Atmospheric chemical reaction calculation substitute model and construction method

Abstract

The present application relates to the replacement and acceleration of atmospheric chemical reaction calculation in atmospheric chemical model, and particularly relates to a replacement model of atmospheric chemical reaction calculation and a construction method, the model has high precision in reproducing chemical reaction process, in the 19-day offline prediction of 279 prediction variables of the case, the proportion of prediction variable determination coefficient exceeding 0.99 is 96%; the model can realize at least 10-day continuous stable prediction; the model has significant acceleration of chemical calculation: compared with the original GEOS-Chem chemical integrator, the time consumption of ChemKNet calculation on single CPU can be saved by about 79%, and only about 1% of the original calculation time is needed on single GPU, greatly improving the calculation efficiency of large-scale simulation; the model supports coupling and continuous prediction; the model has full-chem chemical mechanism and global simulation covering GEOS-Chem.

Description

Technical Field

[0001] This invention relates to computational substitution and acceleration of atmospheric chemical reactions in atmospheric chemical models, and particularly to a computational substitution model for atmospheric chemical reactions and its construction method. Background Technology

[0002] Air pollution, as a major global environmental and public health challenge, profoundly impacts ecosystem stability, climate system evolution, and human health. Its formation and evolution involve the generation, transformation, and transregional transport of pollutants at local to global scales, encompassing complex nonlinear reaction processes between various forms of chemical substances, including gaseous and particulate matter. Currently, the key pathways and mechanisms of atmospheric chemical reaction networks, especially the formation of secondary pollutants such as ozone and fine particulate matter, are not fully understood. Against this backdrop, atmospheric chemical models (ACMs) have become core tools for revealing pollution processes, interpreting observational data, evaluating control strategies, and predicting future scenarios.

[0003] With the development of Earth System Sciences (ESS), Earth System Models (ESMs) emphasize the fully coupled simulation of atmospheric chemical processes with subsystems such as climate, ecology, and ocean. This requires atmospheric chemical models to characterize more comprehensive chemical reaction mechanisms, often covering hundreds of species and hundreds of reactions. However, due to the limitations of the spatiotemporal resolution of Earth System Models, some processes smaller than the model scale, such as atmospheric fluctuations and cloud physics processes, cannot be directly simulated dynamically and must instead be approximated using parametric methods. Although the improvement of computing power has enabled models to couple more physicochemical processes, the increasing complexity of models has led to a significant increase in computing memory and time costs. At the same time, the "post-Moore's Law" era facing traditional semiconductor technology makes the path of simply relying on hardware upgrades to gain performance gains increasingly difficult.

[0004] In atmospheric chemical models, solving chemical dynamic systems often requires repeatedly solving high-dimensional differential equations, especially in large-scale Earth system simulations where chemical mechanisms may involve hundreds of reactants and related meteorological variables. Using traditional numerical integration methods to calculate chemical states is not only computationally expensive, but the integrator typically accounts for 50% to 90% of the total computational cost during forward integration. Furthermore, because different atmospheric chemical reactions have varying reaction rates (timescales typically ranging from microseconds to days, or even longer), solving differential equations presents rigidity issues, leading to poor stability, requiring extremely small integration steps, and resulting in low efficiency for conventional integration methods. Implicit integration methods (such as the inverse Euler method) can guarantee numerical stability, but they are computationally expensive and require significant computational resources. To improve efficiency, the Rosenbrock integration method and its improvements (such as adaptive step sizes) have been proposed to reduce computational costs while maintaining integration accuracy and stability. However, even with adaptive step sizes, intermediate chemical states still require frequent calculations during integration, resulting in considerable redundant computation.

[0005] In recent years, machine learning (ML) technology has been widely used in the simulation of complex systems, providing a low-computational-cost prediction method for highly complex processes. Research shows that neural networks can approximate solutions to complex partial differential equations and predict the dynamic evolution of high-dimensional systems; by introducing physical constraints, the simulation results can be guaranteed to follow conservation laws and fundamental physical laws, achieving effective coupling between physics and artificial intelligence.

[0006] In atmospheric chemistry modeling simulations, machine learning methods can be used to replace chemical integrators, either offline or online. Existing research has attempted to replace atmospheric chemical integrators with random forests, deep neural networks, or Fourier neural operators to achieve rapid predictions of chemical reaction systems and maintain numerical stability over a certain timeframe. Despite these advancements, issues such as decreased accuracy or error accumulation in trace amounts persist in long-term continuous predictions or under complex chemical mechanisms. Furthermore, most research focuses on offline replacements; seamlessly embedding trained machine learning models into online Earth system models for coupled continuous simulations often faces challenges such as numerical instability and increasing errors during long-term continuous predictions.

[0007] Therefore, there is an urgent need for a method that can improve the efficiency of chemical state calculation while ensuring the accuracy of atmospheric chemical integration, reduce unnecessary intermediate calculation steps, and support the smooth migration of offline trained models to online continuous prediction models, so as to achieve efficient and stable atmospheric chemical simulation.

[0008] There is an urgent need to develop an efficient, stable, and easily coupled alternative for atmospheric chemical reaction calculations based on machine learning methods. An ideal method should possess the following characteristics: (1) High computational efficiency: significantly reducing redundant calculations in traditional numerical integration and achieving rapid updates of chemical states. (2) Long-term numerical stability: maintaining controllable errors in continuous multi-step predictions and avoiding unreasonable concentration outputs. (3) Physicochemical consistency: designing a reasonable machine learning model framework to ensure that predictions conform to basic chemical laws. (4) Smooth online integration capability: supporting a smooth transition from offline training to online coupling and being compatible with the programming language environment on which Earth system models rely.

[0009] Therefore, this invention addresses the urgent need to overcome the bottleneck of chemical calculations in Earth system models. It aims to integrate atmospheric chemical mechanisms with advanced machine learning techniques to develop a new generation of intelligent chemical calculation alternative models, providing key technical support for efficient, accurate, and operational atmospheric environment simulation. Summary of the Invention

[0010] The primary objective of this invention is to provide an alternative model for calculating atmospheric chemical reactions, which is constructed through the following method:

[0011] (1) Under the corresponding chemical reaction mechanism, the input and output data of the chemical integrator are collected, and the collected input and output data are converted into Parquet format for data preprocessing. The data obtained after preprocessing is the training data. The training data is Z-score standardized and converted into sample data in TFRecord format.

[0012] (2) Construct ChemKNet, an atmospheric chemical reaction calculation alternative model based on NeuralODE with Autoencoder, ROR and ResNeXt as basic components. Use the sample data obtained in step (1) for model hyperparameter tuning and model training to obtain the ChemKNet model after hyperparameter tuning and training.

[0013] (3) Replace the atmospheric chemical reaction calculation process in the atmospheric chemical transport model with the ChemKNet model obtained in step (2). The ChemKNet model receives the input data of the atmospheric chemical model. After the input data is calculated by ChemKNet, the simulation output of the atmospheric chemical reaction calculation is obtained.

[0014] (4) Online training process: The simulation output data obtained in step (3) and the output data of the integrator at the current time point of the actual chemical transport model are used to construct a training dataset. The training dataset is used to fine-tune the model. The model fine-tuning process continuously adjusts the learning rate used in the training process according to the loss stability of the ChemKNet model described in step (2) to obtain the ChemKNet model in the prediction form after adaptive coupling back to the atmospheric chemical transport model.

[0015] (5) Using the ChemKNet model obtained in step (4), and based on the chemical step size set by the original atmospheric chemical transport model, output the simulation output of chemical reaction calculation after inputting the input data described in step (1).

[0016] Preferably, the Z-score normalization formula in step (1) is expressed as follows: Among them, among them, For the first before standardization One characteristic, For the first The mean of each feature in the training set, For the first The standard deviation of each feature in the training set For the standardized first One characteristic.

[0017] Preferably, the input data in step (1) includes meteorological variables, fixed chemical species concentration variables, and active chemical species concentration variables.

[0018] Preferably, the meteorological variables include 9, the fixed chemical species concentration variables include 4, and the active chemical species concentration variables include 279.

[0019] Preferably, the meteorological variables are TEMP, PRESS, NUMDEN, H2O, SUNCOS, RELHUM, UVALBEDO, PEDGE, and OPTD; the fixed chemical species concentration variables are H2, N2, O2, and RCOOH; TEMP is temperature, PRESS is air pressure, NUMDEN is air number density, H2O is water vapor number concentration, SUNCOS is the cosine of the solar zenith angle, RELHUM is relative humidity, UVALBEDO is ultraviolet surface albedo, PEDGE is moist air pressure at the bottom of the grid, and OPTD is visible optical thickness; H2 is hydrogen, N2 is nitrogen, O2 is oxygen, and RCOOH is organic acid.

[0020] Preferably, step (2) of the model hyperparameter tuning and model training includes the following steps:

[0021] S1. The concentration variables of active chemical species are subjected to feature extraction and dimensionality reduction processing through the encoder in the autoencoder to extract low-dimensional potential features reflecting the chemical reaction system. The expression is as follows: ;

[0022] in, The corresponding active chemical species concentration variable in the model, for 3D real vector space, For variables Dimensions The encoded latent space feature vector. To After encoding The dimension, in which As one of the model structure hyperparameters, denoted as the encoded dimension, This represents the encoder weight matrix, where the number of rows corresponds to the input feature dimension. The number of columns represents the feature dimension after encoding. , This indicates its transpose. , It is the bias vector;

[0023] S2, the concentration variables of active chemical species are obtained after encoding. , It is concatenated with fixed chemical species concentration variables and meteorological variables as input to the first ResNeXt component. The expression that combines the fixed chemical species concentration variable and the meteorological variable is shown below: ;

[0024] in, , For a fixed chemical species concentration variable, , Corresponding meteorological variables, , , and They represent the corresponding variables respectively , and Dimensions Features after splicing This indicates a concatenation operation along the feature dimension;

[0025] S3. In the ResNeXt component. After multiple parallel fully connected layer branches, each branch is followed by a ReLU activation function to introduce non-linear expressive power. The outputs of all branches are concatenated and then integrated through a fully connected layer and a ReLU activation function; the expression is shown below:

[0026] ;

[0027] ;

[0028] in, This indicates the number of parallel branches, or the number of branches. Let be one of the model structure hyperparameters, denoted as the number of parallel branches of ResNeXt; for the th Parallel branches , This is the weight matrix of the fully connected layer in this branch. Let it be its transpose matrix. This is the bias vector of the fully connected layer in this branch. The activation function is used, and the branch output is... , The feature dimension of the output after parallel branching is determined by the number of neurons in the parallel branch and is one of the model structure hyperparameters, denoted as the number of neurons in the parallel branch of ResNeXt. , This indicates that a concatenation operation is performed along the feature dimension. This is the weight matrix of the fully connected layer used to integrate the spliced ​​features. Let it be its transpose matrix. For the corresponding bias vector , for The dimension is determined by the weight matrix of the fully connected layer that integrates and splices features. The value determined by this is one of the model structure hyperparameters, denoted as the number of neurons in the integrated fully connected layer;

[0029] S4 After passing through a fully connected layer, the output dimension is mapped back to... Dimensions:

[0030] ;

[0031] in, These are the features obtained after the preceding integration operations; Here is the weight matrix of the fully connected layer. Let it be its transpose matrix. This corresponds to the bias vector; since this layer is used to map the feature dimensions back... The dimension, therefore , The output features are those obtained after dimension back mapping;

[0032] S5. Each ResNeXt component has its own local residual connections to ensure stable feature propagation within a single component. The formula is as follows: ;

[0033] in, , After local residual connection, and The result obtained by adding them together;

[0034] The process from S1 to S5 involves data in... The computational operations performed within the component can be simplified as follows: ;

[0035] S6 The next ResNeXt with the same structure will be expressed as: ;

[0036] in, , express After the second The output obtained after the component;

[0037] How many specifically? It can be obtained by setting the model structure hyperparameters and passing through the set n ResNeXt modules. Then, perform layer normalization, as shown in the formula below:

[0038] ;

[0039] in, Presentation layer normalization operation, and For trainable scaling and translation parameters, ⊙ denotes element-wise multiplication. To prevent small constants from being divided by zero, Standard deviation, The mean, These are the features obtained after layer normalization;

[0040] S7, obtained Then, feature fusion is performed through global residual connections; the global residual connections contain a set... There are 1 ResNeXt component, each with local residual connections, thus forming a Relational Oorestable Relation (ROR); its formula is expressed as:

[0041] ;

[0042] in, These are low-dimensional latent features obtained by encoding the concentration variables of active chemical species in S1; The final output feature obtained by S6 is the feature after being processed and layer-normalized by multiple ResNeXt components inside the ROR structure. The ROR output features are obtained through global residual connections. ;

[0043] S8. Finally, the decoder outputs a prediction, remapping the features in the high-dimensional latent space back to the original chemical species variable dimension space, as shown in the following formula:

[0044] ;

[0045] in, Here is the weight matrix of the fully connected layer of the decoder. Let it be its transpose matrix. For the corresponding bias vector; the decoder will Mapping to the prediction output space yields the prediction result. , or denoted as Therefore ,in, To define the dimensions of the output predictor variables, i.e., the original active chemical species concentration variables;

[0046] Then, inverse standardization is performed, expressed by the following formula:

[0047] ;

[0048] in, For the first The predicted value of each feature, The first one obtained after inverse standardization Predicted concentration values ​​of a chemical substance, For the first The mean of each feature in the training set, For the first The standard deviation of each feature in the training set.

[0049] A second objective of this invention is to provide a method for constructing the aforementioned alternative model for atmospheric chemical reactions, comprising the following steps:

[0050] (1) Under the corresponding chemical reaction mechanism, the input and output data of the chemical integrator are collected, and the collected input and output data are converted into Parquet format for data preprocessing. The data obtained after preprocessing is the training data. The training data is Z-score standardized and converted into sample data in TFRecord format.

[0051] (2) Construct ChemKNet, an atmospheric chemical reaction calculation alternative model based on NeuralODE with Autoencoder, ROR and ResNeXt as basic components. Use the sample data obtained in step (1) for model hyperparameter tuning and model training to obtain the ChemKNet model after hyperparameter tuning and training.

[0052] (3) Replace the atmospheric chemical reaction calculation process in the atmospheric chemical transport model with the ChemKNet model obtained in step (2). The ChemKNet model receives the input data of the atmospheric chemical model. After the input data is calculated by ChemKNet, the simulation output of the atmospheric chemical reaction calculation is obtained.

[0053] (4) Online training process: The simulation output data obtained in step (3) and the output data of the integrator at the current time point of the actual chemical transport model are used to construct a training dataset. The training dataset is used to fine-tune the model. The model fine-tuning process continuously adjusts the learning rate used in the training process according to the loss stability of the ChemKNet model described in step (2) to obtain the ChemKNet model in the prediction form after adaptive coupling back to the atmospheric chemical transport model.

[0054] (5) Using the ChemKNet model obtained in step (4), and based on the chemical step size set by the original atmospheric chemical transport model, output the simulation output of chemical reaction calculation after inputting the input data described in step (1).

[0055] The beneficial effects of this invention are as follows: This invention provides an alternative model for calculating atmospheric chemical reactions, which accurately reproduces chemical reaction processes. By replacing the chemical reaction calculation part of the atmospheric chemical transport model (such as GEOS-Chem) with the machine learning model ChemKNet, the model can accurately predict the concentration evolution of chemical species such as O3, NO, NO2, and OH while maintaining the accuracy of the original chemical integrator. In the case study, 96% of the 279 predictor variables had a coefficient of determination greater than 0.99 in 19 days of offline prediction. After coupling the model back to GEOS-Chem, it can achieve continuous and stable predictions for at least 10 days (720 steps). The model significantly accelerates chemical calculations: compared to the original GEOS-Chem chemical integrator, ChemKNet can save approximately 79% of the computation time on a single CPU and only about 1% of the original computation time on a single GPU, greatly improving the computational efficiency of large-scale simulations. The model supports coupling and continuous prediction: it is not only applicable to offline chemical box prediction, but also, after coupling ChemKNet back to GEOS-Chem, enables continuous temporal evolution prediction of the global three-dimensional atmospheric chemical state, maintaining trends and distributions highly consistent with the original model. The model possesses full-chem chemical mechanisms covering GEOS-Chem and global simulation: the ChemKNet alternative used in this invention can cover multiple types of chemical reactions, is applicable to full-chem chemical mechanisms, and is suitable for accelerating and replacing chemical processes under global simulation environments, exhibiting strong versatility and generalizability. Attached Figure Description

[0056] Figure 1 A flowchart of a computational alternative model for atmospheric chemical reactions;

[0057] Figure 2 This is a structural diagram of the GEOS-Chem case model in this embodiment;

[0058] Figure 3 A scatter plot showing the correlation between offline concentration predictions of chemical species from the ChemKNet model and concentration predictions from GEOS-Chem chemical reaction simulations;

[0059] Wherein, a is a scatter plot of the correlation between predicted and actual ozone values; b is a scatter plot of the correlation between predicted and actual nitric oxide values; c is a scatter plot of the correlation between predicted and actual nitrogen dioxide values; and d is a scatter plot of the correlation between predicted and actual hydroxyl radical values.

[0060] Figure 4 The 10-day zonal ozone concentration distribution is predicted online by the original GEOS-Chem model and coupled with the ChemKNet model.

[0061] The upper part of the figure represents the region from the ground to near-zero atmospheric pressure, while the lower part represents the region from the ground to near the stratosphere. Figure a shows the zonal average ozone concentration distribution of the GC, figure b shows the zonal average ozone concentration distribution of the GC coupled with ChemKNet, figure c shows the absolute error distribution between the GC and the GC coupled with ChemKNet, figure d shows the fractional error distribution between the GC and the GC coupled with ChemKNet, figure e shows the zonal average ozone concentration distribution of the GC, figure f shows the zonal average ozone concentration distribution of the GC coupled with ChemKNet, figure g shows the absolute error distribution between the GC and the GC coupled with ChemKNet, and figure h shows the fractional error distribution between the GC and the GC coupled with ChemKNet.

[0062] Figure 5 The distribution map of 10-day average ozone concentration near the ground is predicted online after the original GEOS-Chem model is used to predict and the ChemKNet model is coupled.

[0063] Where, a is the global time-averaged ozone concentration distribution predicted by the near-surface GC, b is the global time-averaged ozone concentration distribution predicted by the near-surface GC coupled with ChemKNet, c is the absolute error distribution between the GC and the GC coupled with ChemKNet, and d is the fractional error distribution between the GC and the GC coupled with ChemKNet.

[0064] Figure 6 The original GEOS-Chem model was used to predict the 10-day average ozone concentration distribution at 500 hPa, and then the ChemKNet model was coupled to predict it online.

[0065] Wherein, a is the global time-averaged ozone concentration distribution predicted by GC at 500 hPa, b is the global time-averaged ozone concentration distribution predicted by GC coupled with ChemKNet at 500 hPa, c is the absolute error distribution between GC and GC coupled with ChemKNet, and d is the fractional error distribution between GC and GC coupled with ChemKNet.

[0066] Figure 7 Comparison of regional time series predictions of ozone and related substances made online after the original GEOS-Chem model and coupled ChemKNet model;

[0067] Among them, a) is a comparison of ozone time series predictions between the GC in the Beijing area and the GC coupled with ChemKNet; b) is a comparison of ozone time series predictions between the GC in the Legionovo area and the GC coupled with ChemKNet; c) is a comparison of ozone time series predictions between the GC in the Los Angeles area and the GC coupled with ChemKNet; d) is a comparison of nitric oxide time series predictions between the GC in the Beijing area and the GC coupled with ChemKNet; e) is a comparison of nitric oxide time series predictions between the GC in the Legionovo area and the GC coupled with ChemKNet; f) is a comparison of nitric oxide time series predictions between the GC in the Los Angeles area and the GC coupled with ChemKNet; g) is a comparison of ozone time series predictions between the GC in the Beijing area and the GC coupled with ChemKNet; and g) is a comparison of nitric oxide time series predictions between the GC in the Beijing area and the GC coupled with ChemKNet. Comparison of nitrogen dioxide time series predictions between GC in the Beijing region and GC coupled with ChemKNet; h is the comparison of nitrogen dioxide time series predictions between GC in the Legionovo region and GC coupled with ChemKNet; i is the comparison of nitrogen dioxide time series predictions between GC in the Los Angeles region and GC coupled with ChemKNet; j is the comparison of hydroxyl radical time series predictions between GC in the Beijing region and GC coupled with ChemKNet; k is the comparison of hydroxyl radical time series predictions between GC in the Legionovo region and GC coupled with ChemKNet; l is the comparison of hydroxyl radical time series predictions between GC in the Los Angeles region and GC coupled with ChemKNet.

[0068] Figure 8 This is a schematic diagram illustrating the acceleration effect achievable in this embodiment; Detailed Implementation

[0069] The present invention will be further described in detail below with reference to the embodiments.

[0070] In the following examples, the atmospheric chemical transport model GEOS-Chemv14.1.1 was selected. This model operates using the full-chem chemical reaction mechanism, is a global model, and has latitude and longitude resolutions of 5° and 4°, respectively. The mass transport step size is 10 min, and the chemical step size is 20 min. Meteorological variables and chemical concentration data related to the full-chem chemical mechanism calculations are extracted every 20 min.

[0071] In the following embodiments, the fixed chemical species concentration variable is the concentration variable whose concentration remains unchanged during the chemical reaction calculation process, and the active chemical species concentration variable is the concentration variable whose concentration value changes accordingly during the chemical reaction calculation process.

[0072] In the following embodiments, an atmospheric chemical reaction calculation alternative model based on the neural ordinary differential equation is used to efficiently simulate the chemical calculation process in the atmospheric chemical transport model, thereby achieving rapid prediction and simulation of atmospheric chemical substance concentrations.

[0073] Example 1: An Alternative Model for Calculating Atmospheric Chemical Reactions

[0074] like Figure 1 As shown, an alternative model for atmospheric chemical reaction calculations is proposed to accelerate the chemical calculation process. The method includes:

[0075] 1. Collect the input and output data of the chemical integrator, as well as the variables required for relevant chemical reaction calculations. Specifically, this data includes the meteorological variables, fixed chemical species concentration variables, and active chemical species concentration variables required for calculating the chemical reaction system under the full-chem chemical mechanism; the relevant variables required for calculating the numerical evolution of chemical species concentrations under this reaction mechanism; and the chemical species concentration data calculated by the chemical integrator after setting the chemical step size (the training target for the machine learning model). The collected data is converted to Parquet format for data preprocessing. Then, the Parquet format data is Z-score standardized and converted to TFRecord format for efficient training of the machine learning model. The formula for Z-score standardization is: ;in, For the first before standardization One characteristic, For the first The mean of each feature in the training set, For the first The standard deviation of each feature in the training set For the standardized first One characteristic.

[0076] Because a single chemical integral calculation involves 145,728 grid points, the extracted data consists of 145,728 × 2 rows. Each row contains 9 meteorological state variables, 4 fixed chemical species concentration variables, and 279 active chemical species concentration variables. Processing the data row by row is very time-consuming. Therefore, the data is first converted to Parquet format and then processed column by column for efficient processing. After obtaining the Parquet data, it is standardized and then converted to TFRecord format for machine learning training.

[0077] The training data mentioned above consists of nine meteorological variables: TEMP (temperature), PRESS (air pressure), NUMDEN (air number density), H2O (water vapor number concentration), SUNCOS (solar zenith angle cosine), RELHUM (relative humidity), UVALBEDO (ultraviolet albedo), PEDGE (moist air pressure at the bottom of the grid), and OPTD (visible optical thickness).

[0078] Four fixed chemical species concentration variables: H2, N2, O2, and RCOOH;

[0079] Concentration variables of 279 active chemical species: O3, NO, NO2, OH, A3O2 (propane primary peroxide radical), ACET (acetone), ACTA (acetic acid), AERI (dissolved iodine), ALD2 (acetaldehyde), ALK4 (≥C4 alkanes), AONITA (aerosol-phase organic nitrates derived from aromatic hydrocarbon precursors), AROMRO2 (universal peroxide radical from aromatic hydrocarbon oxidation), AROMP4 (universal C4 product from aromatic hydrocarbon oxidation), AROMP5 (universal C5 product from aromatic hydrocarbon oxidation), ATO2 (acetone peroxide radical), ATOOH (peroxide of ATO2), B3O2 (propane secondary peroxide radical), BALD (benzaldehyde and methylbenzaldehyde), BENZ (benzene), BENZO (phenoxy radical), BENZO2 (benzene peroxide radical), BENZP (hydroperoxide generated from BENZO2), Br, Br2, BrCl, BrNO2, BrNO3, BrO, BRO2 (benzene peroxide radical), Br2 (acetone ... Oxygen free radicals), BrSALA (fine sea salt bromine), BrSALC (crude sea salt bromine), BZCO3 (benzoyl peroxy radical), BZCO3H (perbenzoic acid), BZPAN (perbenzoyl nitrate), C2H2 (acetylene), C2H4, C2H6, C3H8, C4HVP1 (C4 hydroxy-vinyl-peroxy radicals from HPALDs), C4HVP2 (C4 hydroxy-vinyl-peroxy radicals from HPALDs), C4HVP2 (C4 hydroxy-vinyl-peroxy radicals from HPALDs). oxyradicalsfromHPALDs), CCl4, CFC11 (CCl3F; CFC-11, R-11, Freon11), CFC12 (CCl2F2; CFC-12, R-12, F reon12), CFC113 (C2Cl3F3; CFC-113, Freon113), CFC114 (C2Cl2F4; CFC-114, Freon114), CFC115 (C2ClF5;CFC-115, Freon115), CH2Br2, CH2Cl2, CH2I2, CH2IBr, CH2ICl, CH2O, CH2OO (ethylene crescendo intermediate), CH3Br, CH3CCl3, CH3CHOO (propylene crescendo intermediate), CH3Cl, CH3I, CH4, CHBr3, CHCl3, Cl, Cl2, Cl2O2, ClNO2, ClNO3, ClO, ClOO, CO, CO2, CSL (cresol and xylene), DMS (dimethyl sulfide), EOH (ethanol), ETHLN (acetaldehyde nitrate), ETHN (stable hydroxyethyl nitrate), ETHP (stable hydroxyethyl peroxide). ETNO3 (ethyl nitrate), ETO (hydroxyalkoxyethylene radical), ETOO (hydroxyperoxyethylene radical, generated by the reaction of ethylene and hydroxyl radicals), ETO2 (ethyl peroxy radical), ETP (ethyl hydrogen peroxide), GLYC (glyoxal), GLYX (glyoxal), H, H1211 (CBrClF2; H-1211), H1301 (CBrF3; H-1301), H2402 (C2Br2F4; H-2402), H2O2, HAC (hydroxyacetone), HBr, HC5A (C5H8O2; Isoprene-4, 1-hydroxyaldehyde), HCFC123 (C2HCl2F3; H CFC-123, R-123, Freon123), HCFC141b (C(CH3)Cl2F; HCFC-141b, R-141b, Freon141b), HCFC142b (C(CH3)ClF2; HCFC-142b, R-142b, Freon142b), HCFC22 (CHClF2; HCFC-22, R-22, Freon22), HCl, HCOOH (formic acid), HI, HMHP (hydroxymethyl hydroperoxide), HMML (C4H6O3; Hydroxymethyl-methyl-a-lactone), HMS (hydroxymethylsulfonate), HNO2 HNO3, HNO4, HO2 (peroxyhydroxyl radical), HOBr, HOCl, HOI, HONIT (second-generation monoterpenoid organic nitrates), HPALD1 (O=CHC(CH3)=CHCH2OOH; d-4,1-C5-hydroperoxyaldehyde), HPALD1OO (peroxy radical generated from HPALD1), HPALD2 (HOOCH2C(CH3)=CHCH=O; d-1,4-C5-hydroperoxyaldehyde), HPALD2OO (peroxy radical generated from HPALD2), HPALD3 (O=CHC(CH3)OOHCH=CH2;b-2,1-C5-hydroperoxyaldehyde), HPALD4 (CH2=C(CH3)CHOOHCH=O; b-3,4-C5-hydroperoxyaldehyde), HPETHNL (hydroxyperoxyacetaldehyde), I, I2, I2O2, I2O3, I2O4, IBr, ICHE (isoprene hydroxy-carbonyl-epoxide), ICHOO (peroxy radical generated by IEPOXD), ICl, ICN (isoprene carbonyl nitrates), ICNOO (peroxy radical generated by ICN), ICPDH (isoprene dihydroxyhydroperoxycarbonyl compounds), IDC (isoprene dicarbonylation) Compounds), IDCHP (isoprene dicarbonyl-hydroxy-dihydroperoxide), IDHDP (isoprene dihydroxy-dihydroperoxide), IDHNBOO (peroxide radical generated from INPB), IDHNDOO1 (peroxide radical generated from INPD), IDHNDOO2 (peroxide radical generated from INPD), IDHPE (isoprene dihydroxyhydroperoxide epoxide), IDN (isoprene dinitrate ester merged species), IDNOO (peroxide radical generated from IDN), IEPOXA (trans-β-isoprene epoxy glycol), IEPOXAOO (peroxide radical generated from IEPOXA), IEPOXB (cis-β-isoprene epoxy glycol). IEPOXBOO (peroxy radical generated from IEPOXB), IEPOXD (δ-isoprene epoxy glycol), IHN1 (C5H9NO4; Isoprene-d-4-hydroxy-1-nitrate), IHN2 (C5H9NO4; Isoprene-b-1-hydroxy-2-nitrate), IHN3 (C5H9NO4; Isoprene-b-4-hydroxy-3-nitrate), IHN4 (C5H9NO4; Isoprene-d-1-hydroxy-4-nitrate), IHOO1 (peroxyradical fr omOHadditiontoisopreneatC1), IHOO4 (peroxyradical from OHadditiontoisopreneatC4), IHPNBOO (peroxy radical generated from INPB), IHPNDOO (peroxy radical generated from INPD), IHPOO1 (peroxy radical generated from ISOPOOH), IHPOO2 (peroxy radical generated from ISOPOOH), IHPOO3 (peroxy radical generated from ISOPOOH), INA (peroxy radical generated from INO2D), INDIOL (general product of hydrolysis of aerosol-phase organic nitrate esters), INO (INO;Nitrosyliodide), INO2B (beta-peroxyradicals from isoprene + NO3), INO2D (delta-peroxyradicals from isoprene + NO3), INPB (isoprene β-hydroperoxynitrate merged species), INPD (isoprene δ-hydroperoxynitrate merged species), IO (IO; Iodinemonoxide), IONITA (aerosol isoprene precursor organic nitrate), IONO (IONO; Nitryliodide), IONO2 (IONO2;Iodinenitrate), IPRNO3 (isopropyl nitrate), ISALA (fine-modal sea salt iodine), ISALC (coarse-modal sea salt iodine), ISOP (isoprene), ISOPNOO1 (peroxy radical generated from IHN2), ISOPNOO2 (peroxy radical generated from IHN3), ITCN (tetrapene carbonyl nitrate merged species), ITHN (tetrapene hydroxy nitrate merged species), KO2 (RO2 from 3ketones), LBRO2H (virtual oxidant used to track the oxidation of BRO2 by HO2). Species), LBRO2N (virtual species used to track the oxidation of BRO2 by NO), LIMO (limonene), LIMO2 (RO2 from LIMO), LISOPOH (virtual species used to track the oxidation of isoprene (ISOP) by OH), LISOPNO3 (virtual species used to track the oxidation of isoprene (ISOP) by NO3), LNRO2H (virtual species used to track the oxidation of NRO2 by HO2), LNRO2N (virtual species used to track the oxidation of NRO2 by NO), LTRO2H (virtual species used to track the oxidation of TRO2 by HO2), L TRO2N (a virtual species used to track the oxidation of TRO2 by NO), LVOC (a low-volatility non-IEPOX aerosol phase product generated by the oxidation of ISOPOOH (RIP)), LVOCOA (a low-volatility non-IEPOX aerosol phase product generated by the oxidation of ISOPOOH (RIP)), LXRO2H (a virtual species used to track the oxidation of XRO2 by HO2), LXRO2N (a virtual species used to track the oxidation of XRO2 by NO), MACC (methacrylaldehyde), MACC1OO (a peroxyacyl radical generated by the reaction of MACC with OH), MACC1OOH (from... MACR generated peracid), MCRNO2 (product of the reaction of MCRHN with OH), MAP (peracetic acid), MCO3 (peroxyacetyl radical), MCRDH (dihydroxy-methacrylaldehyde), MCRENOL (enols merged species generated by MVK / MACR), MCRHN (hydroxynitrates from MACR), MCRHNB (hydroxynitrates from MACR), MCRHP (hydroxy-hydroperoxide type MACR), MCRHOO (peroxy radical generated by the reaction of MACR with OH), MCT (methylcatechols), MEK (RC(O)R;Methyl ethyl ketone (MEO), MENO3 (methyl nitrate), MGLY (methylglyoxal), MO2 (methyl peroxide radical), MOH (methanol), MONITA (aerosol-phase monoterpene precursor organic nitrate), MONITS (saturated first-generation monoterpene organic nitrate), MONITU (unsaturated first-generation monoterpene organic nitrate), MP (methyl hydroperoxide), MPAN (peroxymethacryloyl nitrate), MPN (methyl peroxy nitrate), MSA (methanesulfonic acid), MTPA (combined monoterpenes: α-pinene, β-pinene, sabinene, carene), MTPO (other monoterpenes: terpinene, terpinolene, myrcetin) ne, ocimene, other monoterpenes), MVK (methyl vinyl ketone), MVKDH (dihydroxy-methyl vinyl ketone), MVKHC (MVK hydroxy-carbonyl product), MVKHCB (MVK hydroxy-carbonyl product), MVKHP (MVK hydroxy-hydroperoxide), MVKN (MVK-derived hydroxynitrate), MVKOHOO (peroxide radical generated by the reaction of MVK with OH), MVKPC (MVK hydroperoxy-carbonyl product), N (atomic nitrogen), N2O, N2O5, NAP (naphthalene), NIT (fine-mode inorganic nitrate), NITs (coarse-mode inorganic nitrate), NO3 (nitrate radical), NPHEN (Nitrophenols), NPRNO3 (n-propyl nitrate), NRO2 (peroxide radical generated by naphthalene oxidation), O (ground state atomic oxygen), O1D (excited state atomic oxygen), OClO (chlorine dioxide), OCS (carbonyl sulfide), OIO (iodine dioxide), OLND (NO3-olefin adduct from monoterpenes), OLNN (NO3 adduct from monoterpenes), OTHRO2 (other C2 peroxide radicals not generated by C2H6 oxidation), PAN (peroxyacetyl nitrate), PHEN (phenol), PIO2 (peroxide radical generated by MTPA), PIP (peroxides generated by MTPA), PO2 (peroxide radicals from propylene), PP (peroxides generated by PO2), P PN (peroxypropionyl nitrate), PRN1 (peroxy radical generated from propylene and NO3), PROPNN (acetone nitrate), PRPE (≥C3 olefins), PRPN (peroxide generated from PRN1), PYAC (pyruvic acid), R4N1 (peroxy radical generated from R4N2), R4N2 (≥C4 alkyl nitrates), R4O2 (peroxy radical generated from ALK4), R4P (peroxide generated from R4O2), RA3P (peroxide generated from A3O2), RB3P (peroxide generated from B3O2), RCHO (≥C3 aldehydes), RCO3 (peroxypropionyl radical), RIPA (HOCH2C(OOH)(CH3)CH=CH2;1,2-ISOPOOH), RIPB (HOCH2C(OOH)(CH3)CH=CH2; 4,3-ISOPOOH), RIPC (C5H10O3; d(1,4)-ISOPOOH), RIPD (C5H10O3; d(4,1)-ISOPOOH), ROH (>C2 alcohols), RP (peroxides generated from RCO3), SALAAL (cumulative modal sea salt aerosol basicity), SALCAL (coarse modal sea salt aerosol basicity) The concentrations of these 279 active chemical species are: SALACL (fine-mode chloride), SALCCL (coarse-mode chloride), SO2, SO4 (sulfates), SO4s (sulfates on sea salt), SOAGX (glyoxal in aerosol phase), SOAIE (IEPOX in aerosol phase), TOLU (toluene), TRO2 (peroxy radical generated by toluene oxidation), XRO2 (peroxy radical generated by xylene oxidation), and XYLE (xylene).

[0080] 2. Construct ChemKNet (Chemical Kinetics Network), a NeuralODE-based alternative model for atmospheric chemical reaction calculation, using Autoencoder, ROR (Residual-of-Residual), and ResNeXt (Aggregated Residual Transformations for Deep Neural Networks) as basic components. Hyperparameter tuning and training of the model were performed using sample data in TFRecord format, resulting in the hyperparameter-tuned and trained ChemKNet model. The constructed ChemKNet model, by inputting three variables—active chemical species concentration, fixed chemical species concentration, and meteorological variables—outputs the active chemical species concentration for the next chemical step, as follows:

[0081] S1. The concentration variables of active chemical species are subjected to feature extraction and dimensionality reduction processing through the encoder in the autoencoder to extract low-dimensional potential features reflecting the chemical reaction system. The expression is as follows: ;

[0082] in, The corresponding active chemical species concentration variable in the model, for 3D real vector space, For variables Dimensions The encoded latent space feature vector. To After encoding The dimension, in which As one of the model structure hyperparameters, denoted as the encoded dimension, This represents the encoder weight matrix, where the number of rows corresponds to the input feature dimension. The number of columns represents the feature dimension after encoding. , This indicates its transpose. , It is the bias vector;

[0083] S2, the concentration variables of active chemical species are obtained after encoding. , It is concatenated with fixed chemical species concentration variables and meteorological variables as input to the first ResNeXt component. The expression that combines the fixed chemical species concentration variable and the meteorological variable is shown below: ;

[0084] in, , For a fixed chemical species concentration variable, , Corresponding meteorological variables, , , and They represent the corresponding variables respectively , and Dimensions Features after splicing This indicates a concatenation operation along the feature dimension;

[0085] S3. In the ResNeXt component. After multiple parallel fully connected layer branches, each branch is followed by a ReLU activation function to introduce non-linear expressive power. The outputs of all branches are concatenated and then integrated through a fully connected layer and a ReLU activation function; the expression is shown below:

[0086] ;

[0087] ;

[0088] in, This indicates the number of parallel branches, or the number of branches. Let be one of the model structure hyperparameters, denoted as the number of parallel branches of ResNeXt; for the th Parallel branches ( ), This is the weight matrix of the fully connected layer in this branch. Let it be its transpose matrix. This is the bias vector of the fully connected layer in this branch. The activation function is used, and the branch output is... , The feature dimension of the output after parallel branching is determined by the number of neurons in the parallel branch and is one of the model structure hyperparameters, denoted as the number of neurons in the parallel branch of ResNeXt. , This indicates that a concatenation operation is performed along the feature dimension. This is the weight matrix of the fully connected layer used to integrate the spliced ​​features. Let it be its transpose matrix. For the corresponding bias vector , for The dimension is determined by the weight matrix of the fully connected layer that integrates and splices features. The value determined by this is one of the model's structural hyperparameters, denoted as the number of neurons in the integrated fully connected layer.

[0089] S4, then After passing through a fully connected layer, the output dimension is mapped back to... Dimensions:

[0090] ;

[0091] in, These are the features obtained after the preceding integration operations; Here is the weight matrix of the fully connected layer. Let it be its transpose matrix. This corresponds to the bias vector; since this layer is used to map the feature dimensions back... The dimension, therefore , The output features are those obtained after dimension back mapping;

[0092] S5. Each ResNeXt component has its own local residual connections to ensure stable feature propagation within a single component. The formula is as follows: ;

[0093] in, , After local residual connection, and The result obtained by adding them together;

[0094] The process from S1 to S5 involves data in... The computational operations performed within the component can be simplified as follows: ;

[0095] S6 The next ResNeXt with the same structure will be expressed as: ;

[0096] in, , express After the second The output obtained after the component;

[0097] How many specifically? This can be set by the model structure hyperparameters and denoted as the number of ResNeXt. Two are used in this example. The component is obtained after passing through two specified ResNeXt components. At the continuous-time modeling level, multiple cascaded ResNeXt components collectively constitute a discrete approximation NeuralODE component, used to learn the time-step integral increment during the chemical integration process. That is, differential operators In the integral result with step size Δt, where, Indicates the system in time The state variables under, Represents a time variable. Using integral variables embodies the modeling concept of continuous-time dynamics. Then, a layer normalization operation is performed, and the normalized result is compared with the result obtained in the initial encoding stage. The characteristics are added element by element to complete one round of chemical kinetic characteristic evolution, as expressed in the following formula:

[0098] ;

[0099] in, Presentation layer normalization operation, and For trainable scaling and translation parameters, ⊙ denotes element-wise multiplication. To prevent small constants from being divided by zero, Standard deviation, The mean, These are the features obtained after layer normalization;

[0100] S7, obtained Then, feature fusion is performed through global residual connections; each global residual connection contains two defined ResNeXt components, and each ResNeXt component has local residual connections, thus forming the ROR structure with ResNeXt components as the basic residual components; its formula is expressed as:

[0101] ;

[0102] in, These are low-dimensional latent features obtained by encoding the concentration variables of active chemical species in S1; The final output feature obtained by S6 is the feature after being processed and layer-normalized by multiple ResNeXt components inside the ROR structure. The ROR output features are obtained through global residual connections. ;

[0103] S8. Finally, the decoder outputs predictions, remapping the features in the high-dimensional latent space back to the original chemical species variable dimension space, as shown in the following formula:

[0104] ;

[0105] in, This is the weight matrix of the fully connected layer of the decoder. Let it be its transpose matrix. For the corresponding bias vector; the decoder will Mapping to the prediction output space yields the prediction result. , or denoted as Therefore ,in, The dimension of the output predictor variable, i.e. the original active chemical species concentration variable, is determined by the model structure hyperparameters and is denoted as the decoded dimension.

[0106] Then, inverse standardization is performed, expressed by the following formula:

[0107] ;

[0108] in, For the first The predicted value of each feature, The first one obtained after inverse standardization Predicted concentration values ​​of a chemical substance, For the first The mean of each feature in the training set, For the first The standard deviation of each feature in the training set.

[0109] In this embodiment, by tuning the model hyperparameters, the machine learning model can maintain accuracy in understanding the chemical mechanisms of the cases while achieving higher computational efficiency than the chemical calculation process in the atmospheric chemical transport model. As mentioned earlier, the model structure hyperparameters mainly include the encoded dimension and the decoded dimension of the Autoencoder. The decoded dimension is generally set according to the number of chemical substance concentration variables to be predicted. Other model hyperparameters include the number of ResNeXts in ROR, the number of parallel branches of ResNeXts, the number of neurons in the parallel branches of ResNeXts, and the number of neurons in the fully connected layer that the hidden features after merging parallel paths need to pass through.

[0110] The optimal combination of hyperparameters can be found through Bayesian optimization or grid search. After determining the model hyperparameters, the model is trained using TFRecord sample data. The training process uses the mean squared error loss function, and the concentration values ​​of 279 chemical species predicted by the machine learning model are used in the loss calculation. Concentration data of the 279 active chemical species from the output of the original chemical integrator. The specific formula is as follows ;in, For the first Predicted values ​​for each chemical species, For the first The output value of the original chemical integrator This is the calculated mean square error.

[0111] 3. Replace the chemical calculation process of the atmospheric chemical transport model with the trained ChemKNet model. The alternative model receives the input data of the atmospheric chemical model, and the input data is calculated by the alternative model to obtain the simulation output of atmospheric chemical reaction calculation.

[0112] like Figure 2As shown, the calculation sequence of the GEOS-Chem model is as follows: model initialization → mass transport process → calculation of planetary boundary layer height → calculation of dry deposition rate → emission process → mixing within the planetary boundary layer → cloud convection → chemical integrator calculation process (in this case, cloud water chemistry, gas phase chemistry, heterogeneous chemistry, and photochemistry) → wet deposition → model diagnosis. After model diagnosis, the model returns to the mass transport process to start the simulation of the next time step. The machine learning model ChemKNet replaces the chemical calculation part (chemical reaction constants and integrator forward integration process) in the GEOS-Chem model. The trained ChemKNet model calls a Python program in the GEOS-Chem model to input GEOS-Chem data into the Python program. After data standardization, the ChemKNet model predicts the active chemical species concentration values ​​(standardized values) for the next chemical step. Then, the values ​​are destandardized, and finally, the predicted values ​​are returned to the GEOS-Chem model to update the chemical species concentration values ​​at the target time.

[0113] 4. In this embodiment, in order to ensure that the coupled machine learning model can make stable and continuous predictions, an online training process is also required. The online training process involves constructing a training set for model fine-tuning by combining the simulation output predicted by the alternative model ChemKNet with the output of the integrator at the current time point in the actual chemical calculation process. During the model fine-tuning process, the learning rate is gradually reduced according to the stability of the loss. By continuously reducing the learning rate during the online training process, the model continuously learns from its own prediction errors and gradually gets rid of the need for "large parameter tuning to ensure stable predictions". This allows the model to gradually reach more stable weights, thereby achieving stable and continuous multi-step predictions.

[0114] 5. After online training is completed, the alternative model ChemKNet will completely replace the atmospheric chemical reaction calculation part of the atmospheric chemical transport model. The model can output the simulation output of the chemical reaction calculation in the original atmospheric chemical transport model according to the chemical step size set by the original atmospheric chemical transport model, through meteorological variables, chemical species concentration data and other relevant variables required for the calculation of chemical species concentration numerical evolution under the full-chem chemical reaction mechanism.

[0115] like Figure 3 As shown, this figure is a scatter plot showing the correlation between predicted and actual values ​​for relevant chemical species. A total of 1,165,824 offline predictions were made, and the actual values ​​are the output of the original chemical integrator. Both predicted and actual values ​​are in molecules per cubic centimeter. The R in the figure... 2 The coefficient of determination is 1.000, and the root mean square error (RMSE) is 5.17 × 10⁻⁶. Subplot a is a scatter plot showing the correlation between predicted and actual ozone values, with a coefficient of determination of 1.000 and an RMSE of 5.17 × 10⁻⁶.9 Subplot b, showing the correlation between predicted and actual nitric oxide values ​​per cubic centimeter, has a coefficient of determination of 0.998 and a root mean square error of 3.13 × 10⁻⁶. 7 The number of molecules per cubic centimeter is shown in subplot c, which is a scatter plot comparing the predicted and actual values ​​of nitrogen dioxide. The coefficient of determination is 1.000, and the root mean square error is 5.10 × 10⁻⁶. 7 The number of molecules per cubic centimeter is used for subplot d, which is a scatter plot showing the correlation between predicted and actual values ​​of hydroxyl radicals. The coefficient of determination is 1.000, and the root mean square error is 5.00 × 10⁻⁶. 4 The molecular count per cubic centimeter shows that ChemKNet's predictions for ozone, nitric oxide, nitrogen dioxide, and hydroxyl radicals are highly consistent with the predictions of the original GEOS-Chem model's chemical integrator. Furthermore, in this implementation case, a total of 279 predictor variables were used, and in the 19-day offline prediction, 96% of the predictor variables had a coefficient of determination greater than 0.99, demonstrating the excellent predictive performance of the ChemKNet model in offline prediction.

[0116] Figure 4In this diagram, GC stands for GEOS-Chem. "GC coupled with ChemKNet" represents the GEOS-Chem model after coupling ChemKNet to the GEOS-Chem model. Subplots a, b, c, and d correspond to the near-surface to near-zero atmospheric pressure range. Subplot a shows the zonal average ozone concentration distribution of GC, and subplot b shows the zonal average ozone concentration distribution of the ChemKNet-coupled GC, both in parts per billion (parts per billion) volume fraction. Subplot c shows the absolute error distribution between GC and the ChemKNet-coupled GC, and subplot d shows the fractional error distribution between GC and the ChemKNet-coupled GC. Subplots e, f, g, and h correspond to the near-surface to near-stratospheric range. Subplot e shows the zonal average ozone concentration distribution of GC, and subplot f shows the zonal average ozone concentration distribution of the ChemKNet-coupled GC. All units are in parts per billion (ppm) of volume. Subplot g shows the absolute error distribution between GC and the GC coupled with ChemKNet, and subplot h shows the fractional error distribution between GC and the GC coupled with ChemKNet. By comparing subplots a and b, and subplots e and f, the ozone distribution predicted by the GC coupled with ChemKNet shows good consistency with the ozone distribution predicted by the original GC in different latitude zones and different atmospheric pressure layers. Observing the corresponding absolute error distribution plots c and g and fractional error plots d and h, obvious error patterns can be found. Areas with larger absolute errors mainly appear in the upper atmosphere, while areas with larger fractional errors mainly appear near the equator. This may be related to the lower background ozone concentration near the equator and the higher background ozone concentration in the upper atmosphere. Fractional errors are more sensitive to areas with lower background concentrations, while areas with higher background concentrations are also prone to larger absolute errors.

[0117] Figure 5 In the diagram, subplot a shows the global temporal average ozone concentration distribution predicted by the near-surface GC; subplot b shows the global temporal average ozone concentration distribution predicted by the near-surface GC coupled with ChemKNet; subplot c shows the fractional error distribution between the GC and the ChemKNet-coupled GC; and subplot d shows the absolute error distribution between the GC and the ChemKNet-coupled GC. Comparing subplots a and b, it can be seen that the ChemKNet-coupled GC can reproduce the near-surface ozone distribution of the GC very well. Analyzing the corresponding absolute error distribution map c and fractional error distribution map d, it can be seen that the absolute error distribution varies relatively little globally, with larger error values ​​appearing in a few areas. Overall, the prediction performance of the ChemKNet-coupled GC is better. As for the fractional error distribution map, the larger positive fractional errors are mainly concentrated near the equator, which is also a region with low background ozone concentration. The fractional error is more sensitive in this region. In conclusion, through the analysis of the concentration distribution map and the error distribution map, the ChemKNet-coupled GC has good overall prediction performance for the near-surface.

[0118] Figure 6 In the diagram, subplot a shows the global temporal average ozone concentration distribution predicted by GC at 500 hPa; subplot b shows the global temporal average ozone concentration distribution predicted by GC coupled with ChemKNet at 500 hPa; subplot c shows the fractional error distribution between GC and GC coupled with ChemKNet; and subplot d shows the absolute error distribution between GC and GC coupled with ChemKNet. Comparing subplots a and b, it can be seen that GC coupled with ChemKNet can capture both high and low ozone concentrations in the global ozone concentration distribution around 500 hPa. The spatial distribution trends are consistent, but the mid-latitude regions of the Northern Hemisphere tend to overestimate high values. Analyzing the corresponding absolute error distribution map c and fractional error distribution map d, we can see that areas with larger absolute errors are mainly distributed in the mid-latitude regions of the Northern Hemisphere. As for the fractional error distribution map, the distribution of larger positive fractional errors is similar to that of near-surface fractional errors, mainly concentrated near the equator. In summary, through the analysis of concentration distribution maps and error distribution maps, the GC prediction coupled with ChemKNet near 500 hPa can capture the spatial distribution trend of ozone.

[0119] Figure 7This study presents the prediction performance of relevant chemical species over a continuous 10-day period (720 predictions, 20-minute intervals) in three regions: Beijing, Legionovo, and Los Angeles. Concentration units are all in parts per billion (ppm). The blue line represents the time-series prediction of GC coupled with ChemKNet, and the black line represents the time-series prediction of GC. Subfigure a compares the ozone time-series predictions of GC in the Beijing region and those coupled with ChemKNet, with a corresponding Pearson correlation coefficient of 0.840. Subfigure b compares the ozone time-series predictions of GC in the Legionovo region and those coupled with ChemKNet, with a corresponding Pearson correlation coefficient of 0. Subplot c shows a comparison of ozone time-series predictions between the Los Angeles region GC and the GC coupled with ChemKNet, with a Pearson correlation coefficient of 0.782; subplot d shows a comparison of nitric oxide time-series predictions between the Beijing region GC and the GC coupled with ChemKNet, with a Pearson correlation coefficient of 0.974; subplot e shows a comparison of nitric oxide time-series predictions between the Legionovo region GC and the GC coupled with ChemKNet, with a Pearson correlation coefficient of 0.985; subplot f shows a comparison of nitric oxide time-series predictions between the Los Angeles region GC and the GC coupled with ChemKNet, with a Pearson correlation coefficient of 0.930. The correlation coefficient is 0.981; subplot g compares the nitrogen dioxide time series predictions of GC in the Beijing region and GC coupled with ChemKNet, with a corresponding Pearson correlation coefficient of 0.983; subplot h compares the nitrogen dioxide time series predictions of GC in the Legionovo region and GC coupled with ChemKNet, with a corresponding Pearson correlation coefficient of 0.997; subplot i compares the nitrogen dioxide time series predictions of GC in the Los Angeles region and GC coupled with ChemKNet, with a corresponding Pearson correlation coefficient of 0.996; subplot j compares the hydroxyl radical time series predictions of GC in the Beijing region and GC coupled with ChemKNet. The corresponding Pearson correlation coefficient is 0.928. Subplot k is a comparison of the time series predictions of hydroxyl radicals for the GC in the Legionovo region and the GC coupled with ChemKNet, with a corresponding Pearson correlation coefficient of 0.980. Subplot l is a comparison of the time series predictions of hydroxyl radicals for the GC in the Los Angeles region and the GC coupled with ChemKNet, with a corresponding Pearson correlation coefficient of 0.992. In summary, through comparison, it can be seen that the predicted trends of these four chemical species—ozone, nitric oxide, nitrogen dioxide, and hydroxyl radicals—in the three regions of Beijing, Legionovo, and Los Angeles are in good agreement with the predicted trends of the original GEOS-Chem model.

[0120] Figure 8The figure shows a comparison between the prediction time using the PyTorch machine learning model and the prediction time of the chemical calculation part of the original GEOS-Chem model. As shown in the figure, the original calculation of this part of the chemical process in GEOS-Chem, which is to predict the concentration changes of chemical species in 145,728 computational grid points, including the calculation of reaction constants, takes 185.54 seconds. If only the integrator calculation is performed, it takes 159.54 seconds on a single CPU. However, when this part of the calculation is performed using a neural network model, the calculation time on a single CPU is only 34.05 seconds, and the calculation time on a single GPU is only 2.36 seconds, which can save 99% of the original time.

[0121] In summary, this invention provides an alternative model for calculating atmospheric chemical reactions based on the neural network's ordinary differential equations. This model accurately reproduces chemical reaction processes: by replacing the chemical reaction calculation part of the atmospheric chemical transport model (GEOS-Chem) with the machine learning model ChemKNet, it can accurately predict the concentration evolution of chemical species such as O3, NO, NO2, and OH while maintaining the accuracy of the original chemical integrator. In the case study, 96% of the 279 predictor variables had a coefficient of determination exceeding 0.99 in 19 days of offline prediction. The model significantly accelerates chemical calculations: compared to the original GEOS-Chem chemical integrator, ChemKNet saves approximately 79% of the computation time on a single CPU and only about 1% of the original computation time on a single GPU, greatly improving the computational efficiency of large-scale simulations. The model supports coupling and continuous prediction: not only is it suitable for offline chemical box prediction, but it can also be coupled back to GEOS-Chem by ChemKNet to achieve continuous temporal evolution prediction of the global three-dimensional atmospheric chemical state, maintaining a trend and distribution highly consistent with the original model. The model described has full-chem chemical mechanisms covering GEOS-Chem and global simulation: The ChemKNet alternative used in this invention can cover multiple types of chemical reactions, is applicable to full-chem chemical mechanisms, is suitable for accelerating and replacing chemical processes in a global simulation environment, and has strong versatility and extensibility.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing an alternative model of atmospheric chemical reaction calculation, characterized by, Includes the following steps: (1) Under the corresponding chemical reaction mechanism, the input and output data of the chemical integrator are collected, and the collected input and output data are converted into Parquet format for data preprocessing. The data obtained after preprocessing is the training data. The training data is Z-score standardized and converted into sample data in TFRecord format. (2) Construct ChemKNet, an atmospheric chemical reaction calculation alternative model based on NeuralODE with Autoencoder, ROR and ResNeXt as basic components. Use the sample data obtained in step (1) for model hyperparameter tuning and model training to obtain the ChemKNet model after hyperparameter tuning and training. (3) Replace the atmospheric chemical reaction calculation process in the atmospheric chemical transport model with the ChemKNet model obtained in step (2). The ChemKNet model receives the input data of the atmospheric chemical model. After the input data is calculated by ChemKNet, the simulation output of the atmospheric chemical reaction calculation is obtained. (4) Online training process: The simulation output data obtained in step (3) and the output data of the integrator at the current time point of the actual chemical transport model are used to construct a training dataset. The training dataset is used to fine-tune the model. The model fine-tuning process continuously adjusts the learning rate used in the training process according to the loss stability of the ChemKNet model described in step (2) to obtain the ChemKNet model in the prediction form after adaptive coupling back to the atmospheric chemical transport model. (5) Using the ChemKNet model obtained in step (4), and based on the chemical step size set by the original atmospheric chemical transport model, output the simulation output of chemical reaction calculation after inputting the input data described in step (1); Step (2) of model hyperparameter tuning and model training includes the following steps: S1, the active chemical species concentration variable is subjected to feature extraction and dimension reduction processing by the encoder in the autoencoder, and low-dimensional potential features reflecting the chemical reaction system are extracted, and the expression is as follows: ; in, The corresponding active chemical species concentration variable in the model, for 3D real vector space, For variables Dimensions The encoded latent space feature vector. To After encoding The dimension, in which As one of the model structure hyperparameters, denoted as the encoded dimension, This represents the encoder weight matrix, where the number of rows corresponds to the input feature dimension. The number of columns represents the feature dimension after encoding. , This indicates its transpose. , It is the bias vector; S2, the concentration variables of active chemical species, are obtained after encoding. , It is concatenated with fixed chemical species concentration variables and meteorological variables as input to the first ResNeXt component. The expression that combines the fixed chemical species concentration variable and the meteorological variable is shown below: ; in, , For a fixed chemical species concentration variable, , Corresponding meteorological variables, , , and They represent the corresponding variables respectively , and Dimensions Features after splicing This indicates a concatenation operation along the feature dimension; S3. In the ResNeXt component. After multiple parallel fully connected layer branches, each branch is followed by a ReLU activation function to introduce non-linear expressive power. The outputs of all branches are concatenated and then integrated through a fully connected layer and a ReLU activation function; the expression is shown below: ； ； in, This indicates the number of parallel branches, or the number of branches. Let be one of the model structure hyperparameters, denoted as the number of parallel branches of ResNeXt; for the th Parallel branches , This is the weight matrix of the fully connected layer in this branch. Let it be its transpose matrix. This is the bias vector of the fully connected layer in this branch. The activation function is used, and the branch output is... , The feature dimension of the output after parallel branching is determined by the number of neurons in the parallel branch and is one of the model structure hyperparameters, denoted as the number of neurons in the parallel branch of ResNeXt. , This indicates that a concatenation operation is performed along the feature dimension. This is the weight matrix of the fully connected layer used to integrate the spliced ​​features. Let it be its transpose matrix. For the corresponding bias vector , for The dimension is determined by the weight matrix of the fully connected layer that integrates and splices features. The value determined by this is one of the model structure hyperparameters, denoted as the number of neurons in the integrated fully connected layer; S4 After passing through a fully connected layer, the output dimension is mapped back to... Dimensions: ； in, These are the features obtained after the preceding integration operations; Here is the weight matrix of the fully connected layer. Let it be its transpose matrix. This corresponds to the bias vector; since this layer is used to map the feature dimensions back... The dimension, therefore , The output features are those obtained after dimension remapping. S5. Each ResNeXt component has its own local residual connections to ensure stable feature propagation within a single component. The formula is as follows: ; in, , After local residual connection, and The result obtained by adding them together; The process from S1 to S5 involves data in... The computational operations performed within the component can be simplified as follows: ; S6 The next ResNeXt with the same structure will be expressed as: ; in, , express After the second The output obtained after the component; How many specifically? It can be obtained by setting the model structure hyperparameters and passing through the set n ResNeXt modules. Then, perform layer normalization, as shown in the formula below: ； in, Presentation layer normalization operation, and For trainable scaling and translation parameters, ⊙ denotes element-wise multiplication. To prevent small constants from being divided by zero, Standard deviation, The mean, These are the features obtained after layer normalization; S7, obtained Then, feature fusion is performed through global residual connections; the global residual connections contain a set... There are 1 ResNeXt component, each with local residual connections, thus forming a Relational Oorestable Relation (ROR); its formula is expressed as: ； in, These are low-dimensional latent features obtained by encoding the concentration variables of active chemical species in S1; The final output feature obtained by S6 is the feature after being processed and layer-normalized by multiple ResNeXt components inside the ROR structure. The ROR output features are obtained through global residual connections. ; S8. Finally, the decoder outputs a prediction, remapping the features in the high-dimensional latent space back to the original chemical species variable dimension space, as shown in the following formula: ； in, Here is the weight matrix of the fully connected layer of the decoder. Let it be its transpose matrix. For the corresponding bias vector; the decoder will Mapping to the prediction output space yields the prediction result. , or denoted as Therefore ,in, To define the dimensions of the output predictor variables, i.e., the original active chemical species concentration variables; Then, inverse standardization is performed, expressed by the following formula: ； in, For the first The predicted value of each feature, The first one obtained after inverse standardization Predicted concentration values ​​of a chemical substance, For the first The mean of each feature in the training set, For the first The standard deviation of each feature in the training set.

2. The method for constructing a computational alternative model for atmospheric chemical reactions as described in claim 1, characterized in that, The formula for Z-score standardization mentioned in step (1) is expressed as follows: ;in, For the first before standardization One characteristic, For the first The mean of each feature in the training set, For the first The standard deviation of each feature in the training set For the standardized first One characteristic.

3. The method for constructing a computational alternative model for atmospheric chemical reactions as described in claim 1, characterized in that, The input data in step (1) includes meteorological variables, fixed chemical species concentration variables, and active chemical species concentration variables.

4. The method for constructing a computational alternative model for atmospheric chemical reactions as described in claim 3, characterized in that, The meteorological variables include 9, the fixed chemical species concentration variables include 4, and the active chemical species concentration variables include 279.

5. The method for constructing a computational alternative model for atmospheric chemical reactions as described in claim 4, characterized in that, The meteorological variables are TEMP, PRESS, NUMDEN, H2O, SUNCOS, RELHUM, UVALBEDO, PEDGE, and OPTD; the fixed chemical species concentration variables are H2, N2, O2, and RCOOH; TEMP is temperature, PRESS is air pressure, NUMDEN is air number density, H2O is water vapor number concentration, SUNCOS is the cosine of the solar zenith angle, RELHUM is relative humidity, UVALBEDO is the ultraviolet albedo, PEDGE is the moist air pressure at the bottom of the grid, and OPTD is visible optical thickness; H2 is hydrogen, N2 is nitrogen, O2 is oxygen, and RCOOH is organic acid.

Images