model of the deep learning model is used to predict the corresponding reflectance spectrum based on the input micropillar array geometric parameters, and the backward prediction model of the deep learning model is used to predict the corresponding micropillar array geometric parameters based on the input reflectance spectrum.
[0014] S3, positive spectral prediction;
[0015] The geometric parameters of the micropillar array in the test set are input into the trained forward prediction model, which outputs the predicted reflectance spectrum; the predicted reflectance spectrum is then compared with the corresponding reflectance spectrum in the test set.
[0016] S4, Inverse Structure Prediction: Input the reflectance spectrum of the test set into the trained inverse prediction model, output the predicted geometric parameters of the micropillar array, and compare the predicted geometric parameters of the micropillar array with the corresponding geometric parameters of the micropillar array in the test set.
[0017] The experimental dataset is derived from the high-throughput fabrication and spectroscopic measurements of micro / nano structure arrays under consistent process conditions, and includes process non-ideal features resulting from the fabrication process.
[0018] The micropillar array in S1 is a micropillar array containing ten cell types, different sizes, and different periodic parameters.
[0019] The specific steps for obtaining data in S1 are:
[0020] S101, photomask design; designing photomasks by drawing two-dimensional pattern arrays of different shapes, sizes and period parameters; integrating pattern arrays of various shapes on the photomask to process micropillar arrays corresponding to various shapes;
[0021] S102, Fabrication of the micropillar array: Using the photolithography mask designed above, a two-dimensional pattern is transferred to the surface of a semiconductor material through photolithography. Then, the semiconductor material outside the pattern area is etched away using inductively coupled plasma etching to form the corresponding micropillar array structure.
[0022] S103, Reflectance Spectroscopy Measurement: Place the micropillar array sample prepared in S102 into the micro Fourier transform infrared spectroscopy testing system, measure the reflectance spectra of all micropillar arrays in sequence, and record the obtained spectral curve data; combine the reflectance spectral data with the micropillar array geometric parameters in S101 to form an experimental dataset.
[0023] S104, Dataset Partitioning and Preprocessing: The experimental dataset obtained in S103 is divided into a training set and a test set using a random seed method, with 80% used as the training set and 20% used as the test set for model training and final performance evaluation.
[0024] The two-dimensional pattern units contained in the photomask in S101 include asymmetric parallelograms, asymmetric rounded parallelograms, asymmetric ellipses, parallelograms, rounded parallelograms, rectangles, ellipses, squares, squares with holes, and circles. By changing the size and period parameters of the pattern, the position of the optical resonance peak of the final fabricated micropillar array can be controlled. Changing the size and period parameters over a large range allows for adjustment of the position of the optical resonance peak over a wider spectral range, thereby expanding the coverage of the optical response in the spectral database. The semiconductor material used to fabricate the micropillar array in S102 is gallium arsenide.
[0025] In S104, all data undergoes standardization preprocessing, normalizing the mean to 0 and the standard deviation to 1 for each feature dimension, while maintaining the shape of the data distribution. The formula for standardization is: , where x is the original data value, μ is the mean of the feature in all samples, σ is the standard deviation of the feature, and z is the standardized output value.
[0026] The specific steps for training the S2 model are as follows:
[0027] S201, Forward prediction model training and spectral prediction; a forward prediction deep learning model is trained using the training set data processed in step S104; the forward prediction model adopts a multilayer perceptron neural network structure, the input of which is the geometric parameters of the micropillar array, and the output is the corresponding reflectance spectrum; the input geometric parameters include the shape type of the micropillar array, whether there are connecting lines, the number of units per period, and continuous parameters in terms of short side length, long side length, unit area, included angle, rotation angle, x-direction period, y-direction period, and duty cycle;
[0028] S202, Inverse prediction model training and geometric parameter inversion; The inverse prediction deep learning model is trained using the training set data processed in step S104; The inverse prediction model adopts a convolutional neural network structure. The input of the inverse prediction model is the reflectance spectrum, and the output is the corresponding micropillar array geometric parameters; The convolutional neural network extracts the feature patterns in the spectrum by performing convolution operations on the input data layer by layer; The inverse prediction task needs to predict discrete category parameters and continuous geometric parameters simultaneously; The classification task includes predicting three parameters: the shape type of the micropillar array unit, whether there are connecting lines, and the number of units per period; The regression task includes predicting the parameters of the short side length, long side length, unit area, included angle, rotation angle, x-direction period, y-direction period, and duty cycle of the micropillar array.
[0029] The S201 multilayer perceptron model consists of multiple fully connected layers, with all neurons in each layer fully connected to all neurons in the next layer. The model training uses mean squared error as the loss function, as shown in the following formula: Where n is the number of data points, Y i It is the true value of the i-th data point. is the predicted value of the i-th data point; the model parameters are optimized by minimizing the above loss function.
[0030] In S202, the back prediction model selects different loss functions for different tasks: the cross-entropy loss function is used for classification tasks, and the mean squared error loss function is used for regression tasks; the overall loss function of the back prediction model can be defined by the following formula: L total It is the weighted sum of the loss functions of all tasks; The weights of the loss function for continuous values; L cont It is a continuous value loss function used to measure the difference between the predicted continuous parameters and the true values. The weights of the loss function for the shape unit; L shape It is the cross-entropy loss function for shape unit types, used to measure the difference between the predicted unit type category and the true category; The weights of the loss function with and without connecting lines; L conn It is the cross-entropy loss function with and without connections, used to measure the difference between the predicted class with and without connections and the true class; The weights of the loss function for the number of units per cycle; L units It is the cross-entropy loss function for the number of units per cycle, used to measure the difference between the predicted number of units per cycle and the true number; by adjusting the above weights, the learning effect of the model on each task can be balanced.
[0031] In view of the above technical features, the present invention has the following beneficial effects: 1. The present invention trains a neural network model using high-throughput micro / nano structure experimental reflectance spectrum datasets, enabling forward prediction of the reflectance spectrum corresponding to the micro / nano structure and reverse prediction of the geometric parameters of the micro / nano structure corresponding to the target reflectance spectrum; 2. High prediction accuracy: For forward prediction, the multilayer perceptron neural network model constructed in this invention achieves a spectral prediction accuracy of over 99% using only a relatively small amount of 1840 sets of reflectance spectrum data, reaching a high level of prediction accuracy; 3. Strong inversion capability: For reverse prediction, the convolutional neural network model constructed in this invention can effectively extract features from high-dimensional spectral data, enabling the prediction of the geometric parameters of the micropillar array. The global optimization search in the parameter space overcomes the defect of traditional spectral inversion design being prone to getting trapped in local optima. The model achieves an accuracy of over 89% in classifying and predicting shape unit type, presence or absence of connecting lines, and number of periodic units. It also exhibits high correlation in predicting continuous geometric parameters, with a determination coefficient R² exceeding 0.89, significantly improving the reliability of micro / nano structure spectral inversion design. 4. Good process consistency: This invention uses a stepper lithography machine to obtain 2300 sets of micropillar arrays and their spectral data in a single fabrication experiment. All data are acquired at once under uniform process conditions, avoiding data deviation caused by multiple batch experiments. The resulting experimental spectral dataset has high consistency and reliability. 5. Fast data acquisition speed: Traditional numerical simulation for acquiring micro / nano structure spectral data takes an average of about 1 hour per spectrum. In contrast, this invention acquires a spectrum in about 15 seconds using automated measurement after sample preparation, completing the acquisition of all 2300 spectra within 10 hours. The acquisition speed of the experimental spectral dataset is approximately 240 times that of the simulation dataset, significantly improving data acquisition efficiency. Attached Figure Description
[0032] Figure 1 A schematic diagram of the overall process of this invention;
[0033] Figure 2 Schematic diagram of various two-dimensional graphic units contained in a gallium arsenide micropillar array;
[0034] Figure 3 A cross-sectional schematic diagram of a gallium arsenide micropillar array structure;
[0035] Figure 4 A comparison chart of the spectrum output by the positive prediction model and the actual measured spectrum;
[0036] Figure 5 The confusion matrix of shape unit type classification for the inverse prediction model;
[0037] Figure 6 The confusion matrix of the inverse prediction model for the presence or absence of connecting lines;
[0038] Figure 7 The classification confusion matrix of the inverse prediction model for the number of units per cycle;
[0039] Figure 8 The regression prediction results of the inverse prediction model for the periodic parameters in the x-direction;
[0040] Figure 9 The regression prediction results of the inverse prediction model for the periodic parameters in the y-direction;
[0041] Figure 10 The regression prediction results of the inverse prediction model for the short-side parameter;
[0042] Figure 11 The regression prediction results of the inverse prediction model for the included angle parameter;
[0043] Figure 12 The regression prediction results of the inverse prediction model for the area parameter;
[0044] Figure 13 The regression prediction results of the inverse prediction model for the duty cycle parameter;
[0045] Figure 14 The regression prediction results of the inverse prediction model for the long side parameter;
[0046] Figure 15 The regression prediction results of the inverse prediction model for the rotation angle parameter. Detailed Implementation
[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that these drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that some components well-known to those skilled in the art but not related to the main content of the present invention may be omitted in the drawings or description. Additionally, for ease of description, some components in the drawings may be omitted, enlarged, or reduced, but this does not represent the actual size or complete structure of the product.
[0048] A method for calculating the relationship between micro / nano structures and their spectral features based on a deep learning model is proposed. This method trains a deep learning model using a high-throughput reflectance spectroscopy experimental dataset, enabling the prediction of the spectral feature relationship between micro / nano structures and their geometric parameters. The high-throughput reflectance spectral data is obtained by fabricating a large-scale gallium arsenide micropillar array and acquiring the reflectance spectrum of each array individually. The overall flowchart is shown below. Figure 1 As shown, the specific steps are:
[0049] S1, Acquisition and partitioning of the dataset;
[0050] Based on the designed micropillar array geometric parameters, a high-throughput micropillar array containing one or more unit patterns was prepared, and reflectance spectral data corresponding to each set of micropillar array geometric parameters were collected one by one to construct an experimental dataset containing the micropillar array geometric parameters and their corresponding reflectance spectra.
[0051] The experimental dataset was then divided into a training set and a test set;
[0052] Preferably, the micropillar array is a micropillar array containing ten unit types, different sizes, and different periodic parameters.
[0053] The experimental dataset is derived from the high-throughput fabrication and spectral measurement of micro / nano structure arrays under consistent process conditions, and the reflectance spectroscopy experimental dataset includes process non-ideal features resulting from the fabrication process.
[0054] S2, Model Training;
[0055] A deep learning model is trained using a training set to establish a bidirectional mapping relationship between the geometric parameters of the micropillar array and the corresponding experimental reflectance spectra; the deep learning model includes a forward prediction model and a backward prediction model.
[0056] The forward prediction model of the deep learning model is used to predict the corresponding reflectance spectrum based on the input micropillar array geometric parameters, and the backward prediction model of the deep learning model is used to predict the corresponding micropillar array geometric parameters based on the input reflectance spectrum.
[0057] S3, positive spectral prediction;
[0058] The geometric parameters of the micropillar array in the test set are input into the trained forward prediction model, which outputs the predicted reflectance spectrum; the predicted reflectance spectrum is then compared with the corresponding reflectance spectrum in the test set.
[0059] S4, Inverse Structure Prediction: Input the reflectance spectrum of the test set into the trained inverse prediction model, output the predicted geometric parameters of the micropillar array, and compare the predicted geometric parameters of the micropillar array with the corresponding geometric parameters of the micropillar array in the test set.
[0060] The specific implementation examples are as follows:
[0061] The specific steps for obtaining data in S1 are:
[0062] S101, photomask design; designing photomasks by drawing two-dimensional pattern arrays of different shapes, sizes and period parameters; integrating pattern arrays of various shapes on the photomask to process micropillar arrays corresponding to various shapes;
[0063] Micropillar arrays of different shapes possess different optical resonance modes, thus enriching the variety of optical modes contained in spectral datasets. For example... Figure 2 As shown, the two-dimensional pattern units contained in the photomask include: asymmetric parallelograms, asymmetric rounded parallelograms, asymmetric ellipses, parallelograms, rounded parallelograms, rectangles, ellipses, squares, squares with holes, and circles. By changing the size and periodic parameters of the patterns, the position of the optical resonance peaks in the final fabricated micropillar array can be controlled. Changing the size and periodic parameters over a large range allows for adjustment of the optical resonance peak positions over a wider spectral range, thereby expanding the coverage of the optical response in the spectral database.
[0064] S102, fabrication of the micropillar array; gallium arsenide is preferred as the semiconductor material for fabricating the micropillar array. Using the photolithography mask designed above, the two-dimensional pattern is transferred to the surface of the gallium arsenide material through photolithography. Then, the gallium arsenide material outside the pattern area is etched away using inductively coupled plasma etching process to form the corresponding micropillar array structure.
[0065] The etching depth is 3.5 micrometers. To ensure the verticality of the etching on the sidewalls of the micropillars and to avoid excessive lateral etching, nitrogen gas is added to the etching gas to form a passivation layer on the sidewalls of the pillars. Figure 3 A cross-sectional schematic diagram of the gallium arsenide micropillar array device structure in this embodiment is shown.
[0066] S103, Reflectance Spectroscopy Measurement: The gallium arsenide micropillar array sample prepared in S102 is placed into a micro Fourier transform infrared spectroscopy testing system, and the reflectance spectra of all micropillar arrays are measured in sequence. The obtained spectral curve data are recorded. The reflectance spectral data and the geometric parameters of the micropillar array in S101 are combined to form an experimental dataset.
[0067] The micro-Fourier transform infrared spectroscopy testing system is equipped with an airtight enclosure. Nitrogen gas is slowly introduced before testing to stabilize the ambient humidity below 10%. First, the reflectance spectrum of a standard gold mirror is measured as a background reference signal. Then, an automated, high-precision moving stage is used to sequentially measure the reflectance spectrum of each micropillar array unit. After each measurement, a control script drives the stage to move, aligning the infrared light source spot with the center of the next micropillar array for measurement. The reflectance spectra of all 2300 micropillar arrays on the photomask are measured sequentially using the above steps, and the resulting spectral curves are recorded (each spectrum contains 1737 wavenumber sampling points, covering a wavenumber range of approximately 650–4000 cm⁻¹). -1 .
[0068] S104, Dataset Partitioning and Preprocessing: The experimental dataset containing all 2300 sets of spectral data obtained in S103 is divided into a training set and a test set using a random seed method. 80% (1840 sets) is used as the training set and 20% (460 sets) is used as the test set for model training and final performance evaluation.
[0069] All data undergoes standardization preprocessing to normalize the mean to 0 and the standard deviation to 1 for each feature dimension, while preserving the shape of the data distribution. The formula for standardization is as follows: , where x is the original data value, μ is the mean of the feature in all samples, σ is the standard deviation of the feature, and z is the standardized output value.
[0070] The specific steps for training the S2 model are as follows:
[0071] S201, Forward prediction model training and spectral prediction; The forward prediction deep learning model is trained using the training set data processed in step S104; The forward prediction model adopts a multilayer perceptron neural network structure, the input of which is the geometric parameters of the micropillar array, and the output is the corresponding reflectance spectrum; The input geometric parameters include the shape type of the micropillar array, whether there are connecting lines, the number of units in each period, and continuous parameters such as the short side length, long side length, unit area, included angle, rotation angle, x-direction period, y-direction period, and duty cycle;
[0072] Shape types are represented by integer codes 1 to 10, representing asymmetric parallelograms, asymmetric rounded parallelograms, asymmetric ellipses, parallelograms, rounded parallelograms, rectangles, ellipses, squares, squares with holes, and circles, respectively. The presence or absence of connecting lines is represented by Boolean values, with 1 indicating the presence of connecting lines and 0 indicating the absence of connecting lines. The number of units contained in each cycle is an integer, either 1 or 2. Continuous parameters such as short side length, long side length, unit area, included angle, rotation angle, x-direction cycle, y-direction cycle, and duty cycle are also included, and these continuous parameters are represented by 32-bit floating-point numbers.
[0073] A multilayer perceptron model consists of multiple fully connected layers, where all neurons in each layer are fully connected to all neurons in the next layer. Although the structure of a multilayer perceptron is relatively simple, given a sufficient number of neurons, it can fit arbitrarily complex nonlinear function relationships.
[0074] The model training uses mean squared error (MSE) as the loss function, and its formula is shown below: Where n is the number of data points, Y i It is the true value of the i-th data point. is the predicted value of the i-th data point; the model parameters are optimized by minimizing the above loss function.
[0075] After training, the performance of the forward model was evaluated using test set data. The geometric parameters of different micropillar arrays were input into the model to predict the corresponding experimental reflectance spectra. The results show that the model-predicted spectra are in high agreement with the actual measured spectra, such as... Figure 4 The image shows a comparison between the predicted spectral curve and the corresponding true spectrum of a representative micropillar array sample, demonstrating a very close match. This indicates that the experimental reflectance spectrum dataset constructed in this invention is of high quality, and the trained forward deep learning model possesses high spectral prediction accuracy.
[0076] S202, Training of the inverse prediction model and inversion of geometric parameters; The inverse prediction deep learning model is trained using the training set data processed in step S104; The inverse prediction model adopts a convolutional neural network structure. The input of the inverse prediction model is the reflectance spectrum, and the output is the corresponding micropillar array geometric parameters; The convolutional neural network extracts the feature patterns in the spectrum by performing convolution operations on the input data layer by layer. In photonics data processing, it is a suitable model choice for processing high-dimensional spectral data.
[0077] The inverse prediction task requires the simultaneous prediction of discrete class parameters and continuous geometric parameters, which is a typical multi-task learning problem. The classification task includes predicting three parameters: the shape type of the micropillar array cells, whether there are connecting lines, and the number of cells per period. The regression task includes predicting the short side length, long side length, cell area, included angle, rotation angle, x-direction period, y-direction period parameters, and duty cycle of the micropillar array.
[0078] Therefore, different loss functions are selected for different tasks: cross-entropy loss function is used for classification tasks, and mean squared error loss function is used for regression tasks; the overall loss function of the backpropagation model can be defined by the following formula: L total It is the weighted sum of the loss functions of all tasks; The weights of the loss function for continuous values; L cont It is a continuous value loss function used to measure the difference between the predicted continuous parameters and the true values. The weights of the loss function for the shape unit; L shape It is the cross-entropy loss function for shape unit types, used to measure the difference between the predicted unit type category and the true category; The weights of the loss function with and without connecting lines; L conn It is the cross-entropy loss function with and without connections, used to measure the difference between the predicted class with and without connections and the true class; The weights of the loss function for the number of units per cycle; L unitsIt is the cross-entropy loss function for the number of units per cycle, used to measure the difference between the predicted number of units per cycle and the true number; by adjusting the above weights, the learning effect of the model on each task can be balanced.
[0079] After the convolutional neural network model is trained, it is evaluated using test set data. The reflectance spectrum of the test set is input into the trained CNN model, and the output is the predicted geometric parameters of the micropillar array. For classification prediction results, Figures 5-7 The prediction results for three parameters in the test set—shape type, presence or absence of connecting lines, and number of units—are shown respectively: the overall accuracy rate for shape type prediction is 89.13%, the accuracy rate for presence or absence of connecting lines is 99.78%, and the accuracy rate for the number of units per cycle is 98.26%. For the regression prediction results of continuous geometric parameters, the coefficient of determination R² is used to evaluate the model performance. Figure 8-15 As shown, the R² values of the predicted values and the true values of the periodic parameters in the x-direction, y-direction, short side length, long side length, cell area, included angle, rotation angle, and duty cycle on the test set are all close to 1, indicating that the model has good accuracy and linear correlation in predicting geometric continuity parameters.
[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. All equivalent changes and modifications made within the scope of the claims of this invention should be considered within the technical scope of this invention.
Claims
1. A method for calculating the relationship between micro / nano structures and spectral features based on a deep learning model, characterized in that: The specific steps are as follows: S1, Acquisition and partitioning of the dataset; Based on the designed micropillar array geometric parameters, a high-throughput micropillar array containing one or more unit patterns was prepared, and reflectance spectral data corresponding to each set of micropillar array geometric parameters were collected one by one to construct an experimental dataset containing the micropillar array geometric parameters and their corresponding reflectance spectra. The experimental dataset was then divided into a training set and a test set; S2, Model Training; A deep learning model is trained using a training set to establish a bidirectional mapping relationship between the geometric parameters of the micropillar array and the corresponding experimental reflectance spectra; the deep learning model includes a forward prediction model and a backward prediction model. The forward prediction model of the deep learning model is used to predict the corresponding reflectance spectrum based on the input micropillar array geometric parameters, and the backward prediction model of the deep learning model is used to predict the corresponding micropillar array geometric parameters based on the input reflectance spectrum. S3, positive spectral prediction; The geometric parameters of the micropillar array in the test set are input into the trained forward prediction model, which outputs the predicted reflectance spectrum; the predicted reflectance spectrum is then compared with the corresponding reflectance spectrum in the test set. S4, Inverse Structure Prediction: Input the reflectance spectrum of the test set into the trained inverse prediction model, output the predicted geometric parameters of the micropillar array, and compare the predicted geometric parameters of the micropillar array with the corresponding geometric parameters of the micropillar array in the test set.
2. The calculation method as described in claim 1, characterized in that: The experimental dataset is derived from the high-throughput fabrication and spectroscopic measurements of micro / nano structure arrays under consistent process conditions, and includes process non-ideal features resulting from the fabrication process.
3. The calculation method as described in claim 1, characterized in that: The micropillar array in S1 is a micropillar array containing ten cell types, different sizes, and different periodic parameters.
4. The calculation method as described in claim 3, characterized in that: The specific steps for obtaining data in S1 are: S101, photomask design; designing photomasks by drawing two-dimensional pattern arrays of different shapes, sizes and period parameters; integrating pattern arrays of various shapes on the photomask to process micropillar arrays corresponding to various shapes; S102, Fabrication of the micropillar array: Using the photolithography mask designed above, a two-dimensional pattern is transferred to the surface of a semiconductor material through photolithography. Then, the semiconductor material outside the pattern area is etched away using inductively coupled plasma etching to form the corresponding micropillar array structure. S103, Reflectance Spectroscopy Measurement: Place the micropillar array sample prepared in S102 into the micro Fourier transform infrared spectroscopy testing system, measure the reflectance spectra of all micropillar arrays in sequence, and record the obtained spectral curve data. The reflectance spectral data and the geometric parameters of the S101 micropillar array were combined to form an experimental dataset; S104, Dataset Partitioning and Preprocessing: The experimental dataset obtained in S103 is divided into a training set and a test set using a random seed method, with 80% used as the training set and 20% used as the test set for model training and final performance evaluation.
5. The calculation method as described in claim 4, characterized in that: The two-dimensional pattern units contained in the photomask in S101 include asymmetric parallelograms, asymmetric rounded parallelograms, asymmetric ellipses, parallelograms, rounded parallelograms, rectangles, ellipses, squares, squares with holes, and circles. By changing the size and period parameters of the pattern, the position of the optical resonance peak of the final fabricated micropillar array can be controlled. Changing the size and period parameters over a large range allows for adjustment of the position of the optical resonance peak over a wider spectral range, thereby expanding the coverage of the optical response in the spectral database. The semiconductor material used to fabricate the micropillar array in S102 is gallium arsenide.
6. The calculation method as described in claim 4, characterized in that: In S104, all data undergoes standardization preprocessing, normalizing the mean to 0 and the standard deviation to 1 for each feature dimension, while maintaining the shape of the data distribution. The standardization calculation formula is as follows: , where x is the original data value, μ is the mean of the feature in all samples, σ is the standard deviation of the feature, and z is the standardized output value.
7. The calculation method as described in claim 4, characterized in that: The specific steps for training the S2 model are as follows: S201, Forward prediction model training and spectral prediction; a forward prediction deep learning model is trained using the training set data processed in step S104; the forward prediction model adopts a multilayer perceptron neural network structure, the input of which is the geometric parameters of the micropillar array, and the output is the corresponding reflectance spectrum; the input geometric parameters include the shape type of the micropillar array, whether there are connecting lines, the number of units per period, and continuous parameters in terms of short side length, long side length, unit area, included angle, rotation angle, x-direction period, y-direction period, and duty cycle; S202, Inverse prediction model training and geometric parameter inversion; The inverse prediction deep learning model is trained using the training set data processed in step S104; The inverse prediction model adopts a convolutional neural network structure. The input of the inverse prediction model is the reflectance spectrum, and the output is the corresponding micropillar array geometric parameters; The convolutional neural network extracts the feature patterns in the spectrum by performing convolution operations on the input data layer by layer; The inverse prediction task needs to predict discrete category parameters and continuous geometric parameters simultaneously; The classification task includes predicting three parameters: the shape type of the micropillar array unit, whether there are connecting lines, and the number of units per period; The regression task includes predicting the parameters of the short side length, long side length, unit area, included angle, rotation angle, x-direction period, y-direction period, and duty cycle of the micropillar array.
8. The calculation method as described in claim 7, characterized in that: The S201 multilayer perceptron model consists of multiple fully connected layers, with all neurons in each layer fully connected to all neurons in the next layer. The model training uses mean squared error as the loss function, as shown in the following formula: Where n is the number of data points, Y i It is the true value of the i-th data point. is the predicted value of the i-th data point; the model parameters are optimized by minimizing the above loss function.
9. The calculation method as described in claim 7, characterized in that: In S202, the back prediction model selects different loss functions for different tasks: the cross-entropy loss function is used for classification tasks, and the mean squared error loss function is used for regression tasks; the overall loss function of the back prediction model can be defined by the following formula: L total It is the weighted sum of the loss functions of all tasks; The weights of the loss function for continuous values; L cont It is a continuous value loss function used to measure the difference between the predicted continuous parameters and the true values. The weights of the loss function for the shape unit; L shape It is the cross-entropy loss function for shape unit types, used to measure the difference between the predicted unit type category and the true category; The weights of the loss function with and without connecting lines; L conn It is the cross-entropy loss function with and without connections, used to measure the difference between the predicted class with and without connections and the true class; The weights of the loss function for the number of units per cycle; L units It is the cross-entropy loss function for the number of units per cycle, used to measure the difference between the predicted number of units per cycle and the true number; by adjusting the above weights, the learning effect of the model on each task can be balanced.