Method for establishing evolution model of influence of seabed leakage CO2 on phaeodactylum tricornutum

Abstract

The invention provides a method for establishing an influence evolution model of seabed leakage CO2 on phaeodactylum tricornutum, and belongs to the technical field of marine pollution. The method comprises the following steps: firstly, obtaining acidification sensitivity parameters of phaeodactylum tricornutum through an experiment, constructing a two-phase flow near-region transport and diffusion model to simulate a CO2 leakage process, and then establishing a far-region transport and diffusion model based on an FVCOM; a near-region result is adopted as a source item to calculate pH change, an OceanAttn model is introduced to optimize a pH distribution prediction result, and the model adopts a hierarchical fusion gating mechanism to dynamically adjust weight distribution according to hydrological feature complexity, leakage scale and water layering strength. And establishing a dose effect function by combining the exposure quantity and acidification sensitivity parameters of the phaeodactylum tricornutum, and finally constructing a population dynamic evolution equation, thereby realizing the accurate prediction of the influence of the seabed leakage CO2 on the spatio-temporal dynamic change of the phaeodactylum tricornutum population.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine pollution, and in particular relates to a method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum. Background Art

[0002] Seafloor carbon capture and storage (SC) is an important means of mitigating climate change, but assessing the risks of CO2 leakage and its impact on marine ecosystems is crucial. As a primary producer in the ocean, population fluctuations in P. tricornutum reflect the health of marine ecosystems. Traditional assessment methods rely primarily on a combination of laboratory toxicology experiments and numerical simulations, measuring the effects of varying pH on algal growth, photosynthesis, and calcification rates to establish simple dose-effect relationships. However, these methods have significant limitations. First, laboratory conditions differ significantly from the actual marine environment, making it difficult to simulate complex ocean currents and water mass mixing. Second, existing numerical simulation models are designed separately for near- and far-flung areas, lacking effective connectivity mechanisms, making it difficult to integrate predictions across different spatial scales. Furthermore, traditional models fail to adequately consider the spatiotemporal variability of marine environmental parameters, making it difficult to accurately characterize the impact of pH changes on P. tricornutum population dynamics. A core challenge currently hindering these technologies is the lack of a model architecture that can effectively integrate multi-scale spatial information and temporal features, hindering accurate simulation and prediction of the spatiotemporal dynamics of P. tricornutum populations. Especially in complex marine environments, traditional models find it difficult to accurately capture the complex correlations between changes in hydrological conditions, biological responses, and population dynamics. Summary of the Invention

[0003] In view of this, the present invention provides a method for establishing an evolutionary model of the impact of submarine CO2 leakage on the triangular cyanobacteria, which can solve the technical problem in the existing technology that it is difficult to accurately predict the spatiotemporal dynamic impact of submarine CO2 leakage on the triangular cyanobacteria population.

[0004] The present invention is implemented as follows: the present invention provides a method for establishing an evolution model of the impact of submarine CO2 leakage on triangular ochre, including: collecting and analyzing toxicological experimental data of triangular ochre on seawater acidification with different pH values, obtaining the acidification sensitivity parameters of triangular ochre; constructing a two-phase flow near-zone transport and diffusion model to obtain the near-zone transport and diffusion results of submarine CO2 leakage; constructing a far-zone transport and diffusion model based on FVCOM to simulate and calculate the far-zone transport and diffusion and pH changes of submarine CO2 leakage; introducing the OceanAttn model to optimize the pH change results calculated by the far-zone transport and diffusion model. The results show that the OceanAttn model fuses ocean current field data and pH field data through a multi-head attention mechanism to improve the prediction accuracy of pH distribution; calculates the exposure dose affected by P. tricornutum; uses the acidification sensitivity parameters of P. tricornutum and the exposure dose to establish a dose-response function for the change in P. tricornutum population size; combines ocean current field data with P. tricornutum growth and reproduction characteristics data to establish a dynamic evolution equation for P. tricornutum populations, and uses a hierarchical fusion gating mechanism to integrate multi-scale spatiotemporal characteristics to form a complete spatiotemporal evolution model of the impact of submarine CO2 leakage on P. tricornutum.

[0005] Among them, the acidification sensitivity parameters of P. tricornutum specifically refer to the growth rate variation coefficient, photosynthesis efficiency reduction rate and calcification rate variation coefficient of P. tricornutum under different pH conditions.

[0006] Among them, the two-phase flow near-zone transport and diffusion model specifically refers to the PBM module of the Euler model established using the CFD software Fluent. The PBM module simulates micro-scale processes such as bubble breakage and merging, and is suitable for simulating CO2 leakage processes.

[0007] The dose-effect function specifically refers to the mathematical relationship that describes the change in the population size of P. tricornutum with pH value, which is usually expressed as an S-shaped curve and includes pH threshold and maximum inhibition rate parameters.

[0008] Among them, the population dynamic evolution equation specifically refers to a set of partial differential equations that describes the spatiotemporal changes in population size by combining the growth rate, mortality rate, diffusion migration rate and environmental factor influence coefficient of the triangular brown finger algae.

[0009] Among them, the exposure amount specifically refers to the product of the degree to which the pH value of the water body where the triangular kelp is located deviates from the normal value and the duration, which is used to quantify the cumulative effect of acidification on the triangular kelp.

[0010] Among them, ocean current field data specifically refers to the horizontal and vertical flow velocity and direction information at each point in the sea area, which is used to calculate the water mass transport path and diffusion rate.

[0011] Among them, the leakage scenario specifically refers to a combination of different leakage rates, leakage durations and leakage tide times, which is used to simulate various possible leakage events.

[0012] Among them, the specific structure of the OceanAttn model is an ocean environment prediction model designed based on the Transformer architecture. It consists of two parts: an encoder and a decoder. The encoder adopts a six-layer multi-head attention mechanism, with each layer containing eight attention heads. The decoder also adopts a six-layer architecture with cross-layer connection characteristics.

[0013] The core component of the OceanAttn model is a hierarchical attention network mechanism, consisting of a spatial attention module and a temporal attention module. The spatial attention module captures the spatial correlation between ocean currents and pH distribution, while the temporal attention module processes the temporal evolution of ocean environmental parameters. The OceanAttn model inputs include local transport and diffusion results, ocean current field data, historical pH distribution data, and time series of environmental parameters; its output is an optimized pH distribution prediction. The OceanAttn model employs a hierarchical fusion weight adjustment mechanism, dynamically adjusting the fusion weights of attention modules at different levels to enable the model to adaptively optimize prediction performance based on different ocean characteristics and spill scenarios. The hierarchical fusion weights are determined by three parameters: the hydrological complexity index, the spill scale coefficient, and the intensity of water stratification. The hierarchical fusion gating mechanism specifically refers to a multi-level feature fusion mechanism that utilizes these three parameters to regulate the performance. This mechanism is controlled by a gating weight function, which dynamically calculates the weight distribution at each level based on the input parameters. The gating weight function first normalizes the three parameters, mapping them to a range of 0 to 1, and then calculates an overall balance value. Based on the range of the overall balance value, the values are divided into three intervals: low sensitivity, moderate sensitivity, and high sensitivity, and weight distribution is calculated using different weighting functions. The training dataset for the OceanAttn model was established by collecting historical submarine spill data from multiple sea areas; augmenting the collected historical data using numerical simulation methods; constructing the training dataset by mixing actual monitoring data with simulated data in an 8:2 ratio; performing spatiotemporal alignment of the data; and normalizing the pH data.

[0014] The present invention optimizes the pH distribution results calculated by the far-zone transport and diffusion model using the OceanAttn model and integrates multi-scale spatiotemporal features using a hierarchical fusion gating mechanism to form a complete impact evolution model. This method organically connects the near-zone and far-zone models, addressing the difficulty in fusing prediction results at different spatial scales. By introducing a hierarchical fusion gating mechanism to dynamically adjust the fusion weights of attention modules at different levels, the present invention effectively integrates multi-scale spatial information and temporal features. This mechanism adaptively adjusts the weight allocation strategy based on three parameters: the hydrological complexity index, the leak scale coefficient, and the intensity of water stratification, addressing the inability of traditional models to accurately capture the complex correlation between changing hydrological conditions and biological responses. By accurately simulating the transport and diffusion of leaked CO2 in the marine environment and its impact on pH, combined with the sensitivity parameters of the triangular algae to acidification and the population dynamics equation, the present invention achieves accurate prediction of the spatiotemporal dynamics of the triangular algae population affected by submarine CO2 leaks, providing a reliable scientific basis for environmental risk assessment of submarine carbon sequestration technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of the method of the present invention.

[0016] Figure 2 The figure is a structural diagram of the OceanAttn model involved in the present invention.

[0017] Figure 3 The present invention relates to a schematic structural diagram of a two-phase flow near-zone transport and diffusion model. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] like Figure 1 FIG. 1 is a flow chart of a method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum provided by the present invention. The method comprises the following steps:

[0020] S01. Collect and analyze the toxicological experimental data of Phaeodactylum tricornutum on seawater acidification at different pH values to obtain the acidification sensitivity parameters of Phaeodactylum tricornutum;

[0021] S02. Construct a two-phase flow near-zone transport and diffusion model, and couple it with the gas-liquid mass transfer influence law obtained from the experiment to obtain the near-zone transport and diffusion results of submarine CO2 leakage;

[0022] S03. Construct a far-area transport and diffusion model based on FVCOM, use the near-area transport and diffusion results as source terms, and simulate and calculate the far-area transport and diffusion and pH changes of the submarine leaked CO2;

[0023] S04. Introducing the OceanAttn model to optimize the pH change results calculated by the far-field transport and diffusion model. The OceanAttn model fuses ocean current field data and pH field data through a multi-head attention mechanism to improve the pH distribution prediction accuracy;

[0024] S05. Calculate the exposure to Phaeodactylum triangularis based on the pH distribution prediction results under different leakage scenarios and the spatial distribution data of Phaeodactylum triangularis;

[0025] S06. establishing a dose-response function for the change in the population of Phaeodactylum tricornutum using the acidification sensitivity parameter of Phaeodactylum tricornutum and the exposure dose;

[0026] S07. Combining the ocean current data with the growth and reproduction characteristics of P. tricornutum, establishing a dynamic evolution equation for the P. tricornutum population, and using a hierarchical fusion gating mechanism to integrate multi-scale spatiotemporal characteristics to form a complete spatiotemporal evolution model of the impact of submarine CO2 leakage on P. tricornutum;

[0027] Among them, the acidification sensitivity parameters of P. tricornutum specifically refer to the growth rate variation coefficient, photosynthesis efficiency reduction rate and calcification rate variation coefficient of P. tricornutum under different pH conditions.

[0028] like Figure 3 As shown in FIG, the two-phase flow near-zone transport and diffusion model specifically refers to the PBM module of the Euler model established using the CFD software Fluent. The PBM module simulates micro-scale processes such as bubble breakage and merging, and is suitable for simulating the CO2 leakage process.

[0029] The dose-effect function specifically refers to the mathematical relationship that describes the change in the population size of P. tricornutum with pH value, which is usually expressed as an S-shaped curve and includes pH threshold and maximum inhibition rate parameters.

[0030] Among them, the population dynamic evolution equation specifically refers to a set of partial differential equations that describes the spatiotemporal changes in population size by combining the growth rate, mortality rate, diffusion migration rate and environmental factor influence coefficient of the triangular brown finger algae.

[0031] Among them, the exposure amount specifically refers to the product of the degree to which the pH value of the water body where the triangular kelp is located deviates from the normal value and the duration, which is used to quantify the cumulative effect of acidification on the triangular kelp.

[0032] Among them, ocean current field data specifically refers to the horizontal and vertical flow velocity and direction information at each point in the sea area, which is used to calculate the water mass transport path and diffusion rate.

[0033] Among them, the leakage scenario specifically refers to a combination of different leakage rates, leakage durations and leakage tide times, which is used to simulate various possible leakage events.

[0034] like Figure 2 As shown in the figure, the specific structure of the OceanAttn model is an ocean environment prediction model designed based on the Transformer architecture, which consists of two parts: an encoder and a decoder. The encoder adopts a six-layer multi-head attention mechanism, with eight attention heads in each layer. The decoder also adopts a six-layer architecture with cross-layer connection characteristics. The core part is a hierarchical attention network mechanism, which includes a spatial attention module and a temporal attention module. The spatial attention module captures the spatial correlation between ocean currents and pH distribution, and the temporal attention module processes the evolution of ocean environmental parameters over time. The model input includes the near-zone transport and diffusion results, the ocean current field data, the historical pH distribution data and the environmental parameter time series. The output is the optimized pH distribution prediction result. The most notable feature of the model is the use of a hierarchical fusion weight adjustment mechanism. By dynamically adjusting the fusion weights of attention modules at different levels, the model can adaptively optimize the prediction performance according to different sea area characteristics and leakage scenarios. The hierarchical fusion weight is jointly determined by three parameters: the hydrological characteristic complexity index, the leakage scale coefficient and the water body stratification intensity.

[0035] The steps for establishing the training data set of the OceanAttn model specifically include collecting historical submarine leakage event data from multiple sea areas, including leakage location, leakage rate, leakage duration, pH distribution data at multiple time points before and after the leakage, the ocean current field data, and temperature and salinity data; using numerical simulation methods to expand the collected historical data, and generating simulated data for different leakage scenarios by changing the leakage parameters; mixing actual monitoring data and simulated data in a ratio of 8:2 to form a training data set; performing spatiotemporal alignment processing on the data to ensure consistency of data from different sources in time and space dimensions; normalizing the pH value data and uniformly adjusting the value range to between 0 and 1; converting the ocean current field data into two components of flow velocity and flow direction, and performing vector normalization; performing time window sliding segmentation on the training data set to generate sample pairs containing input feature sequences and corresponding labels; finally, stratified sampling of the data set according to different sea areas and different leakage scales to ensure that the training set covers various environmental conditions and the leakage scenarios.

[0036] The OceanAttn model training steps specifically include first initializing model parameters using mini-batch stochastic gradient descent, with a batch size of 64 and an initial learning rate of 0.001, and dynamically adjusting the learning rate using a cosine annealing strategy. The training process is divided into two stages: in the first stage, the hierarchical fusion weights are fixed and the basic model parameters are trained until convergence; in the second stage, the hierarchical fusion weights are unfrozen and all parameters are simultaneously optimized through backpropagation. A weighted combination of mean squared error and relative pH change rate is used as the loss function, with the relative pH change rate being given a higher weight in low pH regions to improve the model's prediction accuracy in areas with severe acidification. L2 regularization is introduced to mitigate the risk of overfitting, with the regularization coefficient set to 0.0001. A validation set evaluation is performed every 10 training cycles, and an early stopping mechanism is triggered when the performance of the validation set stops improving after three consecutive training cycles. A model ensemble technique is used to train five models with the same structure but different initializations, and the weighted average of their prediction results is taken as the final output. Finally, the model performance is evaluated using a test set to ensure that the expected pH distribution prediction accuracy can be achieved under different marine environments.

[0037] The hierarchical fusion gating mechanism specifically refers to a multi-level feature fusion mechanism that utilizes the three parameters of the hydrological characteristic complexity index, the leakage scale coefficient and the water body stratification intensity to jointly regulate; the hierarchical fusion gating mechanism is controlled by a gating weight function, and the weight distribution of each level is dynamically calculated according to the input parameters; the gating weight function first standardizes the three parameters, maps them to the range of 0 to 1, and then calculates the comprehensive balance value. The comprehensive balance value calculation considers the linear combination of the three parameters and introduces a nonlinear activation function to enhance the expression ability; according to the range of the comprehensive balance value, it is divided into three intervals: low sensitivity interval, moderate sensitivity interval and high sensitivity interval; when the comprehensive balance value is in the low sensitivity interval, the linear weight function is used , the linear weight function assigns similar weights to low-level features and high-level features; when the comprehensive balance value is in the moderately sensitive interval, an exponential weight function is adopted, and the exponential weight function increases the weight distribution exponentially with the increase of the layer depth; when the comprehensive balance value is in the highly sensitive interval, a logarithmic weight function is adopted, and the logarithmic weight function assigns higher weights to the underlying features to retain detailed information; in addition, the gating weight function also includes an adaptive adjustment mechanism, which dynamically fine-tunes the weight distribution strategy by monitoring the deviation between the pH distribution prediction result and the actual observation value, so that the system continuously optimizes performance during long-term operation; the hierarchical fusion gating mechanism significantly improves the model's prediction accuracy for the dynamic changes of the triangular algae population under different marine environmental conditions.

[0038] Among them, the hydrological characteristic complexity index specifically refers to a dimensionless parameter that characterizes the complexity of the hydrological conditions in the sea area, which is calculated by the ocean current velocity gradient, thermohaline jump layer intensity and terrain change rate.

[0039] Among them, the leakage scale coefficient specifically refers to the product of the leakage rate and the leakage duration and is a dimensionless parameter after normalization, which is used to characterize the overall scale of the leakage event.

[0040] Among them, water stratification intensity specifically refers to the gradient of water density changing with depth, which is a dimensionless parameter that characterizes the difficulty of vertical mixing of water.

[0041] The relative pH change rate specifically refers to the percentage change of the pH value in the acidified area relative to the background value, which is a parameter used to measure the degree of acidification.

[0042] The specific implementation of the above steps is described in detail below.

[0043] The specific implementation of step S01 is to collect and analyze toxicological experimental data on P. tricornutum in response to seawater acidification at different pH values to obtain the acidification sensitivity parameter of P. tricornutum. This step first designs a pH gradient experiment, culturing P. tricornutum in a seawater environment with a pH range of 5.5 to 8.5, with an interval of 0.2 pH units, and setting up five parallel groups for each pH condition. Then, the biomass change of P. tricornutum is measured using optical density, the photosynthetic efficiency is measured using pulse-modulated chlorophyll fluorescence technology, and the calcification rate is measured by alkalinity difference subtraction. Next, the growth rate variation coefficient, photosynthetic efficiency reduction rate, and calcification rate variation coefficient are calculated for each pH condition relative to the control group (pH 8.1). Finally, all experimental data are normalized and the functional relationship between pH value and acidification sensitivity parameter is fitted using nonlinear regression analysis, where a generalized logistic function is used as the fitting function. The acidification sensitivity parameters of P. tricornutum obtained through this step are of great ecological significance and can be used as basic data for evaluating the impact of acidification. The pH threshold (i.e., the pH value when the growth rate is reduced by 50%) is usually between pH 7.3 and 7.5.

[0044] The specific implementation of step S02 is to construct a two-phase flow near-zone transport and diffusion model, and couple it with the gas-liquid mass transfer influence law obtained from the experiment to obtain the near-zone transport and diffusion results of the submarine leaked CO2. This step first uses the CFD software Fluent to establish the PBM module of the Euler model, sets the simulation area to a three-dimensional space of 100m×100m×50m, uses unstructured tetrahedral mesh for grid division, and performs grid encryption in the area near the leakage source; then, the bubble dynamics equations are implemented in the PBM module, including the bubble rising velocity model, the bubble breakage and merging model, and the bubble size distribution evolution equation; then, the gas-liquid two-phase interface mass transfer model is used to describe the CO2 transfer process from bubbles to water bodies, considering the CO2 hydration reaction kinetics and the acid-base equilibrium reaction of seawater, and the mass transfer coefficient is calculated according to the Higbie penetration theory; finally, the boundary conditions are set and numerical solutions are performed, which include the CO2 flow rate of the leakage source (0.1~10kg / s), the leakage aperture (0.01~0.1m), the seawater background flow rate (0~0.5m / s) and the ambient temperature (5~25℃). The calculation results of this step include the near-zone CO2 concentration distribution, pH value distribution, and bubble plume rising trajectory, which can be used as source input for the far-zone model and provide initial conditions for subsequent simulations. The simulation results show that the near-zone pH value can drop below 6.5 near the leakage source.

[0045] The specific implementation of step S03 is to construct a far-field transport and diffusion model based on FVCOM, using the near-field transport and diffusion results as source terms to simulate and calculate the far-field transport and diffusion of submarine CO2 leaks and pH changes. This step first constructs a three-dimensional unstructured grid for the study area with a horizontal resolution range of 50 to 500 meters and is divided into 10 to 20 vertical layers using a sigma coordinate system. The physical parameters of the model are then set, including horizontal and vertical eddy diffusion coefficients, bottom friction coefficients, and open boundary conditions. The eddy diffusion coefficient is calculated using the Smagorinsky turbulent closed model. The near-field transport and diffusion results obtained in step S02 are then interpolated in time and space as source term input for the far-field model. Finally, the geochemical module for the CO2 dissolution and acidification process is implemented, including the carbonate system equilibrium equation and the seawater acid-base balance governing equation. This step uses the splitting operator method to solve the control equation. The time step is set to 60 to 300 seconds and the simulation duration is 7 to 30 days. The pH value change process in a large range of the study sea area can be obtained, providing basic data for evaluating ecological impacts. The simulation results show that the pH value change range in the far area is usually between 0.1 and 1.0 pH units.

[0046] The specific implementation of step S04 is to introduce the OceanAttn model to optimize the pH change results calculated by the far-field transport and diffusion model. The OceanAttn model fuses ocean current field data and pH field data through a multi-head attention mechanism to improve the accuracy of pH distribution prediction. This step first constructs the input dataset of the OceanAttn model, including the pH distribution results simulated by FVCOM, ocean current field data, and environmental parameter time series. Then, the pH distribution is predicted using the pre-trained OceanAttn model. The model is based on the Transformer architecture and consists of six layers of encoders and decoders, each containing eight attention heads. The core component is the hierarchical attention network mechanism. The model prediction results are then weighted and fused with the original FVCOM simulation results. The weight coefficient is dynamically adjusted according to the hydrological characteristic complexity index, and the OceanAttn model is given a higher weight in areas with a high complexity index. Finally, the model is verified and corrected using measured data. When the root mean square error between the predicted value and the measured value exceeds 0.05 pH units, the model retraining procedure is initiated. This step optimizes the calculation results of the traditional hydrodynamic model through machine learning methods, significantly improving the accuracy of pH distribution prediction, especially in complex hydrological environments, where the prediction accuracy can be improved by more than 30%.

[0047] The specific implementation of step S05 is to calculate the exposure to P. tricornutum based on the pH distribution prediction results under different leakage scenarios and the spatial distribution data of P. tricornutum. This step first collects the spatial distribution data of P. tricornutum in the study area, including the abundance distribution map and seasonal variation characteristics. Then, multiple leakage scenarios are designed, including a low-rate, long-term leakage (0.1-1 kg / s, lasting 30-90 days), a medium-rate, medium-duration leakage (1-5 kg / s, lasting 7-30 days), and a high-rate, short-duration leakage (5-10 kg / s, lasting 1-7 days), and the influence of different tides (high tide, low tide, and slack tide). Then, the pH distribution time series under each scenario is superimposed and analyzed with the spatial distribution data of P. tricornutum, and the pH deviation value (deviation from the normal value of 8.1) and duration within each grid cell are calculated. Finally, the exposure value is calculated using an integral method, that is, the product of the pH deviation value and the duration is accumulated for all affected areas. This step achieves a quantitative assessment of the degree of acidification impact on P. tricornutum by calculating the exposure dose, providing a data basis for the subsequent establishment of a dose-effect relationship. Studies have shown that when the cumulative exposure dose exceeds 2.0 pH units × day, the P. tricornutum population will be significantly affected.

[0048] The specific implementation of step S06 is to use the acidification sensitivity parameter of the triangular algae and the exposure dose to establish a dose-effect function of the population change of the triangular algae. This step first analyzes the relationship between the acidification sensitivity parameter obtained in step S01 and the exposure dose calculated in step S05; then constructs a mathematical expression of the dose-effect function, using a four-parameter logistic function form, including a maximum inhibition rate parameter, a minimum inhibition rate parameter, a half-effect exposure parameter and a curve slope parameter; then calibrates the function parameters using experimental data and field observation data, and uses the least squares method to optimize the parameters, with the error controlled within 5%; finally, verifies the applicability of the dose-effect function under different environmental conditions, considering temperature (10-30°C), light (50-300μmol / m 2 The dose-response function established in this step accurately describes the response of the P. tricornutum population to changes in pH, exhibiting a typical S-shaped curve. The pH threshold is typically between 7.3 and 7.5, and the maximum inhibition rate can reach over 90%, providing important parameters for subsequent population dynamics simulations.

[0049] The specific implementation of step S07 is to combine the ocean current field data with the growth and reproduction characteristics of P. triangularis to establish a dynamic evolution equation for the P. triangularis population. A hierarchical fusion gating mechanism is used to integrate multi-scale spatiotemporal features, forming a complete spatiotemporal evolution model of the impact of submarine CO2 leaks on P. triangularis. This step first constructs a basic set of equations for the dynamics of the P. triangularis population, including an equation for population size change, an equation for spatial diffusion, and an equation for the influence of environmental factors. The dose-response function established in step S06 is then introduced to integrate the effects of pH changes on the P. triangularis population dynamics equation. A hierarchical fusion gating mechanism is then used to integrate the multi-scale spatiotemporal features. This mechanism utilizes three parameters: the hydrological complexity index, the leak scale coefficient, and the water stratification intensity to jointly regulate the feature fusion process. Finally, a numerical solution method is used to calculate the population distribution changes at different time steps. The fourth-order Runge-Kutta method is used for time integration, and the finite volume method is used for spatial discretization. The spatiotemporal evolution model established in this step realizes the full-process simulation of the dynamic changes of the triangular kelp population, and can predict the spatiotemporal distribution characteristics of the population under different leakage scenarios, providing a scientific basis for the ecological risk assessment of submarine CO2 leakage.

[0050] The detailed structure of the OceanAttn model, based on the Transformer architecture, is a deep learning model for ocean environmental prediction. The model consists of two main components: an encoder and a decoder. The encoder employs a six-layer multi-head attention mechanism, with eight attention heads per layer, to extract deep features from the input data. The decoder also employs a six-layer structure with cross-layer connections, enabling the fusion of feature information from different layers. The model's input data includes local transport and diffusion results, ocean current field data, historical pH distribution data, and environmental parameter time series. These data are pre-processed and fed into the model. The output is an optimized pH distribution prediction with higher spatiotemporal resolution and accuracy. The core of the OceanAttn model is a hierarchical attention network mechanism, comprising a spatial attention module and a temporal attention module. The spatial attention module employs a self-attention algorithm to capture the spatial correlation between ocean currents and pH distribution, calculate weight relationships between different locations, and provide spatial dependency information for the model. The temporal attention module employs a bidirectional long-short-term memory network structure to process the temporal evolution of ocean environmental parameters, capturing both long-term trends and short-term fluctuations. The model's most notable feature is its hierarchical fusion weight adjustment mechanism. By dynamically adjusting the fusion weights of attention modules at different levels, the model achieves adaptive optimization for different sea area characteristics and spill scenarios. The hierarchical fusion weights are determined by three parameters: the hydrological characteristic complexity index, the spill scale coefficient, and the water body stratification intensity, and are calculated using a gated weight function. During the training phase, the model adopts a two-stage training strategy. First, the hierarchical fusion weights are fixed and the basic model parameters are trained until convergence. Then, the hierarchical fusion weights are unfrozen and all parameters are optimized simultaneously through backpropagation. The loss function uses a weighted combination of mean squared error and relative pH change rate, assigning higher weights to low pH areas to improve prediction accuracy in areas with severe acidification. In addition, L2 regularization is introduced to mitigate the risk of overfitting, with the regularization coefficient set to 0.0001.

[0051] The detailed steps for establishing the training data set of the OceanAttn model include five steps: data collection, data expansion, data mixing, data alignment, and data preprocessing. First, in the data collection step, historical submarine leakage incident data of multiple sea areas are screened from the Global Ocean Observing System database. The data must include complete leakage location information, leakage rate records (0.1 to 10 kg / s), leakage duration records (1 to 90 days), pH distribution data at at least 5 time points before and after the leakage, ocean current field data (flow rate range 0 to 2 m / s), and temperature and salinity data; the sampling density of pH distribution data is not less than every 10 km 2For one sampling point, the temporal resolution is no less than one sampling every 6 hours; the spatial resolution of the ocean current field data is no less than 1 km, and the temporal resolution is no less than 1 hour. Then, in the data expansion phase, the collected historical data is expanded using numerical simulation methods, and simulated data of different leakage scenarios are generated by systematically changing the leakage parameters (including leakage rate, duration and leakage depth); the simulation is carried out using the aforementioned FVCOM model, and a total of about 200 leakage scenario data are generated, covering three typical leakage scenarios: low-rate long-term, medium-rate medium-term, and high-rate short-term. Then, in the data mixing phase, the actual monitoring data and the simulated data are mixed in a ratio of 8:2 to form a training data set. The proportion of actual data is no less than 20% to ensure that the model can learn the characteristics of real leakage events; the final constructed data set contains about 1,000 spatiotemporal series samples, each of which contains all the monitoring data of a complete leakage event. Then, in the data alignment phase, data from different sources were spatially and temporally aligned. Temporal alignment used linear interpolation to adjust all data to a unified time series with a time step of 1 hour. Spatial alignment used Kriging interpolation to resample data of varying spatial resolutions to a unified spatial grid with a grid resolution of 500 meters. Finally, in the data preprocessing phase, pH data were normalized to a range between 0 and 1 using minimum-maximum scaling. Current field data were converted into velocity and direction components, each of which was normalized. Temperature and salinity data were standardized to conform to a normal distribution with a mean of 0 and a standard deviation of 1. The training dataset was segmented using a sliding time window with a window length of 48 hours and a sliding step of 12 hours to generate sample pairs containing the input feature sequence and its corresponding label. The dataset was stratified by sampling based on different sea areas and spill sizes to ensure that the training set covered a variety of environmental conditions and spill scenarios. The dataset was then divided into training, validation, and test sets in an 8:1:1 ratio.

[0052] It should be noted that the core technical ideas of the present invention are mainly reflected in three aspects. First, the introduction of the OceanAttn model realizes the in-depth processing and analysis of complex spatiotemporal data of the marine environment. The model is based on the Transformer architecture and processes nonlinear relationships in the marine environment through a six-layer multi-head attention mechanism, effectively capturing the spatial correlation between ocean currents and pH distribution and the evolution of environmental parameters over time. Compared with traditional numerical simulation methods, the OceanAttn model can automatically extract deep-level features and establish intrinsic correlations between complex environmental factors, overcoming the limitations of traditional models in expressing complex nonlinear relationships, and significantly improving the model's prediction accuracy and adaptability in a changing marine environment.

[0053] Secondly, the design of a hierarchical fusion gating mechanism provides the model with intelligent adaptive capabilities. This mechanism dynamically adjusts the fusion weights of attention modules at different levels based on three key parameters: the hydrological characteristic complexity index, the leak scale coefficient, and the water body stratification intensity. By intelligently judging the complexity of the environment and adopting linear, exponential, or logarithmic weighting functions under different conditions, the model achieves adaptive integration of multi-scale features. This dynamic weight allocation strategy breaks through the limitations of fixed parameter settings in traditional models, enabling the model to automatically optimize its structure and parameters based on different sea area characteristics and leak scenarios, significantly improving the model's generalization ability and predictive stability in complex and changing environments.

[0054] Third, the integrated connection between near- and far-zone models solves the challenge of integrating information at different spatial scales. By incorporating the results of a two-phase flow near- and far-zone transport and diffusion model as source terms for a far-zone transport and diffusion model, this approach achieves full-scale simulation, from microscopic bubble dynamics to macroscopic water mass motion. Compared to traditional approaches that employ separate near- and far-zone models, this coherent model architecture more accurately describes the complete process of CO2 diffusion from the leak source to the larger-scale diffusion, improving the overall accuracy of predictions of the scope and extent of acidification impacts.

[0055] The synergistic effect of the three core technical ideas mentioned above has formed a multi-scale, multi-level, and adaptive comprehensive model system. By organically combining accurate near-field physical simulation, efficient far-field diffusion calculation, and deep learning optimization prediction, the system is able to simulate and predict the entire chain from microscopic leakage processes to macroscopic ecological impacts. Especially under complex marine environmental conditions, the three technical ideas complement and enhance each other, forming a comprehensive description of the impact process of submarine CO2 leakage. This synergistic effect breaks through the limitations of the limited applicability of traditional single models, creating a comprehensive assessment system that can cope with the changing marine environment and adapt to different leakage scenarios, providing a more reliable and comprehensive scientific basis for the environmental risk assessment of submarine carbon sequestration technology.

[0056] Specifically, the principle of the present invention is that it can accurately predict the spatiotemporal dynamic impact of submarine CO2 leakage on the population of Phaeodactylum tricornutum. The principle is mainly reflected in the following aspects:

[0057] First, this invention establishes a comprehensive technical approach, achieving a complete process design from experimental data collection to final impact prediction. Experiments were used to obtain toxicological data on the effects of Phaeodactylum tricornutum on seawater acidification at varying pH values, ensuring the scientific validity of the biological response parameters. A two-phase flow near-field transport and diffusion model simulates microscale processes such as bubble breakup and merging, providing a rational description of the near-field physical characteristics of CO2 leakage. A far-field transport and diffusion model, based on FVCOM, ensures the accuracy of large-scale transmission processes. This integrated multi-scale model lays the foundation for accurate predictions in complex environments.

[0058] Secondly, the introduction of the OceanAttn model is the core innovation of this invention. The model is based on the Transformer architecture and adopts a six-layer multi-head attention mechanism, with each layer containing eight attention heads, which can effectively process complex spatiotemporal data associations. The model captures the spatial correlation between ocean currents and pH distribution through the spatial attention module, and processes the evolution of marine environmental parameters over time through the temporal attention module, realizing the deep fusion and analysis of multi-source heterogeneous data. This deep learning architecture can automatically extract complex nonlinear features, overcoming the complex environmental response relationships that are difficult to express with traditional numerical models.

[0059] Crucially, the hierarchical fusion gating mechanism designed in this invention provides the model with adaptive capabilities. This mechanism dynamically calculates the fusion weights of features at different levels based on three key parameters: the hydrological complexity index, the leak scale coefficient, and the intensity of water stratification. When environmental conditions are simple, a linear weighting function is used to balance the features of each layer; when the environmental conditions are moderately complex, an exponential weighting function is used to enhance deep-layer features; and when the environment is highly complex, a logarithmic weighting function is used to preserve underlying detailed information. This intelligent weighting strategy enables the model to adaptively optimize for different sea area characteristics and leak scenarios, significantly improving prediction accuracy.

[0060] Furthermore, the model training employed rigorous data processing and optimization strategies, including blending real-world monitoring data with simulated data, stratified sampling, and spatiotemporal alignment. Furthermore, a weighted combined loss function of mean squared error and relative pH change rate was introduced, particularly enhancing its predictive capabilities in areas of severe acidification. These design features ensured the model's robust generalization and predictive stability in complex and changing marine environments.

[0061] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0062] The specific implementation of step S01 is to collect and analyze the toxicological experimental data of P. triangularis on seawater acidification with different pH values, and obtain the acidification sensitivity parameters of P. triangularis. This step first designs a pH gradient experiment, culture P. triangularis in a seawater environment with a pH range of 5.5 to 8.5, with an interval set to 0.2 pH units, and set up 5 parallel groups under each pH condition; then use the optical density method to measure the biomass change of P. triangularis, use the pulse modulation chlorophyll fluorescence technology to measure the photosynthesis efficiency, and use the alkalinity difference subtraction method to measure the calcification rate; then calculate the growth rate variation coefficient, photosynthesis efficiency reduction rate and calcification rate variation coefficient under each pH condition relative to the control group (pH 8.1); finally, normalize all experimental data and use nonlinear regression analysis to fit the functional relationship between pH value and acidification sensitivity parameter. The calculation formula of the growth rate variation coefficient of P. triangularis is as follows:

[0063]

[0064] Where G r is the growth rate variation coefficient, dimensionless; G pH is the growth rate under certain pH conditions, in d -1 ; G 8.1 is the growth rate under pH 8.1 conditions (control group), unit is d -1 .

[0065] The formula for calculating the photosynthesis efficiency reduction rate is as follows:

[0066]

[0067] Where, P r is the rate of decrease in photosynthesis efficiency, dimensionless; F v / F m (pH) is the maximum photochemical quantum yield under certain pH conditions; F v / F m (8.1) is the maximum photochemical quantum yield under pH 8.1 conditions (control group).

[0068] The calculation formula of the calcification rate variation coefficient is as follows:

[0069]

[0070] Where C r is the coefficient of variation of calcification rate, dimensionless; C pH The calcification rate under certain pH conditions, unit is μmolCaCO3·cell -1 ·d -1 ; C 8.1 is the calcification rate at pH 8.1 (control group), unit is μmol CaCO3·cell -1 ·d -1 .

[0071] For fitting the functional relationship between pH value and acidification sensitivity parameter, the generalized logistic function is used:

[0072]

[0073] Where S(pH) is the acidification sensitivity parameter (G r 、P r or C r );S min is the minimum value of the parameter; S max is the maximum value of the parameter; k is the slope parameter of the curve; pH0 is the half-effect pH value, that is, when the parameter value reaches The acidification sensitivity parameters of Phaeodactylum tricornutum obtained through this step are of great ecological significance and can be used as basic data for assessing the impact of acidification. The pH threshold (i.e., the pH value when the growth rate is reduced by 50%) is usually between pH 7.3 and 7.5.

[0074] The specific implementation method of step S02 is to construct a two-phase flow near-zone transport and diffusion model, and couple the gas-liquid mass transfer influence law obtained from the experiment to obtain the near-zone transport and diffusion results of the submarine leaked CO2. This step first uses the CFD software Fluent to establish the PBM module of the Euler model, sets the simulation area to a three-dimensional space of 100m×100m×50m, uses unstructured tetrahedral grids for mesh division, and performs mesh encryption in the area near the leakage source; then implements the bubble dynamics equations in the PBM module, including the bubble rising velocity model, the bubble breakage and merging model, and the bubble size distribution evolution equation; then describes the transfer process of CO2 from bubbles to water bodies through the gas-liquid two-phase interface mass transfer model, taking into account the CO2 hydration reaction kinetics and the acid-base equilibrium reaction of seawater, and the mass transfer coefficient is calculated according to the Higbie penetration theory; finally, the boundary conditions are set and numerical solutions are performed. The bubble rising velocity model uses the modified Schiller-Naumann relationship:

[0075]

[0076] Where U b is the rising velocity of the bubble, in m / s; g is the acceleration due to gravity, in m / s 2 ;d b is the bubble diameter, in m; ρ l is the liquid density, in kg / m 3 ρ g is the gas phase density, in kg / m 3 ;μ l is the liquid phase dynamic viscosity, in Pa·s; C D is the drag coefficient, calculated using the following formula:

[0077]

[0078] Where, Re b is the bubble Reynolds number,

[0079] The bubble size distribution evolution equation adopts the population balance model (PBM):

[0080]

[0081] Where n(v, t) is the number density function of bubbles with a volume of v, and the unit is m -3 ; is the bubble velocity vector, in m / s; B co is the bubble merging rate, in m -3 ·s -1 ;D co is the bubble merging and disappearing rate, in m -3 ·s -1 ; B br is the bubble breakage rate, in m -3 ·s -1 ;D br is the bubble breakage and disappearance rate, unit is m -3 ·s -1 .

[0082] The gas-liquid two-phase interface mass transfer model adopts the double-membrane theory:

[0083]

[0084] Where, is the CO2 concentration in the liquid phase, in mol / m 3 ;k L is the liquid phase mass transfer coefficient, in m / s; a is the specific surface area, in m -1 ; is the equilibrium concentration of the solution, determined by Henry's law, in mol / m 3 . Mass transfer coefficient k L Calculated according to Higbie penetration theory:

[0085]

[0086] Where, is the diffusion coefficient of CO2 in water, in m 2 / s;t c is the contact time, in seconds, The calculation results of this step include the near-zone CO2 concentration distribution, pH value distribution, and bubble plume rising trajectory, which can be used as source input for the far-zone model and provide initial conditions for subsequent simulations. The simulation results show that the near-zone pH value can drop below 6.5 near the leakage source.

[0087] The specific implementation method of step S03 is to construct a far-zone transport and diffusion model based on FVCOM, use the near-zone transport and diffusion results as source terms, and simulate and calculate the far-zone transport and diffusion and pH changes of CO2 leaked from the seabed. This step first constructs a three-dimensional unstructured grid of the study sea area with a horizontal resolution range of 50 to 500m, and is divided into 10 to 20 layers in the vertical direction using the sigma coordinate system; then sets the physical parameters of the model, including horizontal and vertical eddy diffusion coefficients, bottom friction coefficients, and open boundary conditions, where the eddy diffusion coefficient is calculated using the Smagorinsky turbulent closed model; then, the near-zone transport and diffusion results obtained in step S02 are interpolated in time and space as the source term input of the far-zone model; finally, the geochemical module of the CO2 dissolution and acidification process is implemented, including the carbonate system equilibrium equation and the seawater acid-base balance control equation. The horizontal eddy diffusion coefficient is calculated using the Smagorinsky formula:

[0088]

[0089] Where A M is the horizontal eddy diffusion coefficient, in m 2 / s; C is an empirical constant with a value range of 0.1 to 0.2; Ω is the grid unit area in m 2 ;u x 、v y 、u y 、v x are the spatial gradients of the horizontal velocity components, respectively.

[0090] The equilibrium equation of the carbonate system after CO2 dissolution is as follows:

[0091]

[0092] Where K1 and K2 are the first and second dissociation constants, respectively, which are related to temperature, salinity, and pressure. The governing equation for the acid-base balance of seawater is:

[0093]

[0094] Where K W is the ion product constant of water. Combined with the charge balance equation:

[0095]

[0096] The pH value is calculated as follows:

[0097] pH = -log 10 [H + ].

[0098] This step uses the splitting operator method to solve the control equation. The time step is set to 60 to 300 seconds and the simulation duration is 7 to 30 days. The pH value change process in a large range of the study sea area can be obtained, providing basic data for evaluating ecological impacts. The simulation results show that the pH value change range in the far area is usually between 0.1 and 1.0 pH units.

[0099] The specific implementation of step S04 is to introduce the OceanAttn model to optimize the pH change results calculated by the far-field transport and diffusion model. The OceanAttn model fuses ocean current field data and pH field data through a multi-head attention mechanism to improve the accuracy of pH distribution prediction. This step first constructs the input data set of the OceanAttn model, including the pH distribution results simulated by FVCOM, ocean current field data and environmental parameter time series; then uses the pre-trained OceanAttn model to predict pH distribution; then the model prediction results are weighted and fused with the original FVCOM simulation results, and the weight coefficient is dynamically adjusted according to the hydrological characteristic complexity index; finally, it is verified and corrected through measured data. The weighted fusion formula of the OceanAttn model prediction results and the original FVCOM simulation results is as follows:

[0100] pH final =α·pH OceanAttn +(1-α)·pH FVCOM ;

[0101] Where, pH final is the predicted result of pH distribution after final optimization; pH OceanAttn is the prediction result of OceanAttn model; pH FVCOM is the original simulation result of the FVCOM model; α is the weight coefficient, which is determined by the hydrological complexity index (HCI):

[0102]

[0103] Where λ is the adjustment parameter, ranging from 2.0 to 5.0; HCI0 is the complexity threshold, ranging from 0.5 to 0.7. The calculation formula for the hydrological characteristic complexity index is:

[0104]

[0105] Where, is the ocean current velocity gradient, in s -1 ; is the vertical density gradient, in kg / m 4 ; is the rate of topographic change, expressed in m / m; w1, w2, and w3 are weight coefficients, with w1 + w2 + w3 = 1, typically set to w1 = 0.4, w2 = 0.4, and w3 = 0.2. This step optimizes the results of traditional hydrodynamic models through machine learning, significantly improving the accuracy of pH distribution predictions, particularly in complex hydrological environments, where prediction accuracy can be increased by over 30%.

[0106] The specific implementation method of step S05 is to calculate the exposure amount of the affected P. tricornutum based on the pH distribution prediction results under different leakage scenarios and the spatial distribution data of P. tricornutum. This step first collects the spatial distribution data of P. tricornutum in the study sea area, including the abundance distribution map and seasonal variation characteristics; then designs a variety of leakage scenarios, including low-rate long-term leakage (0.1-1 kg / s, lasting 30-90 days), medium-rate medium-time leakage (1-5 kg / s, lasting 7-30 days) and high-rate short-time leakage (5-10 kg / s, lasting 1-7 days), and considers the influence of different tides (high tide, low tide, and flat tide); then the pH distribution time series under each scenario is superimposed and analyzed with the spatial distribution data of P. tricornutum to calculate the pH deviation value and duration in each grid unit; finally, the exposure amount is calculated using the integral method. The pH deviation value calculation formula is:

[0107] ΔpH(x,y,z,t)=|pH(x,y,z,t)-pH normal |;

[0108] Where ΔpH(x, y, z, t) is the pH deviation value of the coordinate point (x, y, z) at time t; pH(x, y, z, t) is the pH value predicted by the model; pH normal The pH value of normal seawater is usually 8.1.

[0109] The exposure calculation formula is:

[0110]

[0111] Where E(x, y, z) is the exposure at the coordinate point (x, y, z), and the unit is pH unit × day; t0 and t1 are the start and end times of the leak, respectively.

[0112] Considering the spatial distribution of P. tricornutum, the total exposure calculation formula is:

[0113] E total =∫∫∫E(x,y,z)·D(x,y,z)dxdydz;

[0114] Where, E totalis the total exposure, with the unit of pH unit × day × number of individuals; D(x, y, z) is the population density of P. tricornutum at the coordinate point (x, y, z), with the unit of number of individuals / m 3 This step achieved a quantitative assessment of the degree of acidification impact on P. tricornutum by calculating the exposure dose, providing a data basis for the subsequent establishment of a dose-effect relationship. Studies have shown that when the cumulative exposure exceeds 2.0 pH units per day, the P. tricornutum population will be significantly affected.

[0115] The specific implementation of step S06 is to use the acidification sensitivity parameter of the triangular algae and the exposure dose to establish a dose-response function for changes in the triangular algae population. This step first analyzes the relationship between the acidification sensitivity parameter obtained in step S01 and the exposure dose calculated in step S05; then constructs a mathematical expression for the dose-response function using a four-parameter logistic function; then calibrates the function parameters using experimental data and field observation data; and finally verifies the applicability of the dose-response function under different environmental conditions. The dose-response function uses a four-parameter logistic function:

[0116]

[0117] Where, R(E) is the relative change rate of the population of P. tricornutum when the exposure dose is E; R min is the minimum inhibition rate parameter, which indicates the survival rate under the maximum inhibition effect, and its value range is 0.05~0.2; R max is the maximum inhibition rate parameter, usually close to 1.0; EC 50 is the half-effect exposure parameter, that is, the exposure when the population size is reduced by 50%, the unit is pH unit × day, and the value range is 1.5 to 2.5; Hill is the curve slope parameter, which characterizes the steepness of the dose-effect curve, and the value range is 1.5 to 4.0.

[0118] Taking into account the regulatory effect of environmental factors, the modified dose-effect function is:

[0119] R′(E)=R(E)·f T ·f I ·f N ;

[0120] Where R′(E) is the relative rate of change of population size after considering the influence of environmental factors; f T 、f I 、f N are the regulation coefficients of temperature, light and nutrients respectively. The temperature regulation coefficient calculation formula is:

[0121]

[0122] Where, T is the actual temperature, in °C; T opt The optimum temperature for P. tricornutum is usually 20-25℃; T tol is the temperature tolerance parameter, with a value of 5 to 10°C. The calculation formula for the light regulation coefficient is:

[0123]

[0124] Where I is the actual light intensity, in μmol / m 2 / s;K I is the half-saturation light constant, with a value of 100 to 150 μmol / m 2 / s. The calculation formula of nutrient salt adjustment coefficient is:

[0125]

[0126] Where N and p are the concentrations of nitrogen and phosphorus, respectively, in μmol / L; K N and K P are the half-saturation constants for nitrogen and phosphorus, with values ranging from 1.0 to 2.0 μmol / L and 0.1 to 0.3 μmol / L, respectively. The dose-response function established in this step accurately describes the response of the P. tricornutum population to changes in pH, exhibiting a typical S-shaped curve. The pH threshold is typically between 7.3 and 7.5, and the maximum inhibition rate can reach over 90%, providing important parameters for subsequent population dynamics simulations.

[0127] The specific implementation of step S07 is to combine the ocean current field data with the growth and reproduction characteristic data of the triangular brown finger algae to establish the dynamic evolution equation of the triangular brown finger algae population, and use a hierarchical fusion gating mechanism to integrate multi-scale spatiotemporal features to form a complete spatiotemporal evolution model of the impact of submarine CO2 leakage on the triangular brown finger algae. This step first constructs the basic equation group of the triangular brown finger algae population dynamics; then introduces the dose-effect function established in step S06 to integrate the effect of pH value changes on the triangular brown finger algae into the population dynamics equation; then uses a hierarchical fusion gating mechanism to integrate multi-scale spatiotemporal features; finally, calculates the population distribution changes under different time steps through a numerical solution method. The basic equation of the triangular brown finger algae population dynamics is:

[0128]

[0129] Where N is the population density of P. triangularis, and the unit is individuals / m 3 ; t is time, unit is d; is the ocean current velocity vector, in m / s; D is the diffusion coefficient, in m 2 / s; G is the growth rate, unit is d -1 ; M is the mortality rate, unit is d-1 .

[0130] The population dynamics equation after considering the influence of pH changes is:

[0131]

[0132] Wherein, R′(E) is the relative change rate of the population after considering the influence of environmental factors determined in step S06.

[0133] The hierarchical fusion gating mechanism is applied to the solution process of the population dynamics equation. By dynamically adjusting the weight coefficients of features at different scales, the model prediction accuracy is improved. The calculation formula of the gating weight function is:

[0134] W l =g(HCI,LSC,SWI,l);

[0135] Where W l is the weight coefficient of the lth layer feature; HCI is the hydrological characteristic complexity index; LSC is the leakage scale coefficient; SWI is the water body stratification intensity; g is the gating function, which is determined according to the comprehensive balance value B:

[0136] B=β1·HCI+β2·LSC+β3·SWI;

[0137] Where β1, β2, and β3 are balance coefficients, and β1+β2+β3=1. The values are usually β1=0.4, β2=0.3, and β3=0.3.

[0138] Depending on the range of B values, the gating weight function takes different forms:

[0139] When 0≤B<0.3 (low sensitivity range), W l =a1·l+b1;

[0140] When 0.3≤B<0.7 (moderately sensitive range),

[0141] When 0.7≤B≤1.0 (highly sensitive range), W l =a3·ln(b3·l+c3);

[0142] Where a1, b1, a2, b2, a3, b3, and c3 are fitting parameters determined by data calibration; l is the layer depth index, ranging from 1 to 6. The calculation method of the hydrological characteristic complexity index is the same as step S04. The leakage scale coefficient calculation formula is:

[0143]

[0144] Where Q is the leakage rate, in kg / s; T is the leakage duration, in d; Q max and T max are the maximum leakage rate and maximum duration considered in the model respectively. The water stratification intensity calculation formula is:

[0145]

[0146] Where H is the water depth in meters and ρ is the seawater density in kg / m 3 z is the vertical coordinate, measured in meters. The spatiotemporal evolution model established in this step simulates the entire dynamics of the P. tricornutum population, predicting the spatiotemporal distribution of populations under different leakage scenarios and providing a scientific basis for ecological risk assessment of submarine CO2 leaks.

[0147] The detailed structure of the OceanAttn model is based on the Transformer architecture and is a deep learning model for marine environmental prediction. The model consists of two major components: an encoder and a decoder. The encoder uses a six-layer multi-head attention mechanism, with each layer containing eight attention heads, to extract deep features from the input data. The decoder also uses a six-layer structure with cross-layer connections, capable of fusing feature information from different levels. The model's input data includes local transport and diffusion results, ocean current field data, historical pH distribution data, and environmental parameter time series. All types of data are preprocessed and then input into the model. The output is an optimized pH distribution prediction result with higher spatiotemporal resolution and prediction accuracy. The core of the OceanAttn model is a hierarchical attention network mechanism, which includes a spatial attention module and a temporal attention module. The spatial attention module uses a self-attention algorithm to capture the spatial correlation between ocean currents and pH distribution, calculate the weight relationship between different locations, and provide spatial dependency information for the model. The temporal attention module uses a bidirectional long-short-term memory network structure to process the evolution of ocean environmental parameters over time, capturing long-term trends and short-term fluctuation characteristics. The calculation formula of the spatial attention mechanism is:

[0148]

[0149] Where Q, K, and v are query matrix, key matrix, and value matrix, respectively. d k and d v are the dimensions of the key vector and value vector respectively; n and m are the lengths of the query sequence and key-value sequence respectively; softmax is the normalization function. The calculation formula of the multi-head attention mechanism is:

[0150] MultHead(Q,K,V)=Concat(head1,head2,...,head h )W O ;

[0151] Where, is the learnable parameter matrix; h is the number of attention heads, which is 8; d model is the hidden layer dimension of the model, and its value is 512. The temporal attention module adopts a bidirectional long short-term memory network (BiLSTM) structure, and the forward LSTM calculation formula is:

[0152] f t =σ(W f ·[h t-1 , x t ]+b f );

[0153] i t =σ(W i ·[h t-1 , x t ]+b i );

[0154]

[0155] o t =σ(W o ·[h t-1 , x t ]+b o );

[0156] h t =o t *tanh(C t );

[0157] Where, f t 、i t 、o t They are forget gate, input gate and output gate respectively; C t is the unit state; h t is the hidden state; x t is the input vector; W f 、W i 、W C 、W o is the weight matrix; b f 、b i 、b C 、b ois the bias vector; σ is the sigmoid activation function; tanh is the hyperbolic tangent activation function; and * is element-wise multiplication. The backward LSTM is calculated in a similar way, and the final BiLSTM output is the concatenation of the forward and backward LSTM outputs:

[0158]

[0159] The most notable feature of the model is its hierarchical fusion weight adjustment mechanism. By dynamically adjusting the fusion weights of attention modules at different levels, the model achieves adaptive optimization for different ocean characteristics and spill scenarios. The hierarchical fusion weights are determined by three parameters: the hydrological complexity index, the spill scale coefficient, and the water stratification intensity, and are calculated using a gated weight function. The loss function of the OceanAttn model is a weighted combination of the mean square error and the relative pH change rate:

[0160] Loss = λ1·MSE+λ2·RPR;

[0161] Where MSE is the mean square error, RPR is the relative pH change rate loss, in is the weight coefficient, γ is a tuning parameter ranging from 0.5 to 1.0; λ1 and λ2 are loss weights, with λ1 + λ2 = 1, typically set to λ1 = 0.3 and λ2 = 0.7. During training, the model employs a two-stage training strategy: first, the layer-wise fusion weights are fixed, and the basic model parameters are trained until convergence. Then, the layer-wise fusion weights are unfrozen, and all parameters are optimized simultaneously through backpropagation.

[0162] The detailed steps for establishing the training data set of the OceanAttn model include five steps: data collection, data expansion, data mixing, data alignment, and data preprocessing. First, in the data collection step, historical submarine leakage incident data of multiple sea areas are screened from the Global Ocean Observing System database. The data must include complete leakage location information, leakage rate records (0.1 to 10 kg / s), leakage duration records (1 to 90 days), pH distribution data at at least 5 time points before and after the leakage, ocean current field data (flow rate range 0 to 2 m / s), and temperature and salinity data; the sampling density of pH distribution data is not less than every 10 km 2At each sampling point, the temporal resolution is no less than once every 6 hours; the spatial resolution of the ocean current field data is no less than 1 km, and the temporal resolution is no less than 1 hour. Then, in the data expansion phase, numerical simulation methods are used to expand the collected historical data. By systematically changing the leakage parameters (including leakage rate, duration, and leakage depth), simulated data for different leakage scenarios are generated. The simulation is performed using the aforementioned FVCOM model to generate approximately 200 leakage scenario data with leakage scale coefficients (LSC) ranging from 0.1 to 1.0. The calculation formula is as follows:

[0163]

[0164] Here, the meanings of the parameters are the same as described above. Next, in the data mixing phase, actual monitoring data and simulated data are mixed in an 8:2 ratio to form a training dataset. The actual data ratio is no less than 20% to ensure that the model can learn the characteristics of real leak events. The final dataset contains approximately 1,000 spatiotemporal sequence samples, each of which contains all the monitoring data for a complete leak event. Then, in the data alignment phase, the data from different sources are spatiotemporally aligned, using linear interpolation for temporal alignment:

[0165]

[0166] Where y(t) is the interpolation result at time t; y(t i ) and y(t i+1 ) are adjacent moments t i and t i+1 The spatial alignment uses Kriging interpolation:

[0167]

[0168] Where, is the interpolation estimate at position s0; Z(s i ) is the position s i The observation value at λ i is the interpolation weight, satisfying Determined by the variogram model:

[0169]

[0170] Where γ(h) is the semivariogram value of the sample point pair separated by a distance of h; C0 is the nugget effect, which represents microscale variability and has a value of 0.01 to 0.05; C is the base value, which represents the overall spatial variability and has a value of 0.2 to 1.0; a is the range, which represents the effective distance of spatial correlation and has a value of 10 to 50 km. Finally, in the data preprocessing step, the pH value data is normalized to adjust the value range to between 0 and 1. The normalization adopts the minimum and maximum value scaling method:

[0171]

[0172] Where, pH norm is the normalized pH value; pH min and pH max are the minimum and maximum pH values in the data set respectively. Standardization of temperature and salinity data:

[0173]

[0174] Where, X std is the standardized data; X is the original data; μ and σ are the mean and standard deviation of the data, respectively. The training dataset was segmented using a sliding time window with a window length of 48 hours and a sliding step of 12 hours to generate sample pairs consisting of input feature sequences and corresponding labels. The dataset was stratified by sampling based on different sea areas and spill sizes to ensure that the training set covers various environmental conditions and spill scenarios. The dataset was then divided into training, validation, and test sets in a ratio of 8:1:1.

[0175] The hierarchical fusion gating mechanism specifically refers to a multi-level feature fusion mechanism that is jointly regulated by the three parameters of the hydrological characteristic complexity index, the leakage scale coefficient, and the water body stratification intensity. The hierarchical fusion gating mechanism is controlled by a gating weight function, which dynamically calculates the weight distribution of each level based on the input parameters. The gating weight function calculation steps are as follows: First, the three parameters of the hydrological characteristic complexity index (HCI), leakage scale coefficient (LSC), and water body stratification intensity (SWI) are standardized and mapped to the range of 0 to 1:

[0176]

[0177]

[0178] Where, HCI norm , LSC norm and SWI norm They are the three parameters after standardization; HCI min , LSC min and SWI minare the minimum values of the three parameters respectively; HCI max , LSC max and SWI max are the maximum values of the three parameters respectively. Then calculate the comprehensive balance value B:

[0179] B=β1·HCl norm +β2·LSC norm +β3·SWI norm ;

[0180] Where β1, β2 and β3 are weight coefficients, and β1+β2+β3=1, which is determined by optimization on the validation set. The values are usually β1=0.4, β2=0.3, and β3=0.3. Then, a nonlinear activation function is introduced to enhance the expressive power:

[0181]

[0182] Where B' is the comprehensive balance value after nonlinear transformation; λ is the shape parameter, ranging from 5 to 10; B0 is the center point parameter, ranging from 0.5. According to the range of B' values, it is divided into three intervals: low sensitivity interval (0 ≤ B' < 0.3), moderate sensitivity interval (0.3 ≤ B' < 0.7), and high sensitivity interval (0.7 ≤ B' ≤ 1.0). Depending on the interval to which B' belongs, the gating weight function adopts different forms:

[0183] When B′ is in the low sensitivity range, a linear weight function is used:

[0184] W l =a1·l+b1;

[0185] When B′ is in the moderately sensitive range, the exponential weight function is used:

[0186]

[0187] When B′ is in the highly sensitive interval, the logarithmic weight function is used:

[0188] W l =a3·ln(b3·l+c3);

[0189] Where W l is the weight of the lth layer feature; l is the layer depth index, ranging from 1 to 6; a1, b1, a2, b2, a3, b3, and c3 are fitting parameters, determined by optimization on the validation set. To ensure weight normalization, the following processing is performed:

[0190]

[0191] Where W l′ is the normalized feature weight of the lth layer; L is the total number of layers, which is 6. In addition, the gated weight function also includes an adaptive adjustment mechanism, which dynamically fine-tunes the weight allocation strategy by monitoring the deviation between the pH distribution prediction result and the actual observation value:

[0192] W l ′(t+1)=W l ′(t)+η·ΔW l ';

[0193] Where W l ′(t) and W l ′(t+1) are the weight values at time t and t+1 respectively; η is the learning rate, ranging from 0.01 to 0.1; ΔW l ′ is the weight adjustment amount, which is determined by the prediction error:

[0194]

[0195] Where α is the adjustment coefficient, ranging from 0.1 to 1.0; E is the prediction error, The hierarchical fusion gating mechanism significantly improves the model's accuracy in predicting the dynamics of P. tricornutum populations under different marine environmental conditions.

[0196] To better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: Researchers conducted a series of simulation experiments to verify the effectiveness and accuracy of the evolution model of the impact of submarine CO2 leakage on the triangular kelp. The simulation experiment selected a typical sea area as the virtual scene. This sea area is a rectangular area with dimensions of 50 km x 30 km, an average water depth of 120 m, and a distinct seasonal thermocline and a variable tidal environment. The simulation experiment first constructed an acidification sensitivity database for the triangular kelp according to step S01. By integrating published experimental data and a self-designed pH gradient experiment (5.6 to 8.4, with an interval of 0.2), a complete set of acidification sensitivity parameters for the triangular kelp was obtained, as shown in Table 1.

[0197] Table 1 Fitting results of acidification sensitivity parameters of Phaeodactylum triangularis in simulation experiments

[0198]

[0199]

[0200] Based on these parameters, the researchers used a generalized logistic function to construct the response relationship of P. tricornutum to acidification, and determined that when the pH was below 7.35, the growth rate would be significantly reduced. Subsequently, in step S02, the researchers used Fluent software to construct a two-phase flow near-zone transport and diffusion model to simulate the near-zone dynamic process after CO2 leakage. The simulation experiment designed three typical leakage scenarios, namely, a low-rate long-term leakage (0.8 kg / s, lasting 60 days), a medium-rate medium-time leakage (3.5 kg / s, lasting 14 days), and a high-rate short-time leakage (8.0 kg / s, lasting 3 days). The main parameter settings and results of the near-zone model are shown in Table 2.

[0201] Table 2 Near-zone model simulation parameters and results

[0202] Parameters / Results Low-rate, long-term leakage Medium rate and medium time leakage High-rate, short-term leakage Leak depth 120m 120m 120m Leakage diameter 0.03m 0.05m 0.08m Background flow rate 0.12m / s 0.12m / s 0.12m / s Grid quantity 486,520 486,520 486,520 Calculating time steps 0.1s 0.1s 0.1s Lowest pH value in 10m range 6.87 6.42 6.15 Lowest pH value in 50m range 7.62 7.23 7.05 Bubble plume height 73m 85m 93m <![CDATA[CO2 dissolution rate]]> 87％ 79％ 72％

[0203] Results from the near-field model show that the pH decreases more significantly in high-rate leak scenarios, but the bubble dissolution rate is lower; in low-rate leak scenarios, the pH decreases more moderately, but the overall dissolution rate is higher. The researchers used the output of the near-field model as a source term and input it into a far-field transport and diffusion model built on FVCOM. The far-field model has a spatial resolution of 200m, is divided into 18 vertical layers, has a time step of 120s, and simulates for the duration of the leak plus a 30-day recovery period. To improve the accuracy of pH distribution predictions, the researchers used the OceanAttn model in step S04 to optimize the far-field model results. The OceanAttn model was trained on a virtual training dataset containing simulated data for 300 different leak scenarios. The training process and model performance are shown in Table 3.

[0204] Table 3 OceanAttn model training process and performance evaluation

[0205]

[0206] Compared with the traditional FVCOM model, the root mean square error of the pH distribution prediction results after the optimization of the OceanAttn model was reduced from 0.075 to 0.029, an improvement of 61.3%. The prediction effect was particularly significant in areas with complex hydrological environments. In step S05 of the simulation experiment, the researchers designed a distribution scenario of triangular algae in a virtual sea area. The distribution density was high near the shore and low far from the shore in the horizontal direction. In the vertical direction, it was most dense in the 20-40m water layer, with an average density of 3.2×10 5 The calculation results of the cumulative exposure in each area under the three leakage scenarios are shown in Table 4.

[0207] Table 4 Regional cumulative exposure under different leakage scenarios (pH unit × day)

[0208] Distance from the leak point Low-rate, long-term leakage Medium rate and medium time leakage High-rate, short-term leakage 0-5km 3.24 2.86 1.92 5-10km 2.18 1.96 1.25 10-15km 1.53 1.31 0.76 15-20km 0.92 0.78 0.43 20-30km 0.46 0.36 0.21 >30km 0.12 0.09 0.05

[0209] Based on the acidification sensitivity parameters and cumulative exposure data, the researchers established a dose-response function in step S06. The R min =0.06, R max =0.98,EC 50 =1.82pH units × day, Hill = 2.76. This indicates that when the cumulative exposure reaches 1.82pH units × day, the population of P. tricornutum will decrease by 50%. In step S07, the researchers constructed a complete dynamic evolution model of the P. tricornutum population and applied a hierarchical fusion gating mechanism to integrate multi-scale spatiotemporal features. For the three leakage scenarios, the researchers calculated the comprehensive balance value based on the hydrological characteristic complexity index (0.58), the leakage scale coefficient (0.32, 0.42, and 0.36, respectively), and the water stratification intensity (0.47), and determined the corresponding gating weight function.

[0210] Simulation results show that a low-rate, long-duration leak has the greatest impact on the population of Phaeodactylum tricornutum, with the total population decreasing by 31.5% after the leak ends and a further 18.3% decrease after the recovery period. A medium-rate, medium-duration leak results in a 25.7% population decrease, followed by a 12.1% decrease after the recovery period. A high-rate, short-duration leak results in a 17.9% population decrease, followed by an 8.2% decrease after the recovery period. Spatially, all three scenarios show the most significant population decrease within a 10 km radius of the leak (72.6%, 65.4%, and 48.9%, respectively), with the impact decreasing rapidly with increasing distance. Traditional methods for assessing the ecological impact of CO2 leaks typically use direct extrapolation of experimental data under a single pH value or use simple Gaussian diffusion models to predict pH distribution, which fail to reflect the complex processes in the real marine environment. Compared with traditional assessment methods, the present invention improves the pH distribution prediction accuracy by 61.3% by integrating the two-phase flow near-zone model, the FVCOM far-zone model and the OceanAttn optimization model; by introducing the dose-effect function and the hierarchical fusion gating mechanism, the consistency between the population dynamics prediction of Phaeodactylum tricornutum and the reference simulation data is improved from 76.8% of the traditional method to 93.5%, an improvement of 16.7%, providing more accurate technical support for the environmental risk assessment of submarine CO2 storage projects.

[0211] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5, 6 and 7 below.

[0212] Table 5 Variable Explanation Table (Part 1)

[0213]

[0214]

[0215] Table 6 Variable Explanation Table (Part 2)

[0216]

[0217]

[0218] Table 7 Variable Explanation Table (Part 3)

[0219]

[0220]

[0221] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum, characterized in that: include: Toxicological experimental data on the effects of P. tricornutum on seawater acidification at different pH values were collected and analyzed to obtain the acidification sensitivity parameters of P. tricornutum. A two-phase flow near-field transport and diffusion model was constructed to obtain the near-field transport and diffusion results of CO2 leaked from the seabed. A far-field transport and diffusion model was constructed based on FVCOM to simulate and calculate the far-field transport and diffusion and pH changes of CO2 leaked from the seabed. The OceanAttn model was introduced to optimize the pH change results calculated by the far-field transport and diffusion model. The OceanAttn model integrates ocean current field data and pH field data through a multi-head attention mechanism to improve the accuracy of pH distribution prediction. Calculate the exposure dose affected by P. triangularis; use the acidification sensitivity parameters and exposure dose of P. triangularis to establish a dose-effect function for the population change of P. triangularis; combine ocean current field data with the growth and reproduction characteristics data of P. triangularis to establish the dynamic evolution equation of P. triangularis population, and use a hierarchical fusion gating mechanism to integrate multi-scale spatiotemporal characteristics to form a complete spatiotemporal evolution model of the impact of submarine CO2 leakage on P. triangularis.

2. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 1, characterized in that: The acidification sensitivity parameters of P. tricornutum specifically refer to the growth rate variation coefficient, photosynthesis efficiency reduction rate and calcification rate variation coefficient of P. tricornutum under different pH conditions.

3. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 2, characterized in that: The two-phase flow near-zone transport and diffusion model specifically refers to the PBM module of the Euler model established using the CFD software Fluent. The PBM module simulates the micro-scale process of bubbles and is suitable for simulating the leakage process of CO2.

4. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 3, characterized in that: The dose-effect function specifically refers to the mathematical relationship that describes the change in the population size of P. tricornutum with pH value, which is usually expressed as an S-shaped curve and includes pH threshold and maximum inhibition rate parameters.

5. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 4, characterized in that: The population dynamics evolution equation specifically refers to a set of partial differential equations that describes the spatiotemporal changes in population size by combining the growth rate, mortality rate, diffusion migration rate and environmental factor influence coefficients of the triangular brown algae.

6. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 5, characterized in that: The exposure dose specifically refers to the product of the degree to which the pH value of the water body in which the triangular lily is located deviates from the normal value and the duration, which is used to quantify the cumulative effect of acidification on the triangular lily.

7. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 6, characterized in that: Ocean current field data specifically refers to the horizontal and vertical flow velocity and direction information at each point in the sea area, which is used to calculate the water mass transport path and diffusion rate.

8. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 7, characterized in that: A leakage scenario specifically refers to a combination of different leakage rates, leakage durations and leakage tide times, which is used to simulate various possible leakage events.

9. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 8, characterized in that: The specific structure of the OceanAttn model is an ocean environment prediction model designed based on the Transformer architecture. It consists of two parts: an encoder and a decoder. The encoder adopts a six-layer multi-head attention mechanism, with each layer containing eight attention heads. The decoder also adopts a six-layer architecture with cross-layer connection characteristics.

10. The method for establishing an evolution model of the impact of submarine CO2 leakage on Phaeodactylum tricornutum according to claim 9, characterized in that: The core part of the OceanAttn model is a hierarchical attention network mechanism, which includes a spatial attention module and a temporal attention module. The spatial attention module captures the spatial correlation between ocean currents and pH distribution, while the temporal attention module processes the evolution of ocean environmental parameters over time.

Images