A Spectral Reconstruction Optimization Method for a Micro-Spectrometer

Abstract

The present invention belongs to the technical field of photodetectors, and specifically relates to a method for optimizing spectral reconstruction of a micro-spectrometer. The present invention is based on a micro-spectrometer that only includes a photodetector array with different light response wavelengths, and uses a machine learning algorithm to reconstruct an unknown spectral curve. The process of reconstructing the spectrum by the machine learning algorithm is as follows: first, a database composed of the light response current of the device and spectral characteristic values is established, then a machine learning algorithm model framework is built, a mapping relationship between the light response current and the spectrum is established through training, and finally, the unknown spectrum is reconstructed according to this mapping relationship and the light response current of the device to the unknown spectrum. In order to improve the spectral reconstruction accuracy, the light response current is extracted from multiple different voltages to increase the data volume in the database. Experiments show that the optimization method of the present invention can significantly improve the spectral reconstruction accuracy. The present invention can be applied to fields such as hyperspectral imaging and flexible electronics.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical detectors, and in particular relates to a spectrum reconstruction optimization method of a micro-spectrometer. Background Art

[0002] In recent years, micro-spectrometers based on spectral reconstruction have been widely researched. Two main approaches exist for achieving spectral reconstruction: mathematical modeling and machine learning algorithms. Mathematical modeling, which relies heavily on the number of devices, their performance, and test accuracy, results in poor error tolerance. Furthermore, achieving the required spectral reconstruction accuracy requires significant effort in device fabrication and performance improvement. With the advancement of computer technology, using software programming to compensate for hardware deficiencies has become a new trend.

[0003] Machine learning algorithms, with their ability to learn, analyze, predict, and provide feedback, have been applied to various fields, including image recognition, spectrum detection, and semiconductor bandgap prediction. Because machine learning algorithms can continuously reduce errors through iteration during training, they have also been gradually applied to spectral reconstruction in recent years. In 2020, Zhang et al. from Tsinghua University used machine learning algorithms to reconstruct the spectrum of a filter system based on 108 quantum dots within the wavelength range of 400 nm to 750 nm, achieving a resolution of 5 nm. However, this work required the preparation of 108 quantum dots with different bandgaps and the integration of additional charge-coupled devices (CCDs), increasing the complexity of the system and requiring further refinement and improvement. Therefore, fully leveraging the software advantages of machine learning algorithms and selecting appropriate optimization methods to reduce the requirements for the number and performance of photodetectors required for spectral reconstruction has become a research issue. Summary of the Invention

[0004] In order to solve the above problems, the object of the present invention is to provide a spectrum reconstruction optimization method for a micro-spectrometer, which can achieve high-precision spectrum reconstruction when the device distribution density is sparse.

[0005] The present invention provides a spectral reconstruction optimization method for a micro-spectrometer. This method utilizes a machine learning algorithm for spectral reconstruction and increases the amount of data in the machine learning algorithm's database to achieve high-precision spectral reconstruction. The micro-spectrometer may comprise a series of photodetectors exhibiting different photoresponse cutoff wavelengths, or a single photodetector whose photoresponse wavelength can be altered by adjusting bias stress or mechanical stress.

[0006] The spectrum reconstruction optimization method of the micro-spectrometer provided by the present invention comprises the following specific steps for spectrum reconstruction:

[0007] (1) Establish a database. The database contains two parts: spectral characteristic values and photoresponse currents of photodetectors. Spectral characteristic values are obtained by using Gaussian distribution function to de-peak the spectrum, and the half-peak width, peak position or peak height is obtained. Photoresponse current is the amount of data added to the database, which is extracted from multiple different voltages by each photodetector with different photoresponse cutoff wavelength;

[0008] (2) Establishing the mapping relationship between photoresponse current and spectrum by training the machine learning algorithm model;

[0009] (3) Input the photodetector's photoresponse current to the spectrum to be measured, and predict the characteristic value of the spectrum to be measured based on the mapping relationship between the photoresponse current and the spectrum;

[0010] (4) Substitute the predicted characteristic values of the spectrum to be measured into the Gaussian distribution function to obtain the spectrum curve to be measured.

[0011] The machine learning algorithm model, see Figure 2 As shown, it contains an input layer, an output layer and two hidden layer networks. The input layer is denoted as D1…D n , represents the light response current of each photodetector, n is the number of photodetectors; the output layer is denoted as S1…S m , represents the characteristic value of each spectrum in the training database, m is the number of spectra; the two hidden layer networks are denoted as A1 ...A n1 and B1…B n2 , that is, it is necessary to use the logistic regression model to train the machine learning algorithm model and iteratively calculate the hidden layer network A1 ...A n1 and B1…B n2 The value of is used to establish a mapping relationship between the input layer and the output layer, that is, a mapping relationship between the photoresponse current of the device and the spectrum.

[0012] Preferably, the method of increasing the data volume of the database in the machine learning algorithm is as follows: for each photodetector having a different photoresponse cutoff wavelength, extracting the photoresponse current under a plurality of different voltages, thereby increasing the data volume in the database.

[0013] Preferably, the structure of the photodetector can be a diode, a triode, a thin film transistor, or a resistor structure, and the photosensitive layer can be a single photosensitive material or a composite material of two or more photosensitive materials.

[0014] Preferably, the photodetector is a photodetector based on a series of heterojunctions of CsPbX3 (X = Cl, Br, I) quantum dots with different band gaps and the organic semiconductor 2,7-dioctyl[1]benzothieno[3,2-B]benzothiophene (C8-BTBT).

[0015] Preferably, the photosensitive material may be nanocrystals, nanowires, two-dimensional materials, thin films or single crystals.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] The spectral reconstruction optimization method of the present invention is applicable to all micro spectrometers that only include an array of photodetectors, and can make full use of the advantages of software programming to reduce the requirements for the device distribution density and device performance in the micro spectrometer. This spectral reconstruction algorithm with low requirements for devices can further miniaturize the size of the micro spectrometer while maintaining high-precision spectral reconstruction, laying a foundation for the integration and flexible application of future micro spectrometers. The present invention can be applied to fields such as hyperspectral imaging and flexible electronics. Description of the Drawings

[0018] Figure 1 (a)-(c) are current-voltage curve graphs of three different photodetection devices in the micro spectrometer in the embodiment under illumination of different wavelengths, and (d) is a relationship curve graph of the spectral responsivity of all photodetection devices in the micro spectrometer changing with wavelength.

[0019] Figure 2 It is a machine learning algorithm model framework for spectral reconstruction proposed by the present invention.

[0020] Figure 3 (a) is the result of the spectral curve reconstructed by the machine learning algorithm, and (b) is the result of the spectral curve reconstructed after using the optimization method proposed by the present invention.

[0021] Figure 4 It is a statistical chart of the spectral errors reconstructed before and after using the optimization method proposed by the present invention. Detailed Embodiment

[0022] The present invention will be further described below with reference to the accompanying drawings through specific embodiments.

[0023] The micro spectrometer adopted in this embodiment is based on a series of photodetectors of heterojunctions of different bandgap CsPbX3 (X = Cl, Br, I) quantum dots and the organic semiconductor 2,7-dioctyl[1]benzothieno[3,2-b]benzothiophene (C8-BTBT). The specific preparation process is as follows:

[0024] (1)CsPbBr3 quantum dots were prepared by the hot injection method. Cesium carbonate powder (Cs2CO3, 0.5 g), oleic acid (OA, 2 mL), and octadecene (ODE, 50 mL) were dissolved at 120 °C under vacuum to obtain a Cs-OA solution. Then, lead bromide powder (PbBr2, 0.8 g), ODE (50 mL), OA (5 mL), and oleylamine (OAm, 5 mL) were dissolved at 120 °C under vacuum to obtain a PbBr2 solution. The PbBr2 solution was heated to 180 °C and then rapidly injected into the preheated Cs-OA solution (5 mL) at 70 °C. After reacting for 5 s, the reaction was quenched in an ice-water mixture. Finally, the resulting solution was mixed with methyl acetate at a volume ratio of 1:3 and centrifuged and purified twice to obtain a CsPbBr3 quantum dot solution.

[0025] (2)CsPbX3 quantum dot solutions were prepared. Lead iodide powder (PbI2, 0.461 g), OA (0.1 mL), OAm (0.1 mL), and n-hexane (10 mL) were dissolved to obtain a 0.1 mmol / mL PbI2 solution. A 0.1 mmol / mL lead chloride (PbCl2) solution was prepared in the same manner, and 2 mL of trioctylphosphine (TOP) was additionally added as a co-solvent during the preparation process. Then, 100 μL, 80 μL, 50 μL, 30 μL, 10 μL of the PbCl2 solution and 30 μL, 50 μL, 70 μL, 90 μL, 110 μL, 130 μL, 145 μL, 1 mL of the PbI2 solution were added to 1 mL of the CsPbBr3 quantum dot solution. After reacting for 10 min, each solution was purified according to the purification method of the CsPbBr3 quantum dot crude solution to obtain a CsPbX3 quantum dot solution.

[0026] (3)Device fabrication. CsPbX3 quantum dots were spin-coated on a cleaned Si / SiO2 substrate. After the quantum dots were completely dried, a 300 nm C8-BTBT film was thermally evaporated. Finally, an Au electrode was prepared by thermal evaporation to obtain 14 photodetectors with different light response wavelengths. The device size was 100 μm × 500 μm.

[0027] Figure 1Shows the current-voltage (I-V) curves of device 3, device 8, and device 13 under illumination with the same light intensity but different wavelengths, as well as the relationship curve of the optical responsivity of 14 devices varying with wavelength. It can be seen from the figure that the optical response wavelengths of the 14 devices are different. When the illumination wavelength is greater than the starting optical response wavelength of the device, there is no obvious change in the device current. When the illumination wavelength is less than the starting optical response wavelength of the device and gradually decreases, the current of the device gradually increases. By comparing the relationship curves of the optical responsivity of the 14 devices varying with wavelength, it is found that the starting wavelengths of the optical responses of the devices are evenly distributed in the wavelength range of 400 - 700 nm, indicating that the microspectrometer composed of these 14 photodetectors with different optical response wavelengths can identify and reconstruct unknown spectra in the range of 400 - 700 nm.

[0028] Figure 2 Shows the framework of the machine learning algorithm model for spectral reconstruction. This framework includes an input layer, an output layer, and two hidden layer networks. Among them, the input layer is the optical response current of the photodetector, and the output layer is the characteristic value of the spectrum irradiated on the device. After the framework is built, a logistic regression model is used for training to establish the mapping relationship between the input layer and the output layer, that is, to establish the mapping relationship between the optical response current of the device and the spectrum.

[0029] Figure 3 (a) Shows a comparison chart of the reconstructed unknown spectral curve and the spectral curve tested by a commercial spectrometer based on the above machine learning algorithm. A total of 14 reconstruction results of unknown spectra evenly distributed in the wavelength range of 400 - 700 nm are shown in the figure. It can be seen from the figure that the machine learning algorithm proposed in the present invention can successfully reconstruct the unknown spectrum, but its accuracy is relatively low. Figure 3 (b) Shows the reconstruction result of the unknown spectrum after adopting the optimization method proposed in the present invention. This optimization method is to increase the amount of data in the database in the machine learning algorithm, that is, to extract the optical response current at multiple different voltages as the input network database, thereby increasing the amount of data and achieving the improvement of the spectral reconstruction accuracy without changing the number of devices. Comparing Figure 3 (a) and (b), it is found that after optimization, the reconstructed spectral curve coincides highly with the spectral curve tested by the commercial spectrometer, indicating that the spectral reconstruction accuracy has been improved after adopting the optimization method.

[0030] Figure 4The figure shows the relationship between the peak position and the full width at half maximum (FWHM) errors of the reconstructed spectrum and the number of voltages when the optimization method proposed in the present invention (extracting the photocurrent response of the device at multiple different voltages to increase the amount of data in the database) is adopted. It can be seen from the figure that as the number of voltages increases, the root mean square error (RMSE) of the peak position of the reconstructed spectrum decreases from 14.4 nm to 3.8 nm, and the RMSE of the FWHM decreases from 0.87 nm to 0.53 nm. This indicates that the proposed optimization method for spectrum reconstruction in the present invention can effectively improve the spectrum reconstruction accuracy without changing the number of devices.

[0031] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be recognized that the above description should not be considered as a limitation of the present invention. After those skilled in the art have read the above content, various modifications and alternatives to the present invention will be obvious. Therefore, the protection scope of the present invention should be defined by the appended claims.

Claims

1. A method for optimizing spectral reconstruction of a microspectrometer, characterized in that, A machine learning algorithm based on spectral reconstruction achieves high-precision spectral reconstruction by increasing the data volume in the database of the machine learning algorithm. Among them, for the micro-spectrometer, its structure only includes a series of photodetectors showing different light response cut-off wavelengths, or only a single photodetector, and this single photodetector can change the light response wavelength of this single photodetector by adjusting the bias stress or mechanical stress. The specific steps for spectral reconstruction are as follows: (1) Establish a database: The database includes two parts, namely spectral characteristic values and the light response current of the photodetector. Among them, the spectral characteristic values are the full width at half maximum, peak position or peak height obtained by peak splitting the spectrum using the Gaussian distribution function; the light response current is the increased data volume in the database, which is extracted from each photodetector with different light response cut-off wavelengths at multiple different voltages. (2) Establish the mapping relationship between the light response current and the spectrum by training the machine learning algorithm model. (3) Input the light response current of the photodetector to the spectrum to be measured, and predict the characteristic values of the spectrum to be measured according to the mapping relationship between the light response current and the spectrum. (4) Substitute the predicted characteristic values of the spectrum to be measured into the Gaussian distribution function to obtain the spectral curve of the spectrum to be measured. The machine learning algorithm model described in step (2) includes an input layer, an output layer, and two hidden layer networks. Among them, the input layer is denoted as D1 … D n , representing the photocurrent responses of each photodetector, where n is the number of photodetectors; the output layer is denoted as S1 … S m , representing the eigenvalue of each spectrum in the training database, where m is the number of spectra; the two hidden layer networks are respectively denoted as A1 … A n1 and B1 … B n2 , that is, the value calculated by iterative calculation is required to train the machine learning algorithm model using a logistic regression model; Use the logistic regression model to train the machine learning algorithm model to establish the mapping relationship between the input layer and the output layer, that is, establish the mapping relationship between the light response current of the device and the spectrum.

2. The spectral reconstruction optimization method of the micro-spectrometer according to claim 1, wherein The structure of the photodetector is a diode, a triode, a thin film transistor or a resistive structure, and the photosensitive layer therein is a single photosensitive material or a composite material of two or more photosensitive materials.

3. The spectral reconstruction optimization method of the microspectrometer according to claim 2, characterized in that The photodetector is a photodetector based on a heterojunction of a series of different bandgap CsPbX3 quantum dots and the organic semiconductor 2,7-dioctyl[1]benzothieno[3,2-b]benzothiophene C8-BTBT, where X is Cl, Br or I.

4. The spectral reconstruction optimization method of the micro-spectrometer according to claim 2, characterized in that The photosensitive material is a nanocrystal, a nanowire, a two-dimensional material, a thin film or a single crystal.

Images