Simulation device and simulation method
The simulation device addresses the challenge of regional and seasonal variations in ocean carbon flux by using a Monte Carlo method with a marine biogeochemical model and surrogate learning, enhancing quantification precision and reducing computational costs.
Patent Information
- Application Number
- PCT/JP2024/012400
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-03-27
- Publication Date
- 2025-10-02
AI Technical Summary
Current global ocean biogeochemical models fail to accurately capture regional and seasonal variations in biological productivity, limiting the precision of ocean carbon flux quantification due to computational constraints and insufficient spatial resolution, which affects the global carbon budget.
A simulation device employing a Monte Carlo method with a marine biogeochemical model that initializes phytoplankton community-dependent parameters randomly, coupled with an ocean general circulation model, and uses a surrogate model trained with machine learning to estimate ocean carbon fixation, allowing for improved quantification and uncertainty analysis.
Enhances the accuracy of ocean carbon flux quantification by reflecting seasonal and regional variations, reducing computational demands, and providing uncertainty estimates for improved policymaking and climate prediction.
Smart Images
Figure JP2024012400_02102025_PF_FP_ABST
Abstract
Description
Simulation device and simulation method
[0001] The present disclosure relates to a simulation device and a simulation method.
[0002] The oceans are a sink for about 25±2% of the atmospheric carbon dioxide emitted by human activities, amounting to 2 petagrams per year (PgCyr). -1 In particular, primary production by marine phytoplankton photosynthesis is thought to account for 40% of the total carbon fixation (anthropogenic and non-anthropogenic) on Earth.
[0003] Actively quantifying the carbon stored in marine carbon sinks and understanding their variability is essential for policymaking and predicting future changes.
[0004] Hauck, Judith, et al., "Seasonally different carbon flux changes in the Southern Ocean in response to the southern annular mode," Global Biogeochemical Cycles 27.4 (2013), pp. 1236-1245Hauck, Judith, et al., "Consistency and challenges in the ocean carbon sink estimate for the global carbon budget," Frontiers in Marine Science, October 2020, Volume 7, Article 571720.Weber, T et al., “Technical note: Deep learning for creating surrogate models of precipitation in Earth system models,” Atmospheric Chemistry and Physics, (2020), Volume 20, pp. 2303-2317
[0005] Estimation of ocean carbon dioxide fluxes for the global carbon budget mainly involves mapping methods and global ocean biogeochemical models.
[0006] The mapping method uses statistical interpolation or neural network regression to map the global sea surface CO2 partial pressure based on measurements from the Surface Ocean CO2 Atlas published by SOCAT and other environmental datasets. However, the accuracy of this method is limited by the lack of temporal continuity and small spatial coverage of the observations.
[0007] The global ocean biogeochemical models used in the global carbon budget are coupled ocean biogeochemical models (GCMs) that can be used to project future changes in ocean carbon fluxes.
[0008] In the Regulated Ecosystem Model, version 2 (REcoM-2), a marine biogeochemical model, parameters related to respiration, grazing, nutrient intake, and excretion of marine microorganisms (phytoplankton, diatoms, zooplankton, etc.) are set as constants.
[0009] Global ocean biogeochemical models do not adequately capture regional and seasonal variations in biological productivity (carbon fixation through photosynthesis), limiting the accuracy of quantifying ocean carbon fluxes for the global carbon budget.
[0010] Earth system models require high-performance computers and are one of the most computationally intensive applications. To make calculations possible within the limitations of current high-performance computers, more efficient algorithms are needed. For example, due to current computational constraints, the maximum resolution of the global ocean physical-biogeochemical coupled model in the Coupled Model Intercomparison Project (CMIP) 6 is 0.25° (approximately 10 4 m scale). However, 3Different marine microbial communities have been observed at the 100-m scale, and this resolution is insufficient to analyze the spatial variation of marine microorganisms and the associated variations in ocean carbon fluxes.
[0011] The present disclosure has been made in light of the above, and aims to improve the quantification of ocean carbon fluxes and ocean carbon fixation.
[0012] A simulation device according to one aspect of the present disclosure is a simulation device that estimates an amount of ocean carbon fixation using a Monte Carlo method, and includes: an initialization unit that initializes functional parameters of a marine biogeochemical model that depend on the community composition of phytoplankton, to random values; a first simulator unit that links the marine biogeochemical model with an ocean general circulation model and executes the model; an estimation unit that estimates a first amount of ocean carbon fixation from simulation results; a learning unit that learns a surrogate model of the first simulator unit using pairs of the functional parameters and the first amount of ocean carbon fixation as learning data; and a second simulator unit that inputs the functional parameters initialized with random values into the surrogate model and estimates a second amount of ocean carbon fixation.
[0013] The present disclosure allows for improved quantification of ocean carbon fluxes and carbon sequestration.
[0014] Fig. 1 is a diagram showing an example of the configuration of a system for quantifying uncertainty in oceanic carbon fluxes using a surrogate model Monte Carlo method. Fig. 2 is a flowchart showing an example of the processing flow of the system for quantifying uncertainty in oceanic carbon fluxes. Fig. 3 is a flowchart showing an example of the processing flow of the system for quantifying uncertainty in oceanic carbon fluxes. Fig. 4 is a diagram showing an example of the hardware configuration of the system for quantifying uncertainty in oceanic carbon fluxes.
[0015] [System Configuration] Hereinafter, an embodiment of the present disclosure will be described with reference to the drawings.
[0016] Figure 1 shows an example of the configuration of a surrogate model Monte Carlo-based ocean carbon flux uncertainty quantification system according to this embodiment. The system shown in the figure is mainly composed of a modified ocean biogeochemistry model module, a Monte Carlo simulation module, and a surrogate Monte Carlo simulation module.The system shown in the figure comprises an experimental observation database (EXPERIMENTAL OBSERVATION DB) 101, an ocean biogeochemistry model unit (OCEAN BIOGEOCHEMISTRY MODEL UNIT) 102, a functional parameterization unit (BIOGEOCHEMISTRY MODEL FUNCTIONAL PARAMETERIZATION UNIT) 103, a randomized parameter initialization unit (RANDOMIZED PARAMETER INITIALIZATION UNIT) 104, an ocean general circulation model unit (OCEAN GENERAL CIRCULATION MODEL UNIT) 105, a coupling unit (COUPLING MODULE) 106, a simulator unit (SIMULATOR MODULE) 107, a carbon flux quantification unit (CARBON FLUX QUANTIFICATION MODULE) 108, a memory unit (MC SIMULATOR MODULE OUTPUT STORAGE UNIT) 109, and an input / output parameter storage unit (INPUT / OUTPUT PARAMETER STORAGE UNIT) 110. The system includes a metamodel training unit (METAMODEL TRAINING UNIT) 121, a metamodel storage unit (METAMODEL STORAGE UNIT) 123, a randomized parameter initialization unit (RANDOMIZED PARAMETER INITIALIZATION UNIT) 124, a metamodel simulator unit (METAMODEL SIMULATOR UNIT) 125, a storage unit (SMC SIMULATOR MODULE OUTPUT STORAGE UNIT) 126, and a carbon flux uncertainty quantification unit (CARBON FLUX UNCERTAINTY QUANTIFICATION UNIT) 130. Hereinafter, the Monte Carlo simulation module will be referred to as the MC simulation module, and the surrogate Monte Carlo simulation module will be referred to as the SMC simulation module.
[0017] [Improved Marine Biochemistry Module] The experimental observation database 101 holds statistical data on carbon uptake, carbon excretion, and growth of phytoplankton obtained through experiments or observations. For example, the experimental observation database 101 holds statistical data on phytoplankton genes related to growth rate, such as genes determining light-harvesting efficiency, uptake rates of various nutrients, nutrient storage amounts, and cell size, obtained through experiments or observations.
[0018] The marine biogeochemical model section 102 includes standard marine biogeochemical model equations, such as REcoM-2, PlankTOM5, HAMOCC, PISCES, and COBALT.
[0019] The functional parameterization unit 103 identifies parameters of the marine biogeochemical model that vary depending on the composition of the phytoplankton community, such as respiration, grazing, nutrient uptake, and excretion of the model organism, and replaces the parameters with statistical distributions based on experimental observation data from the experimental observation database 101.
[0020] The randomized parameter initialization unit 104 randomly initializes the parameters according to the statistical distribution defined by the function parameterization unit 103. The generated random combination of parameters is used as the initial condition for each simulation run.
[0021] An example of randomly initializing parameters is now described. Here, an example of one model equation is shown, but the same technique can be applied to other model equations.
[0022] In REcoM-2, nanophytoplankton carbon biomass increases as a result of carbon assimilation during photosynthesis. The rate of change of nanophytoplankton carbon biomass, S(C phy ) is expressed by the following formula:
[0023]
[0024] where p phy is the photosynthetic rate of nanophytoplankton, ε phy Cis the carbon emission of nanophytoplankton, f phy lim is the limiting function for nitrogen assimilation in nanophytoplankton, r phy is the respiration rate of nanophytoplankton, g is the aggregation rate, and q phy is the intracellular N:C ratio of nanophytoplankton, and G phy is the grazing rate of nanophytoplankton.
[0025] ε phy C and q phy is a constant parameter defined in the model. The function parameterization unit 103 refers to the experiment observation database 101 and calculates ε phy C and q phy is replaced with a statistical distribution. For example, in the experimental observation database 101, phy C is normally distributed with (μ,σ)=(0.05, 0.01), and q phy Suppose we have experimental or observational data that is normally distributed with (μ,σ)=(0.20, 0.05), where μ is the mean and σ is the standard deviation.
[0026] The randomization parameter initialization unit 104 determines ε phy C and q phy For each of the N models, generate random numbers with a normal distribution and a specified mean and standard deviation, and generate initial conditions for the N model runs.
[0027] Note that the above is a very simplified example. In practice, an ensemble of 50 or more parameters is initialized depending on the model. Other probability distributions, such as beta, binomial, and Poisson distributions, can also be applied, in addition to the normal distribution. Furthermore, parameter initialization can be adjusted based on the ocean region and physical conditions of the target grid cell.
[0028] The marine biogeochemistry model unit 102, the function parameterization unit 103, and the randomization parameter initialization unit 104 constitute a modified ocean biochemistry model module.
[0029] [Monte Carlo Simulation Module] The ocean general circulation model unit 105 includes ocean physical model equations. Examples of ocean general circulation models include MITgcm, NEMO, MICOM, MPIOM, CSIRO, and MOM6.
[0030] The coupling section 106 allows data exchange of shared parameters between the improved marine biogeochemical model module and the ocean general circulation model section 105 .
[0031] The simulator section 107 calculates the predictor variables specified in the improved marine biogeochemical model module and the ocean general circulation model section 105 for each grid cell within the model domain at each time step.
[0032] The carbon flux quantification unit 108 calculates the ocean carbon flux defined by the marine biogeochemical model unit 102 based on the simulation results of the simulator unit 107, and outputs the amount of fixed carbon.
[0033] The Monte Carlo simulation module is made up of the coupling unit 106, the simulator unit 107, the carbon flux quantification unit 108, and the randomization parameter initialization unit 104. The Monte Carlo simulation module repeats parameter initialization in the randomization parameter initialization unit 104 and simulation in the simulator unit 107 N times, and obtains the amount of carbon fixation for each time from the carbon flux quantification unit 108.
[0034] The storage device 109 stores the output of the Monte Carlo simulation module. When the Monte Carlo simulation module repeats the simulation N times, the storage device 109 stores Q n=0 From Q n=N Maintain carbon fixation up to
[0035] [Surrogate Monte Carlo Simulation Module] The input / output parameter storage unit 121 stores pairs of input parameters and carbon fixation amounts when the MC simulation module executes the simulation N times. The input parameters are obtained from the randomization parameter initialization unit 104, and the carbon fixation amount is obtained from the storage device 109.
[0036] The metamodel learning unit 122 uses the input / output parameter pairs stored in the input / output parameter storage unit 121 as learning data to learn a metamodel (surrogate model). The surrogate model is a machine learning model trained using the results of a simulation using a physical model, and can execute simulations with lower computational costs than physical models. Methods for learning the metamodel include neural networks, polynomial regression, Gaussian process regression, and support vector machines.
[0037] The metamodel storage unit 123 stores the metamodel learned by the metamodel learning unit 122, that is, the metamodel Q that infers the amount of carbon fixation when parameters are input. * = f(x1,...,xm) where Q * is the estimated amount of ocean carbon fixation, and x1,...,xm are the m input parameters of the ocean physical-biochemical coupled model.
[0038] Similar to the randomization parameter initialization unit 104, the randomization parameter initialization unit 124 randomly initializes parameters according to the statistical distribution defined by the function parameterization unit 103. The generated random combination of parameters is used as the initial condition for each simulation run.
[0039] The metamodel simulator unit 125 inputs the input parameters acquired from the randomization parameter initialization unit 124 into the metamodel, and calculates the carbon fixation amount Q * Estimate.
[0040] The SMC simulation module is configured with an input / output parameter storage unit 121, a metamodel learning unit 122, a metamodel storage unit 123, a randomization parameter initialization unit 124, and a metamodel simulator unit 125. After learning the metamodel, the SMC simulation module initializes the parameters in the randomization parameter initialization unit 124 and performs the simulation in the metamodel simulator unit 125. * The process is repeated times to obtain the amount of carbon fixed each time.
[0041] The storage device 126 stores the output of the SMC simulation module. * When the simulation is repeated Q times, the storage device 126 stores * i=0 From Q * i=N* Maintain carbon fixation up to
[0042] The carbon flux uncertainty quantification unit 130 estimates the uncertainty of the amount of ocean carbon fixation during the simulation period using the statistical distribution of the output values stored in the storage device 109 and the storage device 126. For example, the average value Qav of the amount of carbon fixation is calculated using the following equation:
[0043]
[0044] Here, Q n is the amount of carbon fixation obtained in the nth Monte Carlo simulation, and N is the number of times the simulation was repeated. * i is the amount of carbon fixation obtained in the i-th surrogate Monte Carlo simulation, N * is the number of times the simulation was repeated.
[0045] The carbon flux uncertainty quantification unit 130 outputs the average value Qav of the carbon fixation amount in the ocean during the simulation period and the uncertainty x (Qav±x%). n , Q * i The uncertainty x can be found by calculating the variance and standard deviation of
[0046] [System Operation] Next, an example of the processing flow of this system will be described with reference to the flowcharts of Figures 2 and 3. Figure 2 is a flowchart showing the processing flow by the MC simulation module. Figure 3 is a flowchart showing the processing flow by the SMC simulation module.
[0047] In step S11, the Monte Carlo simulation initializes counters n and i that count the number of times the simulation is executed.
[0048] In step S12, the function parameterization unit 103 refers to the data held in the experimental observation database 101 and replaces the function parameters of the marine biogeochemical model that depend on the microbial community composition with appropriate statistical distributions.
[0049] In step S13, the MC simulation module determines whether the number of simulation executions n is less than N. If the number of simulation executions n is less than N, the process proceeds to step S14, and if the number of simulation executions n is N or more, the process proceeds to step S21 in FIG.
[0050] When a simulation is performed, in step S14, the randomized parameter initialization unit 104 randomly initializes each parameter of the improved marine biogeochemical model according to its respective statistical distribution.
[0051] In step S15, the linking unit 106 links the improved marine biogeochemical model, the parameters of which have been randomly initialized, with the ocean general circulation model.
[0052] In step S16, the simulator unit 107 calculates the predictor variables specified in the improved marine biogeochemical model and the ocean general circulation model for each grid cell in the model domain for each time step.
[0053] In step S17, the carbon flux quantification unit 108 stores the carbon fixation amount Qn in the storage device 109. The MC simulation module also stores the parameters used in the simulation and the calculated carbon fixation amount Qn. n are stored as a pair in the input / output parameter storage unit 121.
[0054] In step S18, the MC simulation module adds 1 to the number of times n the simulation has been executed.
[0055] When the number of times n of simulations performed by the MC simulation module reaches N, in step S21 of FIG. 3, the metamodel learning unit 122 uses N pairs of parameters and carbon fixation amounts stored in the input / output parameter storage unit 121 as learning data to learn a metamodel for calculating the carbon fixation amount.
[0056] In step S22, the SMC simulation module determines whether the number of executions of the simulation by the metamodel i is N * It is determined whether the number of simulation executions i is less than N * If it is less than N, the process proceeds to step S23. * If the number of times is equal to or greater than the number of times, the process proceeds to step S25.
[0057] In step S23, the metamodel simulator unit 125 inputs the parameters initialized with random values by the randomization parameter initialization unit 124 into the metamodel trained in step S21, and calculates the carbon fixation amount Q * i The SMC simulation module stores the estimated amount of carbon fixation in the storage device 126.
[0058] In step S24, the SMC simulation module adds 1 to the number of times the simulation has been executed.
[0059] The number of times the simulation is executed by the SMC simulation module is N. * If the number of times has reached, in step S25, the carbon flux uncertainty quantification unit 130 compares the output values of the N simulations stored in the storage device 109 with the N simulations stored in the storage device 126. * These estimates are used to estimate the uncertainty in ocean carbon fixation and ocean carbon flux during the simulation period.
[0060] In step S26, the system outputs the calculated ocean carbon fixation amount and ocean carbon flux and their uncertainties.
[0061] As described above, the ocean carbon flux uncertainty quantification system using the surrogate model Monte Carlo method of this embodiment includes a randomized parameter initialization unit 104 that initializes functional parameters of an ocean biogeochemical model that depend on the community composition of phytoplankton with random values; a simulator unit 107 that links an ocean biogeochemical model with an ocean general circulation model and executes the model; a carbon flux quantification unit 108 that estimates the amount of ocean carbon fixation Q from the simulation results; a metamodel training unit 122 that trains a metamodel of the Monte Carlo simulation module using pairs of the functional parameters and the amount of ocean carbon fixation Q as training data; and a surrogate model that inputs the functional parameters initialized with random values by the randomized parameter initialization unit 124 into the surrogate model and estimates the amount of ocean carbon fixation Q. * The system includes an experimental observation database 101 that stores statistical data on phytoplankton obtained through experimental observations, and a functional parameterization unit 103 that replaces functional parameters of the marine biogeochemical model that depend on the community composition of phytoplankton with statistical distributions according to the statistical data stored in the experimental observation database 101, and a randomization parameter initialization unit 104 that initializes the functional parameters with random values according to the statistical distributions.
[0062] Marine phytoplankton form genetically diverse communities in the ocean and constantly adapt to the physical conditions of their environment. The composition of phytoplankton communities in a given ocean region is thought to be determined by random events such as random genetic assortment and ocean mixing. Therefore, this system adds flexibility to the model to reflect seasonal and regional variations in phytoplankton primary production, introducing randomness into phytoplankton community composition and function, and reflecting this in the uncertainty of ocean carbon flux estimates. This can improve the quantification of ocean carbon fluxes.
[0063] Additionally, the Surrogate Monte Carlo simulation module allows for larger sample sizes (N + N * ) can reduce the computation time required to obtain
[0064] The above-described ocean carbon flux uncertainty quantification system can be implemented, for example, by a general-purpose computer system including a central processing unit (CPU) 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906, as shown in Fig. 4. In this computer system, the ocean carbon flux uncertainty quantification system is realized by the CPU 901 executing a predetermined program loaded onto the memory 902. This program can be recorded on a computer-readable non-transitory recording medium such as a magnetic disk, an optical disk, or a semiconductor memory, or can be distributed via a network.
[0065] 101 Experimental Observation Database (EXPERIMENTAL OBSERVATION DB) 102 Ocean Biogeochemistry Model Unit (OCEAN BIOGEOCHEMISTRY MODEL UNIT) 103 Functional Parameterization Unit (BIOGEOCHEMISTRY MODEL FUNCTIONAL PARAMETERIZATION UNIT) 104 Randomized Parameter Initialization Unit (RANDOMIZED PARAMETER INITIALIZATION UNIT) 105 Ocean General Circulation Model Unit (OCEAN GENERAL CIRCULATION MODEL UNIT) 106 Coupling Unit (COUPLING MODULE) 107 Simulator Unit (SIMULATOR MODULE) 108 Carbon Flux Quantification Unit (CARBON FLUX QUANTIFICATION MODULE) 109 Storage Unit (MC SIMULATOR MODULE OUTPUT STORAGE UNIT) 121 Input / Output Parameter Storage Unit (INPUT / OUTPUT PARAMETER STORAGE UNIT 122 METAMODEL TRAINING UNIT 123 METAMODEL STORAGE UNIT 124 RANDOMIZED PARAMETER INITIALIZATION UNIT 125 METAMODEL SIMULATOR UNIT 126 SMC SIMULATOR MODULE OUTPUT STORAGE UNIT 130 CARBON FLUX UNCERTAINTY QUANTIFICATION UNIT
Claims
1. A simulation device that estimates marine carbon fixation using the Monte Carlo method, comprising: an initialization unit that initializes functional parameters of a marine biogeochemical model that depend on the community composition of phytoplankton to random values; a first simulator unit that links the marine biogeochemical model with an ocean general circulation model and executes the model; an estimation unit that estimates a first marine carbon fixation amount from the simulation results; a learning unit that learns a surrogate model of the first simulator unit using pairs of the functional parameters and the first marine carbon fixation amount as learning data; and a second simulator unit that inputs the functional parameters initialized to random values into the surrogate model and estimates a second marine carbon fixation amount.
2. A simulation device according to claim 1, comprising: a database that holds statistical data on phytoplankton obtained through experimental observation; and a replacement unit that replaces functional parameters of the marine biogeochemical model that depend on the community composition of phytoplankton with statistical distributions that correspond to the statistical data held in the database, and the initialization unit that initializes the functional parameters with random values that correspond to the statistical distributions.
3. A simulation device according to claim 1, comprising a calculation unit that calculates the uncertainty of the amount of ocean carbon fixation estimated from the first amount of ocean carbon fixation and the second amount of ocean carbon fixation.
4. A simulation method for estimating the amount of ocean carbon fixation using the Monte Carlo method, comprising the steps of: initializing functional parameters of a marine biogeochemical model that depend on the community composition of phytoplankton to random values; linking the marine biogeochemical model with an ocean general circulation model and running the model; estimating a first amount of ocean carbon fixation from the simulation results; training a surrogate model using pairs of the functional parameters and the first amount of ocean carbon fixation as training data; and inputting the functional parameters initialized to random values into the surrogate model to estimate a second amount of ocean carbon fixation.