• Advanced Photonics Nexus
  • Vol. 3, Issue 6, 066014 (2024)
Feiyu Guan1,†, Yuanchao Liu2, Xuechen Niu1, Weihua Huang1..., Wei Li3, Peichao Zheng3, Deng Zhang4, Gang Xu5,* and Lianbo Guo1,*|Show fewer author(s)
Author Affiliations
  • 1Huazhong University of Science and Technology, Wuhan National Laboratory for Optoelectronics, Wuhan, China
  • 2City University of Hong Kong, Department of Physics, Hong Kong, China
  • 3Chongqing University of Posts and Telecommunications, School of Optoelectronic Engineering, Chongqing, China
  • 4Nanjing Normal University, School of Computer and Electronic Information, Nanjing, China
  • 5Huazhong University of Science and Technology, School of Optical and Electronic Information, Wuhan, China
  • show less
    DOI: 10.1117/1.APN.3.6.066014 Cite this Article Set citation alerts
    Feiyu Guan, Yuanchao Liu, Xuechen Niu, Weihua Huang, Wei Li, Peichao Zheng, Deng Zhang, Gang Xu, Lianbo Guo, "AI-enabled universal image-spectrum fusion spectroscopy based on self-supervised plasma modeling," Adv. Photon. Nexus 3, 066014 (2024) Copy Citation Text show less
    Overview of SISTIFD. (a) The schematic diagram of the system. When the plasma is produced, an intensified charge-coupled device (ICCD) camera and a spectrometer collect signals synchronously according to predetermined procedures. (b) The conventional single-spectrum technique in plasma spectroscopy only uses a spectrum, which exhibits some shortcomings, and the results are not satisfactory. (c) The SISTIFD we proposed can synchronously capture images and emission spectra of plasma and extract many deep features, such as area and brightness to predict physical parameters based on an AI model, thus achieving accurate and precise detection. (The size of symbols represents the standard deviation.)
    Fig. 1. Overview of SISTIFD. (a) The schematic diagram of the system. When the plasma is produced, an intensified charge-coupled device (ICCD) camera and a spectrometer collect signals synchronously according to predetermined procedures. (b) The conventional single-spectrum technique in plasma spectroscopy only uses a spectrum, which exhibits some shortcomings, and the results are not satisfactory. (c) The SISTIFD we proposed can synchronously capture images and emission spectra of plasma and extract many deep features, such as area and brightness to predict physical parameters based on an AI model, thus achieving accurate and precise detection. (The size of symbols represents the standard deviation.)
    Workflow of PISA-Net. (a) Auto-preprocessing. After the image-spectrum fusion acquisition, the images and spectra need to be preprocessed automatically as the input of PISA-Net. The table contains examples of data that are influenced by different interferences. The images need to be cropped to keep the plasma centered and then be normalized and subsampled for better training. As for the spectra, the first step is finding the spectral peaks of interest. Similarly, the spectra also need to be normalized and subsampled before training. (b) The architecture of PISA-Net consists of an image pipeline and a spectrum pipeline to process different signals and merges them into a plasma feature stack. The feature stack contains both plasma physical parameters and high dimensional eigenvectors. Finally, PISA-Net outputs the final results based on these features. (c) The residual attention cell (RACell), including the channel-attention network and soft thresholding mechanism, is well-designed for sparsity and less texture of the plasma image.
    Fig. 2. Workflow of PISA-Net. (a) Auto-preprocessing. After the image-spectrum fusion acquisition, the images and spectra need to be preprocessed automatically as the input of PISA-Net. The table contains examples of data that are influenced by different interferences. The images need to be cropped to keep the plasma centered and then be normalized and subsampled for better training. As for the spectra, the first step is finding the spectral peaks of interest. Similarly, the spectra also need to be normalized and subsampled before training. (b) The architecture of PISA-Net consists of an image pipeline and a spectrum pipeline to process different signals and merges them into a plasma feature stack. The feature stack contains both plasma physical parameters and high dimensional eigenvectors. Finally, PISA-Net outputs the final results based on these features. (c) The residual attention cell (RACell), including the channel-attention network and soft thresholding mechanism, is well-designed for sparsity and less texture of the plasma image.
    Results of plasma parameter calculation. (a) Collected plasma images and (b) spectra at different temperatures. (c) The Boltzmann plot of the Fe element for measurement of plasma temperature. A total of 144 spectral lines were selected for fitting. The plasma temperature can be calculated according to the intercept. (d) The comparison results of the plasma temperature T under different energy conditions. The results of t-test indicate that there is no statistical difference. (e) The relative deviation of the total 50 results obtained by the Boltzmann plot method and SISTIFD under the same condition. (f) Spectra of the Fe I 404.581 nm spectral line after Lorentz fitting under different energy conditions. The electron density can be approximate, estimated based on the FWHM of the spectral lines. (g) The comparison results of electron density ne under different energy conditions. (h) The relative deviation of the total 50 results obtained by the Boltzmann plot method and PISA-Net under the same energy. The results of a t-test indicate that there is no statistical difference. (i) Spectra of soil samples for comparison. (j) The quantitative analysis using original data. (k) The quantitative analysis corrected by T and ne. (l) The quantitative analysis corrected by all parameters (T, ne, ns, and δ).
    Fig. 3. Results of plasma parameter calculation. (a) Collected plasma images and (b) spectra at different temperatures. (c) The Boltzmann plot of the Fe element for measurement of plasma temperature. A total of 144 spectral lines were selected for fitting. The plasma temperature can be calculated according to the intercept. (d) The comparison results of the plasma temperature T under different energy conditions. The results of t-test indicate that there is no statistical difference. (e) The relative deviation of the total 50 results obtained by the Boltzmann plot method and SISTIFD under the same condition. (f) Spectra of the Fe I 404.581 nm spectral line after Lorentz fitting under different energy conditions. The electron density can be approximate, estimated based on the FWHM of the spectral lines. (g) The comparison results of electron density ne under different energy conditions. (h) The relative deviation of the total 50 results obtained by the Boltzmann plot method and PISA-Net under the same energy. The results of a t-test indicate that there is no statistical difference. (i) Spectra of soil samples for comparison. (j) The quantitative analysis using original data. (k) The quantitative analysis corrected by T and ne. (l) The quantitative analysis corrected by all parameters (T, ne, ns, and δ).
    The quantitative results in LIBS. (a) The schematic diagram of LIBS. A laser is focused on the target, and the generated plasma will emit a spectrum with its elemental fingerprinting capability. (b) Examples of plasma spectra and images under different interference conditions, including spectral fluctuation, unstable excitation, matrix effect, and self-absorption. The spectra and images show significant inconsistencies. (c) Part of the original images and (d) spectra in the experiment under cross-interference, composite, and high throughput conditions. (e) Calibration curves of K I 766.490 nm of conventional LIBS. Different kinds of samples (potash feldspar and soil) were used; the energy of the excitation laser was unstable. Herein, the calibration curves of conventional LIBS were influenced by all sorts of interferences mentioned above, indicating unacceptable spectral analytical results. (f) The results of SISTIFD. These interferences can be overcome by SISTIFD to achieve accurate quantification. Error bars in (e) and (f) represent standard deviation (s.d.) for each data point (n=100), and points are average values.
    Fig. 4. The quantitative results in LIBS. (a) The schematic diagram of LIBS. A laser is focused on the target, and the generated plasma will emit a spectrum with its elemental fingerprinting capability. (b) Examples of plasma spectra and images under different interference conditions, including spectral fluctuation, unstable excitation, matrix effect, and self-absorption. The spectra and images show significant inconsistencies. (c) Part of the original images and (d) spectra in the experiment under cross-interference, composite, and high throughput conditions. (e) Calibration curves of K I 766.490 nm of conventional LIBS. Different kinds of samples (potash feldspar and soil) were used; the energy of the excitation laser was unstable. Herein, the calibration curves of conventional LIBS were influenced by all sorts of interferences mentioned above, indicating unacceptable spectral analytical results. (f) The results of SISTIFD. These interferences can be overcome by SISTIFD to achieve accurate quantification. Error bars in (e) and (f) represent standard deviation (s.d.) for each data point (n=100), and points are average values.
    The quantitative results in GD-OES. (a) The schematic diagram of GD-OES. High voltage is applied to both ends to excite the plasma, and the sample is passed through a capillary tube. (b) The plasma image of the Li element. (c) The plasma image of the Cs element. (d) The spectra of Li 670.8 nm in different concentrations. (e) Calibration curve of Li 670.8 nm based on conventional GD-OES. (f) Calibration curve of Li 670.8 nm based on SISTIFD. (g) The spectra of Cs 852.1 nm in different concentrations. (h) Calibration curve of Cs 852.1 nm based on conventional GD-OES. (i) Calibration curve of Cs 852.1 nm based on SISTIFD.
    Fig. 5. The quantitative results in GD-OES. (a) The schematic diagram of GD-OES. High voltage is applied to both ends to excite the plasma, and the sample is passed through a capillary tube. (b) The plasma image of the Li element. (c) The plasma image of the Cs element. (d) The spectra of Li 670.8 nm in different concentrations. (e) Calibration curve of Li 670.8 nm based on conventional GD-OES. (f) Calibration curve of Li 670.8 nm based on SISTIFD. (g) The spectra of Cs 852.1 nm in different concentrations. (h) Calibration curve of Cs 852.1 nm based on conventional GD-OES. (i) Calibration curve of Cs 852.1 nm based on SISTIFD.
    Comparison of results obtained by conventional LIBS (left) and SISTIFD (right) under different interference conditions. (a) Calibration curve of Si I 250.690 nm spectral line using conventional LIBS within experimental condition 1 that displayed spectral fluctuation. (b) The corresponding results using SISTIFD. (c) Calibration curves of Si I 250.690 nm spectral line using conventional LIBS within experimental condition 2 that existed in unstable excitation. (d) The corresponding results using SISTIFD. (e) Calibration curves of Mn II 293.931 nm spectral line using conventional LIBS in experimental condition 3 that existed as a matrix effect. (f) The corresponding results using SISTIFD. (g) Calibration curves of K I 766.490 nm spectral line using conventional LIBS in experimental condition 4 that existed as self-absorption. (h) The corresponding results using SISTIFD. All the error bars represent s.d. for each data point (n=100), and points are average values.
    Fig. 6. Comparison of results obtained by conventional LIBS (left) and SISTIFD (right) under different interference conditions. (a) Calibration curve of Si I 250.690 nm spectral line using conventional LIBS within experimental condition 1 that displayed spectral fluctuation. (b) The corresponding results using SISTIFD. (c) Calibration curves of Si I 250.690 nm spectral line using conventional LIBS within experimental condition 2 that existed in unstable excitation. (d) The corresponding results using SISTIFD. (e) Calibration curves of Mn II 293.931 nm spectral line using conventional LIBS in experimental condition 3 that existed as a matrix effect. (f) The corresponding results using SISTIFD. (g) Calibration curves of K I 766.490 nm spectral line using conventional LIBS in experimental condition 4 that existed as self-absorption. (h) The corresponding results using SISTIFD. All the error bars represent s.d. for each data point (n=100), and points are average values.
    MethodR2RMSEMRERSD
    Conventional LIBS0.03590.60320.34710.1428
    PISA-Net0.99960.01090.00620.0037
    Table 1. Evaluation parameters of the calibration curves established with different methods in LIBS.
    Case 1 (Li)Reference value (ppm)Analytical result (ppm)R2MRE
    Conventional GD-OES0.03000.03310.93550.1033
    SISTIFD0.03010.99970.0033
    Case 2 (Cs)Reference value (ppm)Analytical result (ppm)R2MRE
    Conventional GD-OES0.60000.61760.97840.0293
    SISTIFD0.59340.99930.0110
    Table 2. Evaluation parameters of the calibration curves established with different methods in GD-OES.
    ConditionElementEvaluation indexMethod
    ConventionalIA-LIBS36SISTIFD (ours)
    Condition 1 spectral fluctuationSi (250.690 nm)R20.98010.98870.9998
    RMSE0.98130.69010.0091
    MRE0.05890.03520.0014
    RSD0.07430.06450.0014
    Fe (239.563 nm)R20.92100.95870.9984
    RMSE0.05380.04210.0017
    MRE0.34900.06540.0049
    RSD0.09830.06990.0032
    Condition 2 unstable excitationSi (250.690 nm)R20.22100.76010.9959
    RMSE0.51230.25110.0260
    MRE0.80810.36210.0647
    RSD2.24650.63320.0501
    Mg (279.553 nm)R20.13200.84660.9992
    RMSE0.39310.27390.0155
    MRE1.63590.83560.0505
    RSD1.71540.81430.0448
    Condition 3 matrix effectMn (293.931 nm)R20.41510.96130.9952
    RMSE0.43260.20610.0044
    MRE0.44570.08170.0099
    Mg (285.213 nm)R20.02440.94200.9998
    RMSE0.13820.01730.0013
    MRE0.63300.07330.0064
    Condition 4 self-absorptionK (766.490 nm)R2Exp fitting0.96210.9995
    RMSE0.4266 (Exp)0.11710.0041
    MRE0.2138 (Exp)0.09240.0019
    K (769.896 nm)R2Exp fitting0.94670.9989
    RMSE0.3669 (Exp)0.24310.0053
    MRE0.3852 (Exp)0.10130.0046
    Table 3. Evaluation indices of the calibration curves established with different methods. Two different spectral lines were selected for each interference condition. The results show that our SISTIFD achieved SOTA performance in every indicator.
    PropertySmallLargeNo attentionNo soft thresholdingNo skip connectionLinearity loss onlyPISA-Net (pre-trained)PISA-Net
    8.8 M parameters
    26.4 M parameters
    98.2 M parameters
    Attention
    Soft thresholding
    Skip connection
    Loss coefficient α0.6 → 0.20.6 → 0.20.6 → 0.20.6 → 0.20.6 → 0.20.6 → 0.20.60.6 → 0.2
    Loss (relative)15.87511.62339.8558.3319.74420.815102.3241.000
    R20.9230.9740.9820.9910.9890.9980.9910.999
    Table 4. Evaluation of models with varying architectures. Each variation (columns) combined by different components (rows) is indicated by ticks (√), e.g., variation “large” is a model with 98.2 M parameters using attention, skip connection, soft thresholding, and setting loss coefficient α=0.6 in the pre-training stage while α=0.2 in the fine-tune stage.
    Feiyu Guan, Yuanchao Liu, Xuechen Niu, Weihua Huang, Wei Li, Peichao Zheng, Deng Zhang, Gang Xu, Lianbo Guo, "AI-enabled universal image-spectrum fusion spectroscopy based on self-supervised plasma modeling," Adv. Photon. Nexus 3, 066014 (2024)
    Download Citation