A physical information neural network-based limb observation ozone profile inversion method
By embedding the residual loss function of the radiative transfer equation using a physical information neural network method, the problems of low computational efficiency and insufficient physical constraints in existing ozone profile inversion calculations are solved, and efficient and physically consistent ozone profile inversion is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-02-25
- Publication Date
- 2026-06-19
AI Technical Summary
Existing edge-observation ozone profile inversion methods suffer from low computational efficiency, lack of physical constraints, and poor generalization ability, making it difficult to meet the needs of operational use.
A physical information neural network-based approach is adopted. By constructing a loss function that includes the residuals of the radiative transfer equation, the physical information neural network model is trained, and the gradient is calculated using automatic differentiation technology to achieve end-to-end ozone profile inversion.
It improves computational efficiency, ensures the physical consistency of inversion results, enhances the model's generalization ability in sparse regions of training data, and the inversion time for a single profile is only in the millisecond range, meeting business requirements.
Smart Images
Figure CN122241058A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of satellite remote sensing atmospheric composition inversion technology, specifically to a method for inverting ozone profiles based on physical information neural networks. Background Technology
[0002] Stratospheric ozone is an important component of Earth's atmospheric system, playing a crucial role in absorbing solar ultraviolet radiation and protecting life on Earth from harmful ultraviolet rays. Accurately obtaining information on the vertical distribution of ozone is of significant scientific importance for understanding atmospheric chemical processes, monitoring ozone layer recovery, and studying climate change.
[0003] Limb scattering observation is an important means of obtaining ozone profiles. Satellite sensors such as OMPS-LP (Ozone Mapping and Profiler Suite Limb Profiler), OSIRIS (Optical Spectrograph and InfraRed Imaging System), and SCIAMACHY (Scanning Imaging Absorption Spectrometer for Atmospheric Chartography) all use limb scattering mode to obtain information on the vertical distribution of atmospheric components.
[0004] Existing methods for inverting ozone profiles from near-shore observations mainly include:
[0005] (1) Optimal Estimation: This method is based on Bayesian theory and minimizes the difference between observed and simulated radiance through iterative solution. However, this method has low computational efficiency, and each profile inversion requires multiple calls to the radiative transfer model, which is difficult to meet the needs of operational use.
[0006] (2) Look-up Table Method: This method pre-calculates the radiance under different atmospheric conditions and performs inversion by interpolating and looking up a table. However, this method requires a large amount of storage and is difficult to cover all possible combinations of atmospheric conditions.
[0007] (3) Traditional machine learning methods, such as support vector machines and random forests, are fast in computation but lack physical constraints and have poor generalization ability in areas with insufficient training data coverage.
[0008] The above methods all have certain limitations, especially in balancing computational efficiency and physical consistency. Summary of the Invention
[0009] The technical problem this invention aims to solve is: addressing the shortcomings of existing technologies by proposing a near-edge observation ozone profile inversion method based on a physical information neural network, thus resolving the issues of low computational efficiency, lack of physical constraints, and poor generalization ability in existing methods. To solve these problems, this invention adopts the following technical solution:
[0010] This invention proposes a method for inverting ozone profiles based on limb observations using a physical information neural network, comprising a model training phase and a real-time inversion phase:
[0011] The model training phase includes the following steps:
[0012] Based on the atmospheric radiative transfer model, a dataset of limb observation radiance-ozone profiles was constructed. The dataset includes the following parameters: multi-wavelength limb radiance, tangent height, solar zenith angle, observation zenith angle, relative azimuth angle, temperature profile, air pressure profile, and ozone number density profile.
[0013] A physical information neural network model is constructed, and the residual of the radiative transfer equation describing the limb observation geometry is embedded as a physical constraint term into the loss function of the neural network model. The multi-wavelength limb radiance, tangent height, solar zenith angle, observation zenith angle, and relative azimuth angle are used as the input layer of the neural network, and the ozone number density profile is used as the output layer of the neural network.
[0014] The dataset is used as training samples to train the physical information neural network model through a composite loss function, which includes at least: a data loss term, a residual loss term of the radiative transfer equation, a boundary condition loss term, and a physical constraint loss term; after training, an ozone profile inversion model is obtained.
[0015] The real-time inversion stage includes the following steps:
[0016] Acquire observation data from the edge detector and extract parameters such as multi-wavelength edge radiance, tangent height, solar zenith angle, observation zenith angle, and relative azimuth angle;
[0017] The extracted parameters are input into the trained ozone profile inversion model to calculate the ozone number density profile inversion result.
[0018] Preferably, the multi-wavelength lateral radiance includes radiance data in the ultraviolet band and the visible light band, wherein the ultraviolet band wavelengths include 295nm, 302nm, 306nm, 312nm, 317nm, 322nm, and 353nm, and the visible light band wavelengths include 606nm.
[0019] Preferably, the residuals of the radiative transfer equation are calculated in the following manner:
[0020] The radiative transfer equation describing the near-edge observation geometry is as follows:
[0021] ,
[0022] in, For radiance, For height, To observe the cosine value of the zenith angle, It is the azimuth angle. Extinction coefficient, The source function;
[0023] Using automatic differentiation techniques, the partial derivative of the radiance output by the neural network with respect to height is calculated;
[0024] Calculate its residuals based on the radiative transfer equation. :
[0025] ,
[0026] The residual loss term of the radiative transfer equation for:
[0027] ,
[0028] in, This represents the number of configuration points.
[0029] Preferably, the extinction coefficient The calculation method is as follows:
[0030] ,
[0031] in, The Rayleigh scattering coefficient is... This is the ozone absorption coefficient. The aerosol extinction coefficient;
[0032] The ozone absorption coefficient is:
[0033] ,
[0034] in, For temperature-dependent ozone absorption cross section, This is the ozone number density profile.
[0035] Preferably, the boundary condition loss term includes:
[0036] Atmospheric top boundary condition: incident diffuse radiation is zero;
[0037] Surface boundary conditions: Calculate surface reflected radiation based on the Lambert reflection model or the two-way reflection distribution function;
[0038] The boundary condition loss term is calculated using the following formula:
[0039] ,
[0040] in, The number of boundary points. The neural network predicted the value. These are the boundary condition constraint values.
[0041] Preferably, the physical constraint loss term includes:
[0042] Nonnegativity constraint : Ensure that the predicted ozone concentration is non-negative;
[0043] Smoothing constraints Regularize the second derivative of the ozone profile;
[0044] Chapman layer prior constraints The peak ozone concentration is constrained to be within the 20-25 km altitude range;
[0045] upper boundary constraints : To constrain the concentration of ozone in the upper atmosphere to approach zero;
[0046] Physical constraint loss term Calculate using the following formula:
[0047] ,
[0048] in, , , , These are the weighting coefficients for each constraint term.
[0049] Preferably, the composite loss function for:
[0050] ,
[0051] in, For data loss items, This is the residual loss term in the radiative transfer equation. For boundary condition loss terms, For physical constraint loss terms, , , , These are the weighting coefficients for each loss term.
[0052] Preferably, the physical information neural network includes one input layer, one output layer, and four hidden layers, with each of the four hidden layers containing 128, 128, 128, and 128 neurons respectively, and the activation function is the hyperbolic tangent function.
[0053] Preferably, the edge detector selects OMPS-LP, OSIRIS, or SCIAMACHY edge observation mode.
[0054] Preferably, the training employs a two-stage optimization strategy:
[0055] The first stage uses the Adam optimizer for pre-training with a learning rate set to 10%. -3 ;
[0056] The second stage uses the L-BFGS optimizer for refined training.
[0057] Preferably, the atmospheric radiative transfer model is the SCIATRAN model, and the pseudo-spherical approximation method is used to handle the near-edge observation geometry.
[0058] Preferably, the ozone number density profile has an altitude range of 12.5 km to 57.5 km, a vertical resolution of 1 km, and a total of 46 altitude layers.
[0059] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:
[0060] (1) High computational efficiency: After training, the inversion of a single profile only requires milliseconds of computation time, which is 2-3 orders of magnitude faster than the optimal estimation method;
[0061] (2) Strong physical consistency: By embedding the radiative transfer equation into the loss function, the inversion results are ensured to satisfy the physical laws of atmospheric radiative transfer;
[0062] (3) Good generalization ability: Physical constraints are used as regularization terms to improve the model's extrapolation ability in sparse regions of training data;
[0063] (4) End-to-end learning: The gradient is automatically calculated through automatic differentiation technology without the need to pre-calculate the Jacobian matrix. Attached Figure Description
[0064] Figure 1 This invention relates to a flowchart of a method for inverting ozone profiles based on physical information neural networks.
[0065] Figure 2 This is a schematic diagram of the physical information neural network structure involved in the present invention.
[0066] Figure 3 This is a schematic diagram of the composite loss function structure involved in the embodiment.
[0067] Figure 4 This is a schematic diagram of the edge observation geometry involved in the embodiment.
[0068] Figure 5 This is a flowchart illustrating the residual calculation process for the radiative transfer equation involved in the embodiment.
[0069] Figure 6 This is a comparison chart verifying the inversion results involved in the examples.
[0070] Figure 7 This is a comparison chart of the correlation between the PINN inversion results and MLS in the example.
[0071] Figure 8 This is a comparison diagram of the inversion results and the sounding results involved in the example. Detailed Implementation
[0072] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0073] Example 1: This example proposes a method for inverting ozone profiles from limb observations based on a physical information neural network. (Reference) Figure 1 This method includes a model training phase and a real-time inversion phase:
[0074] 1. Model training phase:
[0075] Based on an atmospheric radiative transfer model, a dataset of limb observation radiance-ozone profiles was constructed. This dataset includes the following parameters: multi-wavelength limb radiance, tangent height, solar zenith angle, observed zenith angle, relative azimuth angle, temperature profile, pressure profile, and ozone number density profile. The atmospheric radiative transfer model used is the SCIATRAN model. SCIATRAN is a general-purpose radiative transfer model capable of handling radiative transfer in the ultraviolet-visible-near-infrared bands in planar parallel and pseudo-spherical atmospheres, supporting limb observation geometry. This model considers radiative transfer processes such as Rayleigh scattering, aerosol scattering, gas absorption, and surface reflection. The multi-wavelength limb radiance includes radiance data in the ultraviolet and visible light bands. The ultraviolet wavelengths include 295nm, 302nm, 306nm, 312nm, 317nm, 322nm, and 353nm, and the visible light wavelength includes 606nm. The multi-wavelength selection and corresponding ozone absorption cross-sections are shown in Table 1.
[0076] Table 1
[0077]
[0078] The above values are estimates based on the Serdyuchenko et al. (2014) dataset at around ~223K. In actual inversion, the algorithm dynamically interpolates the precise cross-sectional values for each height layer based on the input temperature profile.
[0079] 353nm is listed in the document and is typically used as a non-absorption reference channel to construct differential absorption pairs (such as 302nm / 353nm).
[0080] 606nm is located at the peak center of the Chappuis band (the ~600-603nm region) and is the core wavelength for visible light inversion.
[0081] The process of constructing the dataset includes:
[0082] (1) Set the ozone profile perturbation range: Based on the climatological ozone profile, apply ±50% random perturbation at each altitude layer to generate diverse ozone profile samples;
[0083] (2) Set the range of observation geometric parameters: the solar zenith angle range is 30°-85°, the tangent point height range is 12.5-57.5km, and the relative azimuth range is 0°-180°;
[0084] (3) Call the SCIATRAN model to calculate the multi-wavelength radiance under each combination condition;
[0085] (4) Construct input-output data pairs to form a training dataset.
[0086] The dataset contains parameter data as shown in Table 2:
[0087] Table 2
[0088]
[0089] A physical information neural network model is constructed, and the residual of the radiative transfer equation describing the limb observation geometry is embedded as a physical constraint term into the loss function of the neural network model. The multi-wavelength limb radiance, tangent height, solar zenith angle, observation zenith angle, and relative azimuth angle are used as the input layer of the neural network, and the ozone number density profile is used as the output layer of the neural network.
[0090] The neural network consists of one input layer, one output layer, and four hidden layers. The input layer has a dimension of [missing value]. ,in The number of wavelength channels is (8). The height of the tangent point is the number of layers (46). The number of geometric parameters is (3). The output layer dimension is 46, corresponding to the ozone number density of 46 height layers. Each of the four hidden layers contains 128 neurons, and the activation function is the hyperbolic tangent function (Tanh). The hyperparameter settings of the neural network are shown in Table 3.
[0091] Table 3
[0092]
[0093] Input layer: According to claim 3, if all accumulated wavelengths (295nm, 302nm, 306nm, 312nm, 317nm, 322nm, 353nm) are used simultaneously, the input feature dimension is 8 wavelength radiance + 4 geometric parameters = 12 input nodes.
[0094] Output layer: Outputs the corresponding ozone profile. Claim 13 specifies the altitude range as 12.5km-57.5km with a step size of 1km, therefore the output layer is fixed at 46 layers.
[0095] Activation function selection: Tanh was chosen instead of ReLU, which is commonly used in deep learning, because this method requires the use of automatic differentiation techniques to calculate the residuals of the radiative transfer equation (involving partial derivatives with respect to height). Tanh is a smooth, differentiable function that supports the calculation of higher-order derivatives, while ReLU is not differentiable at zero and is not suitable for this type of Physical Information Neural Network (PINN).
[0096] The core innovation of the physical information neural network lies in incorporating the residuals of the radiative transfer equation as part of the loss function. The radiative transfer equation under the near-edge observation geometry is:
[0097]
[0098] in:
[0099] Radiance, in W / (m²) 2 .sr -1 .nm -1 );
[0100] z represents altitude, in km;
[0101] , To observe the zenith angle;
[0102] It is the relative azimuth angle;
[0103] Extinction coefficient, unit: km. -1；
[0104] The source function;
[0105] The extinction coefficient is calculated as follows:
[0106] ;
[0107] in:
[0108] Rayleigh scattering coefficient;
[0109] Ozone absorption coefficient;
[0110] is the aerosol extinction coefficient.
[0111] Using PyTorch's automatic differentiation feature, calculate the partial derivative of the radiance output of the neural network with respect to height z. .
[0112] The ozone absorption coefficient is:
[0113] ,
[0114] in, For temperature-dependent ozone absorption cross section, This is the ozone number density profile.
[0115] The dataset is used as training samples to train the physical information neural network model through a composite loss function, which includes at least: a data loss term, a residual loss term of the radiative transfer equation, a boundary condition loss term, and a physical constraint loss term; after training, an ozone profile inversion model is obtained.
[0116] refer to Figure 3 The composite loss function comprises four parts:
[0117]
[0118] (1) Data loss items : Measuring the difference between the ozone profile predicted by the neural network and the true value:
[0119]
[0120] (2) Residual loss term of radiative transfer equation : Measures the degree to which the output of a neural network satisfies the radiative transfer equation:
[0121]
[0122] in The number of configuration points is specified, which are generated using Sobol sequences or Latin hypercube sampling.
[0123] Reference for the calculation process of the residuals of the radiative transfer equation Figure 5 .
[0124] (3) Boundary condition loss term Includes top atmospheric boundary conditions and surface boundary conditions:
[0125] Atmospheric top boundary conditions ( The incident diffuse radiation is zero. ;
[0126] Surface boundary condition z = 0: Lambert reflection ;
[0127] The boundary condition loss term is calculated using the following formula:
[0128] ,
[0129] in, The number of boundary points. The neural network predicted the value. These are the boundary condition constraint values.
[0130] (4) Physical constraint loss term Includes the following constraints:
[0131] Nonnegativity constraint: ;
[0132] Smoothing constraints: ;
[0133] Chapman layer prior constraints: ;
[0134] Upper boundary constraints: ;
[0135] The training employs a two-stage optimization strategy:
[0136] The first stage uses the Adam optimizer for pre-training, with 1000-5000 iterations and a learning rate of 10. -3 This is used to quickly converge to the feasible region;
[0137] The second stage uses the L-BFGS optimizer for fine-tuning training, with 500-1000 iterations, to obtain a high-precision solution.
[0138] The initial values of the weighting coefficients are set as follows: , , , During training, an adaptive weight adjustment strategy can be adopted to dynamically balance the weights based on the gradient norm of each loss term.
[0139] Specifically, the physical information neural network comprises one input layer, one output layer, and four hidden layers. The four hidden layers contain 128, 128, 128, and 128 neurons respectively, and the activation function is the hyperbolic tangent function. The specific structure is as follows: Figure 2 As shown.
[0140] 2. Real-time inversion stage:
[0141] Acquire observation data from the edge detector and extract parameters such as multi-wavelength edge radiance, tangent height, solar zenith angle, observation zenith angle, and relative azimuth angle;
[0142] The extracted parameters are input into the trained ozone profile inversion model to calculate the ozone number density profile inversion result. (See the schematic diagram of limb observation for reference.) Figure 4 .
[0143] Specifically, it includes the following steps:
[0144] Acquire Level 1B observation data from the edge detector and perform inversion. The edge detector can be selected from:
[0145] (1) OMPS-LP (Ozone Mapping and Profiler Suite Limb Profiler): Mounted on Suomi-NPP and JPSS series satellites, covering an altitude range of 0-80km with a vertical resolution of approximately 1km;
[0146] (2) OSIRIS (Optical Spectrograph and InfraRed Imaging System): Mounted on the Odin satellite, with an observation band of 280-800nm;
[0147] (3) SCIAMACHY edge observation mode: carried on the Envisat satellite (now retired), with an observation band of 240-2380nm.
[0148] Data preprocessing steps include:
[0149] (1) Read Level 1B radiance data;
[0150] (2) Select the radiance of the specified wavelength channel;
[0151] (3) Calculate the radiance ratio (relative to the reference wavelength or the radiance of the top of the atmosphere);
[0152] (4) Extract the corresponding geometric parameters (solar zenith angle, observed zenith angle, relative azimuth angle);
[0153] (5) Obtain auxiliary data (temperature profile, air pressure profile).
[0154] The preprocessed data is input into a trained physical information neural network model, which outputs an ozone number density profile. The inversion time for a single profile is approximately 1-10 milliseconds (GPU) or 10-100 milliseconds (CPU). Specifically, the ozone number density profile has an altitude range of 12.5 km to 57.5 km, a vertical resolution of 1 km, and a total of 46 altitude layers.
[0155] 3. Calculation of the source function:
[0156] The source function The calculation of the source function is crucial for the residual loss in the radiative transfer equation. The source function includes both single-scattering and multiple-scattering contributions: ;
[0157] The method for calculating the single scattering source function is as follows: ;
[0158] in:
[0159] This refers to the single-scattering albedo.
[0160] Solar irradiance at the top of the atmosphere;
[0161] Let be the phase function. The scattering angle;
[0162] The optical thickness is the solar slant path thickness.
[0163] The formula for calculating the scattering angle is:
[0164] ;
[0165] in , This is the solar zenith angle.
[0166] The contribution of multiple scattering is approximated using a neural network parameterization method: ;
[0167] in For pre-trained multiple scattering parameterization network, is the phase function asymmetry factor. This parameterized network was trained using accurate multiple scattering calculations from SCIATRAN.
[0168] 4. Multi-wavelength selection strategy:
[0169] The multi-wavelength selection is based on the ozone absorption spectral characteristics and the features of the limb observation signal, as shown in Table 4:
[0170] Table 4
[0171]
[0172] Using the radiance ratio (radiance of each wavelength divided by the radiance of the reference wavelength) as the input to the neural network can eliminate the influence of absolute calibration error and improve inversion stability.
[0173] 5. Verification method:
[0174] The inversion results are verified using the following methods:
[0175] (1) Comparison with MLS ozone profile: Aura satellite MLS (Microwave Limb Sounder) provides independent ozone profile observations, which can be used to verify the inversion results;
[0176] The correlation between the inversion results of the method described in this invention and MLS is shown in the figure below. Figure 7 As shown.
[0177] (2) Comparison with ozone sounding: Ground-based ozone sounding provides high-precision true values of ozone profiles;
[0178] The inversion results are compared with the sounding results in the figure below. Figure 8 As shown.
[0179] (3) Comparison with official products: Compare with the official OMPS-LP V2.5 product;
[0180] (4) Self-consistency test: Substitute the inverted ozone profile into the radiative transfer model, calculate the simulated radiance, and compare it with the observed radiance.
[0181] The scatter plot comparing the inversion results with MLS is shown below. Figure 6 As shown in Table 5, the error statistics are as follows.
[0182] Table 5
[0183]
[0184] The limb observation ozone profile inversion method based on physical information neural network provided in this embodiment achieves an organic combination of physical constraints and data-driven approaches by embedding the radiative transfer equation into the neural network loss function. It has the advantages of high computational efficiency, strong physical consistency, and good generalization ability.
[0185] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0186] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.
Claims
1. A method for inverting ozone profiles from limb observations based on a physical information neural network, characterized in that, Includes a model training phase and a real-time inversion phase: The model training phase includes the following steps: Based on the atmospheric radiative transfer model, a dataset of limb observation radiance-ozone profiles was constructed. The dataset includes the following parameters: multi-wavelength limb radiance, tangent height, solar zenith angle, observation zenith angle, relative azimuth angle, temperature profile, air pressure profile, and ozone number density profile. A physical information neural network model is constructed, and the residual of the radiative transfer equation describing the limb observation geometry is embedded as a physical constraint term into the loss function of the neural network model. The multi-wavelength limb radiance, tangent height, solar zenith angle, observation zenith angle, and relative azimuth angle are used as the input layer of the neural network, and the ozone number density profile is used as the output layer of the neural network. The dataset is used as training samples to train the physical information neural network model through a composite loss function, which includes at least: a data loss term, a residual loss term of the radiative transfer equation, a boundary condition loss term, and a physical constraint loss term; after training, an ozone profile inversion model is obtained. The real-time inversion stage includes the following steps: Acquire observation data from the edge detector and extract parameters such as multi-wavelength edge radiance, tangent height, solar zenith angle, observation zenith angle, and relative azimuth angle; The extracted parameters are input into the trained ozone profile inversion model to calculate the ozone number density profile inversion result.
2. The method according to claim 1, characterized in that, The multi-wavelength lateral radiance includes radiance data in the ultraviolet and visible light bands. The ultraviolet band wavelengths include 295nm, 302nm, 306nm, 312nm, 317nm, 322nm, and 353nm, and the visible light band wavelengths include 606nm.
3. The method according to claim 1, characterized in that, The residuals of the radiative transfer equations are calculated in the following manner: The radiative transfer equation describing the near-edge observation geometry is as follows: ， in, For radiance, For height, To observe the cosine value of the zenith angle, It is the azimuth angle. Extinction coefficient, The source function; Using automatic differentiation techniques, the partial derivative of the radiance output by the neural network with respect to height is calculated; Calculate its residuals based on the radiative transfer equation. : ， The residual loss term of the radiative transfer equation for: ， in, This represents the number of configuration points.
4. The method according to claim 3, characterized in that, The extinction coefficient The calculation method is as follows: ， in, The Rayleigh scattering coefficient is... This is the ozone absorption coefficient. The aerosol extinction coefficient; The ozone absorption coefficient is: ， in, For temperature-dependent ozone absorption cross section, This is the ozone number density profile.
5. The method according to claim 1, characterized in that, The boundary condition loss term includes: Atmospheric top boundary condition: incident diffuse radiation is zero; Surface boundary conditions: Calculate surface reflected radiation based on the Lambert reflection model or the two-way reflection distribution function; The boundary condition loss term is calculated using the following formula: ， in, The number of boundary points. The neural network predicted the value. These are the boundary condition constraint values.
6. The method according to claim 1, characterized in that, The physical constraint loss term includes: Nonnegativity constraint : Ensure that the predicted ozone concentration is non-negative; Smoothing constraints Regularize the second derivative of the ozone profile; Chapman layer prior constraints The peak ozone concentration is constrained to be within the 20-25 km altitude range; upper boundary constraints : To constrain the concentration of ozone in the upper atmosphere to approach zero; Physical constraint loss term Calculate using the following formula: ， in, , , , These are the weighting coefficients for each constraint term.
7. The method according to claim 1, characterized in that, The composite loss function for: ， in, For data loss items, This is the residual loss term in the radiative transfer equation. For boundary condition loss terms, For physical constraint loss terms, , , , These are the weighting coefficients for each loss term.
8. The method according to claim 1, characterized in that, The physical information neural network consists of one input layer, one output layer, and four hidden layers. The four hidden layers contain 128, 128, 128, and 128 neurons respectively, and the activation function is the hyperbolic tangent function.
9. The method according to claim 1, characterized in that, The edge detector is selected from OMPS-LP, OSIRIS, or SCIAMACHY edge observation modes.
10. The method according to claim 1, characterized in that, The training employs a two-stage optimization strategy, wherein: The first stage uses the Adam optimizer for pre-training with a learning rate set to 10%. -3 ; The second stage uses the L-BFGS optimizer for refined training.