Tsunami propagation and submergence real-time forecasting algorithm based on AI acceleration

Abstract

The invention relates to the technical field of tsunami early warning, and discloses a tsunami propagation and submerging real-time forecasting algorithm based on AI acceleration, and the algorithm comprises the steps: obtaining a stability constraint coefficient based on a local CFL number and a Froude number through a stability constraint model; obtaining a dynamic precision coefficient based on the wave height gradient and the dry-wet boundary speed through a dynamic precision model; integrating the coefficients, the distance from the grid point to the seismic source and the remaining calculation time through an adaptation degree model, and outputting a physical-business adaptation degree; outputting a system efficiency coefficient based on the calculated load, the hardware temperature and the power consumption through a system efficiency model; and finally, the model is adjusted to dynamically output the target numerical integration time step length according to the adaptation degree, the efficiency coefficient, the maximum step length and the minimum step length. According to the method, intelligent dynamic balance of the time step among the physical stability, the calculation precision, the business urgency and the system efficiency is realized through multi-model cooperation, so that the calculation efficiency of tsunami real-time forecasting is remarkably improved while the forecasting precision is ensured.

Description

Technical Field

[0001] This invention belongs to the field of tsunami early warning technology, and in particular relates to an AI-accelerated real-time forecasting algorithm for tsunami propagation and inundation. Background Technology

[0002] As a highly destructive marine disaster, rapid and accurate real-time forecasting of tsunamis is crucial for disaster prevention and mitigation, and for protecting the lives and property of people in coastal areas. Currently, the forecasting of tsunami propagation and inundation processes mainly relies on numerical simulation technology, which uses fluid dynamics models such as shallow water wave equations to simulate the generation, propagation, and inundation of tsunami waves.

[0003] However, achieving real-time or near-real-time tsunami forecasting faces enormous challenges, with the core contradiction lying in balancing computational efficiency and simulation accuracy. On the one hand, to ensure timely forecasting, large-scale, high-resolution numerical calculations must be completed within a very limited timeframe. On the other hand, the physical processes of tsunamis are complex, involving various phenomena such as wave propagation, nonlinear interactions, and dynamic changes in wet and dry boundaries, placing extremely high demands on the stability and accuracy of numerical calculations. Traditional numerical simulations typically employ fixed time steps or simple adaptive step-size strategies based on a single physical condition (such as global CFL conditions). Fixed time steps often force the use of conservative, small step sizes to meet the stability and accuracy requirements of the most demanding regions (such as wavefronts and inundation fronts), resulting in a large amount of unnecessary computation in regions with smooth physical processes or non-critical areas, severely restricting overall efficiency. While simple adaptive strategies can offer some improvement, they typically fail to comprehensively consider the operational constraints of real-time forecasting (such as remaining computation time and regional importance) and the system state of heterogeneous parallel computing environments (such as load balancing and hardware health).

[0004] In recent years, artificial intelligence (AI) technology has provided new ideas for accelerating scientific computing. However, in the field of tsunami numerical forecasting, how to deeply integrate AI methods with complex physical models, urgent operational needs, and dynamic computing resources to build a core algorithm that can intelligently and adaptively schedule computing resources and dynamically optimize computing parameters remains a key technical problem that urgently needs to be solved. Existing technologies lack a comprehensive solution that can simultaneously perceive the dynamic characteristics of the physical field, the urgency of forecasting operations, and the real-time status of the computing system, and accordingly optimize the numerical integration step size. This makes it difficult to achieve the optimal balance between forecast accuracy and computational efficiency under strict time constraints. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-accelerated real-time forecasting algorithm for tsunami propagation and inundation, aiming to solve the above-mentioned problems.

[0006] This invention is implemented as follows: an AI-accelerated real-time tsunami propagation and inundation forecasting algorithm, comprising the following steps: obtaining stability constraint coefficients through a stability constraint model based on local CFL numbers and local flow Froude numbers; outputting dynamic accuracy coefficients through a dynamic accuracy model based on wave height spatial gradient norm and dry / wet boundary movement velocity; outputting physical-operational fit through a fit model based on grid point-to-source distance and remaining global computation time under the stability constraint coefficients and dynamic accuracy coefficients; outputting system efficiency coefficients through a system efficiency model based on computational load balancing, hardware temperature, and hardware power consumption; and adjusting the time step of the model output target numerical integration based on the physical-operational fit and system efficiency coefficients.

[0007] A further technical solution is that the operation process of the adjustment model is as follows: based on the physical-service adaptability and system efficiency coefficient, the target numerical integration time step is calculated by weighted synthesis function and combined with minimum and maximum time steps; wherein, the physical-service adaptability participates in the calculation in a positive weighted manner, the system efficiency coefficient participates in the calculation in a negative weighted manner, and the weight allocation between the physical-service adaptability and the system efficiency coefficient is realized through adjustable weight coefficients.

[0008] A further technical solution is that the system efficiency model operates as follows: the hardware health status is obtained based on hardware temperature and hardware power consumption; based on the differences in the computation time distribution of all parallel processes, the computation load balance degree, which reflects the degree of load balance, is obtained by calculating the complement of the relative deviation between the maximum computation time and the average computation time; the computation load balance degree is multiplied by the hardware health status to obtain the system efficiency coefficient.

[0009] A further technical solution involves the following method for obtaining hardware health status: The difference between the current hardware temperature and the nominal operating temperature is compared with the difference between the maximum allowable hardware operating temperature and the nominal operating temperature. The ratio is then limited to a lower limit of 0 using a max function to obtain the hardware temperature index. The difference between the current hardware power consumption and the normal power consumption is compared with the difference between the maximum safe power consumption and the normal power consumption. The ratio is then limited to a lower limit of 0 using a max function to obtain the hardware power consumption index. Finally, the larger of the hardware temperature index and the hardware power consumption index is taken, and their complement is calculated to obtain the hardware health status coefficient, which reflects the overall health status of the hardware.

[0010] A further technical solution involves the following operational flow of the adaptability model: The remaining global computation time is compared with the maximum allowable computation time set by the system. After normalization and upper limit constraints, the complement is calculated to obtain a time urgency index reflecting the degree of time pressure. The distance from the grid point to the seismic source is compared with the maximum impact distance. After normalization and upper limit constraints, the complement is calculated to obtain a target area priority weight reflecting the importance of the area. The complement of the stability constraint coefficient and the dynamic accuracy coefficient is multiplied to obtain a physical constraint factor. The time urgency index is multiplied with the target area priority weight to obtain an operational urgency factor. The physical constraint factor is divided by the sum of the physical constraint factor and the operational urgency factor to obtain the physical-operational adaptability. When physical conditions allow and operational urgency is low, the adaptability value increases, tending to use a larger time step to improve computational efficiency. When operational urgency is high or physical conditions are limited, the adaptability value decreases, tending to use a smaller time step to ensure computational accuracy.

[0011] A further technical solution involves the following operational flow of the stability constraint model: The local CFL number is compared with the upper limit of the Courant number stability allowed by the numerical format to obtain the local CFL number exponent; when the local flow Froude number does not exceed the critical Froude number, the local flow Froude number exponent is obtained through normalization calculation; when the local flow Froude number exceeds the critical Froude number, the local flow Froude number exponent is calculated through a linear penalty function, and the penalty coefficient is proportional to the magnitude of exceeding the critical value; the reciprocals of the local CFL number exponent and the local flow Froude number exponent are calculated respectively, and the smaller of the two is taken as the stability constraint coefficient. The calculation logic of this stability constraint coefficient ensures that both the CFL stability condition and the flow state stability condition are satisfied simultaneously. A higher stability constraint coefficient indicates that the current physical field state allows for a larger time step scaling ratio.

[0012] A further technical solution involves the following operation flow of the dynamic accuracy model: the wave height spatial gradient norm and the dry-wet boundary moving velocity are compared with the historical maximum wave height gradient and the theoretical maximum inundation velocity to obtain the wave height spatial gradient norm index and the dry-wet boundary moving velocity index; the wave height spatial gradient norm index and the dry-wet boundary moving velocity index are then combined in a weighted linear manner to obtain the dynamic accuracy coefficient; the higher the dynamic accuracy coefficient, the higher the complexity of the current flow field and the stronger the demand for high temporal resolution.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Achieving an adaptive optimal balance between computational efficiency and simulation accuracy: By accurately perceiving the stability boundary and complexity requirements of the physical field through a stable constraint model and a dynamic accuracy model, and by coupling the fitness model with business urgency and regional importance, the time step can be finely adjusted according to the spatial heterogeneity of the physical process and business priority, avoiding the problems of overcomputation in stable regions and insufficient accuracy in critical regions by traditional fixed step size or simple adaptive strategies.

[0014] 2. A multi-dimensional decision-making system integrating physical mechanisms, operational requirements, and system status was constructed: For the first time, the algorithm incorporated physical field characteristics (CFL number, Froude number, wave height gradient, flooding velocity), real-time operational constraints (remaining time, regional importance), and computational system status (load balancing, hardware health) into a unified optimization framework. By adjusting the model for collaborative decision-making, it achieved a leap from "single physical drive" to "physical-operation-system" collaborative drive, significantly improving the overall intelligence and environmental adaptability of the forecasting system.

[0015] 3. Significantly improves resource utilization efficiency and system robustness of large-scale parallel computing: The system efficiency model monitors hardware status and load balancing in real time and feeds back the efficiency coefficient to the step size adjustment, ensuring the health and stability of hardware under high load computing, optimizing the scheduling and resource allocation of parallel tasks, thereby improving the overall throughput and reliability of the cluster, and providing solid system support for minute-level real-time forecasting.

[0016] 4. Effectively meets the stringent requirements of timeliness and accuracy for tsunami warnings: Through a dynamic step size adjustment mechanism, the algorithm can use a larger step size to advance quickly in the early stage of the forecast or in non-critical areas, and automatically shrink the step size to ensure detailed capture as the forecast deadline approaches or in critical flooded areas. Thus, within strict time constraints, it maximizes the forecast accuracy of critical areas and provides more reliable and timely technical support for emergency decision-making. Attached Figure Description

[0017] Figure 1 The flowchart shows a real-time forecasting algorithm for tsunami propagation and inundation based on AI acceleration, which is provided for this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In traditional tsunami numerical forecasting, a uniform time step strategy is adopted across the entire region, which fails to adapt to the spatiotemporal heterogeneity of physical processes. Specifically, in deep-sea areas far from the epicenter or where physical processes are relatively calm, a larger time step is sufficient to maintain computational stability and accuracy; however, in nearshore fracture zones, wet-dry boundaries, and areas with steep wavefronts, an extremely small time step is required to avoid numerical instability and capture critical details. Simultaneously, the uneven distribution of computational load in a parallel environment reduces hardware resource utilization efficiency, and dynamic changes in hardware status (including temperature and power consumption) are not incorporated into the computation scheduling mechanism. Furthermore, the real-time constraint of the forecasting task—that is, the computation completion time must meet the warning deadline—is not explicitly modeled as a core input. Thus, the contradiction between computational efficiency, forecast accuracy, system stability, and real-time response capability is exacerbated.

[0020] For example, in simulating a tsunami event triggered by an undersea earthquake, the computational domain covers the area from the epicenter to coastal cities. During the deep-sea propagation phase, wave morphology changes slowly, and a uniform small time step is enforced across the entire domain, resulting in a large amount of redundant computation. When the tsunami wave approaches the coastline, the spatial gradient of wave height increases sharply at the wet-dry boundary, and the wet-dry boundary moves faster. A uniform large time step is then applied across the entire domain, causing numerical oscillations and simulation failure. Simultaneously, some computing nodes in the GPU cluster become overloaded due to processing the complex near-shore area, causing a rapid rise in hardware temperature. This triggers the frequency reduction protection mechanism, extending the overall computation time and preventing the forecast results from being generated within the warning timeframe.

[0021] If the above problems are not addressed, the accuracy of tsunami forecasts will be reduced, and the risk of missed or false alarms in key flooded areas will increase; insufficient computational efficiency will prevent the forecasting process from meeting the timeliness requirements of minute-level early warnings; inefficient use of hardware resources and the risk of overheating will increase the probability of system failure and weaken the reliability of the forecasting system; ultimately, the limitations of disaster early warning capabilities will directly threaten the safety of people and property in coastal areas, and the effectiveness of emergency response will be constrained.

[0022] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0023] like Figure 1 As shown, an AI-accelerated real-time tsunami propagation and inundation forecasting algorithm provided in one embodiment of the present invention includes the following steps: The system acquires local CFL number, local flow Froude number, wave height spatial gradient norm, wet-dry boundary migration velocity, grid point-to-source distance, remaining global computation time, computational load balancing, and hardware health status; these parameters form the basis for subsequent evaluation and decision-making. For example, these parameters can be obtained by directly reading the current computational status from the numerical simulator, acquiring hardware metrics from the system monitoring interface, or through preset configuration information. In one implementation, the system can periodically extract this data from the simulation core, operating system, and preset configuration files.

[0024] Based on the local CFL number and the local flow Froude number, the stability constraint coefficients are obtained through a stability constraint model. Based on the wave height spatial gradient norm and the moving speed of the wet and dry boundary, dynamic accuracy coefficients are output through a dynamic accuracy model. Based on the distance from the grid point to the seismic source under the stability constraint coefficient and dynamic accuracy coefficient, as well as the remaining global calculation time, the physical-business adaptability is output through the adaptability model. Based on the calculation of load balancing, hardware temperature, and hardware power consumption, the system efficiency coefficient is output through the system efficiency model. Based on the physical-business adaptability and system efficiency coefficient, the time step of the model output target value integration is adjusted.

[0025] The following example will provide a more detailed explanation of the above technical solution: Suppose a tsunami forecasting mission requires real-time simulation of a vast ocean area. At the start of the simulation, the system first acquires various parameters for the current moment. For example, in deep-sea regions, the local CFL number and local flow Froude number may be low, the wave height spatial gradient norm and the dry-wet boundary movement velocity are close to zero, the distance from the grid points to the epicenter is large, and the remaining global computation time is relatively ample. Simultaneously, the computational load is well-balanced, and hardware temperature and power consumption remain at normal levels.

[0026] At this point, the stability constraint model outputs a high stability constraint coefficient based on the low local CFL number and local flow Froude number, indicating that the region allows for a larger time step in terms of stability. The dynamic accuracy model outputs a low dynamic accuracy coefficient based on the near-zero wave height spatial gradient norm and the dry / wet boundary movement velocity, indicating that the region does not have high accuracy requirements and can use a larger time step. The fitness model integrates this information and considers the large grid point-to-source distance and ample remaining global computation time, outputting a low physical-operational fitness, indicating that the region's time step requirement is not so urgent. The system efficiency model outputs a high system efficiency coefficient based on high load balancing, normal hardware temperature and power consumption, indicating that the system is operating well. Finally, the tuning model, based on these inputs, outputs a relatively large target numerical integration time step to improve computational efficiency in the deep-sea region.

[0027] As tsunami waves propagate to the nearshore region, the situation changes. Local CFL numbers and local flow Froude numbers may increase due to shallower water and increased flow velocity, the wave height spatial gradient norm increases significantly (waves become steeper), and the wet-dry boundary movement becomes more pronounced (the inundation process begins). Simultaneously, the distance from the grid points to the epicenter decreases, potentially making the remaining global computation time more critical. Under prolonged high-load computation, hardware temperature and power consumption may rise, leading to a slight decrease in computational load balance.

[0028] In this scenario, the stability constraint model outputs a low stability constraint coefficient based on the high local CFL number and local flow Froude number to ensure numerical stability. The dynamic accuracy model outputs a high dynamic accuracy coefficient based on the increased wave height spatial gradient norm and the moving velocity of the wet and dry boundaries, emphasizing the region's need for high temporal resolution. The fit model integrates this information, taking into account the small grid point-to-source distance and limited remaining global computation time, and outputs a high physical-operational fit, indicating a very urgent need for time steps in this region. The system efficiency model outputs a slightly low system efficiency coefficient based on a slightly decreased load balancing degree and increased hardware temperature and power consumption, indicating a decrease in system operating efficiency. Finally, the adjustment model, based on these inputs, outputs a relatively small target numerical integration time step to ensure computational accuracy and stability in the nearshore region while also considering operational urgency.

[0029] Compared to existing localized and static AI enhancement solutions, this application achieves global and dynamic optimization of the computation process. For example, existing AI models may only be used to accelerate a certain submodule or optimize a fixed parameter, without overall scheduling based on the dynamic state of real-time computation. This application, through the synergistic effect of physical-business adaptability and system efficiency coefficients, enables the adjustment of the target numerical integration time step to simultaneously respond to the evolution of the physical process, the priority of business needs, and the availability of hardware resources. This technical concept, which deeply integrates physical mechanisms, real-time computation status, and early warning business needs, effectively solves the prominent problems of traditional methods in terms of computational efficiency and accuracy, hardware resource utilization efficiency, and real-time constraints of forecasting tasks, providing a new technical path for achieving minute-level or near-real-time high-precision tsunami forecasts.

[0030] like Figure 1 As shown, in a preferred embodiment of the present invention, the operation flow of the adjustment model is as follows: Based on physical-service adaptability and system efficiency coefficient, the target numerical integration time step is calculated by weighted synthesis function and combined with minimum and maximum time step; Among them, the physical-service adaptability is calculated in a positive weighted manner, the system efficiency coefficient is calculated in a negative weighted manner, and the weight allocation between the physical-service adaptability and the system efficiency coefficient is achieved through an adjustable weight coefficient. The adjustment model can specifically be: ,in The time step for integrating the target numerical value. For physical-business adaptability, The system efficiency coefficient. and For the minimum and maximum time steps, and The minimum and maximum allowable time steps can be selected based on historical experience; The fit weighting coefficient, ranging from 0 to 1, is used to balance the relative importance of physical-business fit and system efficiency coefficients in time step adjustment. A larger value indicates that the algorithm focuses more on meeting the time step requirements of the physical process and business needs; when A smaller value indicates that the algorithm prioritizes optimizing the time step using system efficiency. This coefficient can be set empirically based on the specific forecasting task, computing resource allocation, and different priorities for accuracy and real-time performance. For example, a larger value can be set when higher accuracy is required in the early stages of forecasting. Value; however, when time is tight in the later stages of the forecast, it can be appropriately reduced. The value is prioritized to ensure calculation speed. Reinforcement learning or adaptive weighting mechanisms can also be introduced to acquire these technologies.

[0031] As a specific implementation method, the operation of the adjustment model can be implemented as follows: In a tsunami numerical forecasting system, the system first periodically obtains the current physical-operational fit and system efficiency coefficient from various sensors and computing modules. For example, the physical-operational fit may be calculated by a dedicated module based on factors such as the current wave height gradient, wet / dry boundary velocity, and remaining forecast time. Its value ranges from 0 to 1, where 1 indicates that the current region or task has the strongest need for a large step size. The system efficiency coefficient may be calculated by another module based on the real-time load, temperature, power consumption, and load balancing of parallel tasks of the CPU / GPU. Its value also ranges from 0 to 1, where 1 indicates that the system operates with the highest efficiency and the hardware is in the best condition. Assuming the preset minimum allowable time step is 5 seconds and the maximum allowable time step is 20 seconds, the fit weight coefficient is set to 0.6. At a certain moment, if the calculated physical-operational fit is 0.8 (indicating that the current region needs a large step size) and the system efficiency coefficient is 0.9 (indicating that the system is operating well), the adjustment model imports these values ​​into the formula to calculate 12.8 seconds. Therefore, the final target numerical integration time step was determined to be 12.8 seconds and passed to the numerical solver for the next round of calculation. In this way, the algorithm can dynamically adjust the time step according to the real-time physical state, business requirements, and system performance, thereby maximizing computational efficiency while ensuring forecast accuracy and stability.

[0032] Through the above technical solution, this algorithm provides a specific and efficient numerical integration time step adjustment mechanism. This mechanism organically combines physical-operational adaptability with system efficiency coefficients and introduces adjustable weighting coefficients, achieving a dynamic balance between physical demands, operational urgency, and computational resource utilization in the time step. This effectively solves the problems of fixed time steps or insufficient local optimization in traditional tsunami forecasting, avoiding instability or distortion caused by excessively large step sizes in critical areas, and preventing computational resource waste caused by excessively small step sizes in flat areas. Overall, this solution enables tsunami propagation and inundation forecasting algorithms to more intelligently adapt to complex physical processes and variable computational environments, significantly improving the real-time performance and reliability of forecasts, and providing solid technical support for minute-level or near-real-time tsunami warnings.

[0033] like Figure 1 As shown, in a preferred embodiment of the present invention, the operation flow of the system efficiency model is as follows: The system acquires information on computational load balancing, hardware temperature, and hardware power consumption. Computational load balancing refers to the evenness of task distribution among computing units in a parallel computing system. It can be obtained in various ways, such as by querying the CPU or GPU utilization, memory usage, and I / O latency of each process in real time through the application programming interface (API) of the operating system or high-performance computing (HPC) scheduler, and calculating the load distribution accordingly. Alternatively, in a parallel computing framework, each computing node periodically reports its current task queue length, number of completed tasks, and remaining task quantity, which is then aggregated and calculated by the master node. Hardware temperature refers to the real-time temperature of the computing device during operation. It can be obtained by reading temperature data through hardware sensors (such as the CPU or GPU's built-in temperature sensors) and acquiring it through interfaces provided by the operating system. Alternatively, it can be achieved by non-contact temperature monitoring of key hardware components using external thermocouples or infrared sensors, and transmitting the data to the control system. Hardware power consumption refers to the real-time electrical energy consumption of computing devices during operation. It can be obtained through the power management unit (PMU) or a dedicated power measurement chip to obtain real-time power consumption data of the CPU or GPU; or, smart meters or power monitoring devices can be installed in the server rack or at the power input end to measure the power consumption of the entire computing node or cluster.

[0034] The hardware temperature index is obtained by comparing the difference between the current hardware temperature and the nominal operating temperature with the difference between the maximum allowable operating temperature and the nominal operating temperature. The ratio is then capped at 0 using the max function to obtain the lower limit of the ratio. This quantifies the degree to which the hardware temperature deviates from its ideal state and standardizes it into a dimensionless exponent for subsequent calculations. The use of the max function ensures that the exponent will not become negative due to the temperature falling below the nominal value, thus avoiding incorrect assessments of the hardware's health. The nominal operating temperature and the maximum allowable operating temperature can be determined from the hardware manufacturer's specifications or through experimental testing.

[0035] The hardware power consumption index is obtained by comparing the difference between the current hardware power consumption and the normal power consumption with the difference between the maximum safe power consumption and the normal power consumption, and then using the max function to limit the lower limit of the ratio to 0. This aims to quantify the degree to which hardware power consumption deviates from the normal state and standardize it into a dimensionless exponent. The use of the max function ensures that the exponent will not become negative due to power consumption falling below the normal value, thus avoiding incorrect assessment of the health status. Normal power consumption can be determined based on the average power consumption when the system is idle or under light load, while the maximum safe power consumption can be set based on the rated power of the power supply unit, the maximum thermal design power (TDP) recommended by the hardware manufacturer, or the system stability test results.

[0036] By taking the larger of the hardware temperature index and the hardware power consumption index and calculating their complement (i.e., subtracting the larger of the two indices from 1), a hardware health status coefficient reflecting the overall health of the hardware is obtained. This approach aims to comprehensively consider the impact of temperature and power consumption on hardware health, prioritizing the worse (i.e., larger index) scenario and converting it into a health status value between 0 and 1, where 1 represents perfect health and 0 represents extreme unhealthiness. This calculation method reflects the "weakest link" effect, where the health of a system is determined by its weakest link. By taking the larger value, it ensures that even if only one indicator is abnormal, it can promptly reflect potential hardware risks.

[0037] Based on the differences in computation time distribution among all parallel processes, the computational load balance degree, reflecting the degree of load balancing, is obtained by calculating the complement of the relative deviation between the maximum computation time and the average computation time (i.e., the ratio of the difference between the maximum computation time of all processes and the average computation time of all processes to the average computation time of all processes, and subtracting the ratio from 1). This aims to quantify the uniformity of the distribution and execution of parallel computing tasks among various processes or computing nodes. The closer the value is to 1, the more balanced the load; the closer it is to 0, the more unbalanced the load. This calculation method can be applied to various parallel computing models, such as Message Passing Interface (MPI) or OpenMP, where computation time can refer to the time required for each process to complete its local grid computation within a time step.

[0038] The system efficiency coefficient is obtained by multiplying the calculated load balancing degree by the hardware health status. , The value ranges from 0 to 1. Multiplication means that if any one of these metrics performs poorly (e.g., poor load balancing or poor hardware health), it will significantly reduce the overall system efficiency coefficient. This prompts the adjustment model to adopt a more conservative time step strategy; This reflects the combined impact of current parallel computing load and hardware status on computing performance. A higher value indicates higher system parallel efficiency and better hardware condition, enabling support for more aggressive (smaller) step size adjustment strategies.

[0039] As a specific implementation, a distributed computing cluster containing multiple GPU nodes can be considered for real-time forecasting of tsunami propagation and inundation. First, real-time temperature and power consumption data for each GPU can be obtained using the NVIDIA NVML library. For example, assuming GPU1 has a temperature of 70°C and a power consumption of 220W, and GPU2 has a temperature of 75°C and a power consumption of 240W, the timing function of the MPI library can be used to record the computation time of each MPI process within a time step. For example, assuming there are four processes with times of 100ms, 110ms, 95ms, and 105ms respectively, the hardware nominal operating temperature is set to 65°C, and the maximum allowable hardware operating temperature is set to 85°C. Then, the hardware temperature index for GPU1 is 0.25, and the hardware temperature index for GPU2 is 0.5. Setting the normal power consumption to 200W and the maximum safe power consumption to 280W, the hardware power consumption index for GPU1 is 0.25, and the hardware power consumption index for GPU2 is 0.5. Next, the hardware health status of each GPU is calculated. GPU1's health status is 0.75, and GPU2's is 0.5. The worst health status among all GPUs is taken as the overall hardware health status, which is 0.5. Simultaneously, the load balancing is calculated. The maximum computation time is 110ms, and the average computation time is 102.5ms, resulting in a load balancing score of 0.927. Finally, the load balancing score is multiplied by the hardware health status to obtain a system efficiency coefficient of 0.4635. This system efficiency coefficient will be used as input to the aforementioned adjustment model, affecting the final target numerical integration time step. In this case, due to the relatively low hardware health status, even with a high load balancing score, the system efficiency coefficient is relatively low. This will cause the adjustment model to adopt a more conservative strategy when calculating the target numerical integration time step, avoiding computational instability or efficiency degradation due to hardware overload.

[0040] Through the above technical solution, this application can accurately quantify and calculate load balancing and hardware health status, and integrate them into a system efficiency coefficient. This coefficient This provides crucial real-time system performance feedback for the aforementioned adjustment model, enabling the algorithm to dynamically adjust the numerical integration time step. Specifically, when the system operates efficiently and the hardware is in good condition, A higher value allows for a smaller step size in the model adjustment to improve accuracy; conversely, a lower value may result in uneven system load or hardware overload risk. When the value is low, the adjustment model adopts a more conservative step-size strategy to ensure the stability and reliability of the calculation. This dynamic and intelligent adjustment mechanism effectively solves the problem of balancing computational efficiency and stability in traditional methods, significantly improving the adaptability and robustness of tsunami forecasting algorithms in complex parallel computing environments. Thus, while ensuring forecast accuracy, it meets the stringent real-time requirements of tsunami warnings.

[0041] like Figure 1 As shown, in a preferred embodiment of the present invention, the operation flow of the fitness model is as follows: The system retrieves the remaining global calculation time until the forecast task's deadline. This can be queried in real time by the system scheduling module or dynamically calculated by monitoring the start time and preset deadline of the forecast task. Simultaneously, it retrieves the maximum allowed calculation time set by the system. This time can be set based on the tsunami warning response level, historical experience data, or user configuration. For example, it might be set to 30 minutes for a Level 1 warning, while for a drill task it might be set to several hours. The remaining global computation time is compared with the maximum allowed computation time set by the system. After normalization and upper limit constraints, the complement is calculated (that is, the ratio of the remaining global computation time to the maximum allowed computation time set by the system is processed, and after applying a min function to limit the upper limit to 1, the limit value is subtracted from 1). This yields the time urgency index, which reflects the degree of time pressure. .

[0042] Obtain the geographic distance from the current grid point to the tsunami epicenter. This can be achieved using the coordinate calculation function of a Geographic Information System (GIS), such as calculating spherical or Euclidean distance. Simultaneously, obtain the preset maximum impact distance, which can be set based on the propagation range of historical tsunami events, the effective simulation range of a numerical model, or expert experience; for example, it could be set to 2000 kilometers from the epicenter. The distance from the grid point to the seismic source is compared with the maximum influence distance. After normalization and upper limit constraints, the complement is calculated (that is, the ratio of the distance from the grid point to the seismic source to the maximum influence distance is processed, and after applying a min function to limit the amplitude to 1, the amplitude limit value is subtracted from 1). The target area priority weight reflecting the importance of the area is obtained. .

[0043] The physical constraint factor is obtained by multiplying the stability constraint coefficient by the complement of the dynamic accuracy coefficient; the business urgency factor is obtained by multiplying the time urgency index by the target area priority weight. The physical constraint factor is divided by the sum of the physical constraint factor and the business urgency factor to obtain the physical-business fit. When physical conditions allow and business urgency is low, the fit value increases, and a larger time step is tended to be used to improve computational efficiency. When business urgency is high or physical conditions are limited, the fit value decreases, and a smaller time step is tended to be used to ensure computational accuracy. The specific formula for calculating physical-business compatibility can be: ,in, For physical-business adaptability, Output range 0-1, A larger value indicates that a larger step size is needed for that region. For time urgency index, Priority weight for target region For stability constraint coefficients, For dynamic accuracy coefficients, It is a very small positive number, such as 0.000001.

[0044] Numerator of the formula This reflects the size of the time step allowed by the numerical simulation under the current physical field conditions: stability constraint coefficients. The larger the value (the more stable the physical field), the higher the dynamic accuracy coefficient. The smaller the value (the smoother the physical field, the lower the precision requirement), the larger the numerator, meaning a larger time step is physically permissible. In the denominator... The item represents the "inhibitory" effect of business on time step: the more urgent the time ( Larger) or higher regional priority ( When the value is larger, the product term becomes larger, leading to an increase in the denominator, which in turn affects the physical-business fit. Reduced. Through this mechanism, when the physical field is stable and the precision requirement is not high, and if time is ample and the region is not critical, then... A larger value indicates that a larger time step can be used to improve computational efficiency. Conversely, when the physics field is complex or unstable, time is extremely tight, or the current grid point is located in a critical region near the epicenter, even if a slightly larger step size is physically permissible, operational requirements may override this. The increase in the item "lowers" The value, thereby ensuring the final physical-business fit. The smaller time step size indicates that a smaller time step is needed to ensure timely forecasting and accuracy in key areas. This dynamic, adaptive calculation method allows subsequent adjustment models to adapt accordingly. The value, while satisfying physical stability and accuracy, can more intelligently adjust the time step of the target numerical integration, thereby optimizing the computational efficiency and accuracy of tsunami forecasting as a whole. Especially when dealing with real-time early warning tasks, it can better balance computing resources and forecast timeliness.

[0045] As a specific implementation, suppose that in a real-time tsunami forecasting mission, the system's maximum allowable calculation time is set to 60 minutes. At a specific calculation moment, the remaining global calculation time is 15 minutes. The calculation process for the time urgency index is as follows: First, the ratio of the remaining global calculation time of 15 minutes to the maximum allowable calculation time of 60 minutes is calculated, resulting in 0.25. Then, a minus function is applied with an upper limit of 1, resulting in another 0.25. Finally, 0.25 is subtracted from 1, yielding a time urgency index of 0.75. This indicates that the current time is extremely tight. Simultaneously, consider a specific grid point located in a near-shore area, with a distance of 500 kilometers from the epicenter. The system's preset maximum influence distance is 2000 kilometers. The calculation process for the target area priority weight is as follows: First, the ratio of the grid point's distance from the epicenter (500 kilometers) to the maximum influence distance (2000 kilometers) is calculated, resulting in 0.25. Then, a minus function is applied with an upper limit of 1, resulting in another 0.25. Finally, subtracting 0.25 from 1 yields a target area priority weight of 0.75. This indicates that the grid point is located in a critical region and has a high priority. Assuming that at the same time, the stability constraint coefficient calculated using the stability constraint model is 0.8, and the dynamic accuracy coefficient calculated using the dynamic accuracy model is 0.6, substituting these values ​​along with the smallest positive number into the fitness formula yields a physical-operational fitness of approximately 0.3626. Since this value is relatively small, it suggests that although the physical field may allow for a certain time step, due to time constraints and the high priority of this region, the system tends to use a smaller time step to ensure that the forecast is completed within a limited time and the details of the critical region are accurately captured.

[0046] Through the above technical solution, this application effectively addresses the problem that traditional tsunami forecasting algorithms fail to fully consider the real-time constraints of forecasting tasks and the priority of key areas when calculating physical-operational adaptability. By introducing a time urgency index and target area priority weights, and comprehensively calculating them along with the stability constraint coefficient and dynamic accuracy coefficient of the physical field, the physical-operational adaptability can dynamically and comprehensively reflect the current computational status and operational needs. This allows subsequent time step adjustments to no longer rely solely on the local characteristics of the physical field, but to be intelligently adjusted according to the overall timeliness requirements of the forecasting task and the importance of different areas. For example, when forecasting time is tight or in key areas (such as nearshore inundation zones), even if the physical field is relatively stable, the system will tend to use a smaller time step to ensure the timeliness and accuracy of the forecast; while in areas with ample time or non-critical areas, the time step can be appropriately increased to improve computational efficiency. This adaptive adjustment mechanism significantly improves the targeting and overall efficiency of real-time tsunami propagation and inundation forecasting algorithms, ensuring that computational resources are maximized while prioritizing forecast accuracy for key areas, all while meeting the requirements for early warning timeliness.

[0047] like Figure 1 As shown, in a preferred embodiment of the present invention, the operation flow of the stability constraint model is as follows: The local CFL number and local flow Froude number can be obtained by real-time monitoring of fluid velocity, water depth, mesh size, and the current time step, combined with the internal state of the numerical simulator. The local CFL number is a dimensionless parameter that measures the stability of the numerical scheme, indicating the stability condition of the numerical solution at the current time step. The local flow Froude number is a dimensionless parameter that measures the ratio of inertial force to gravity, characterizing the flow state of the fluid, such as subcritical or supercritical flow.

[0048] The local CFL number exponent is obtained by comparing the ratio of the local CFL number to the upper limit of stability of the Courant number allowed by the numerical format. The upper limit of stability of the CFL number allowed by the numerical scheme is a theoretical or empirical value determined by the numerical discretization method and specific scheme used. It represents the maximum value that the CFL number can reach while ensuring the stability of the numerical solution.

[0049] When the local flow Froude number does not exceed the critical Froude number, the local flow Froude number exponent is obtained by normalization calculation; when the local flow Froude number exceeds the critical Froude number, the local flow Froude number exponent is calculated by a linear penalty function, and the penalty coefficient is proportional to the extent of exceeding the critical value. The formula for calculating the Froude number exponent in local flow can be: ,in, For local flow, the Froude number exponent is used. For local flow Froude number, The critical Froude number is the key threshold for distinguishing between subcritical and supercritical flows, and is usually set to 1. The penalty coefficient is a positive parameter used to... Exceed At that time, Further scaling is performed to enhance the model's sensitivity to supercritical flows; this coefficient can be adjusted empirically or through numerical experiments.

[0050] The inverses of the local CFL number exponent and the local flow Froude number exponent are calculated separately, and the smaller of the two is taken as the stability constraint coefficient. The calculation logic of this stability constraint coefficient ensures that both the CFL stability condition and the flow state stability condition are satisfied at the same time. The higher the stability constraint coefficient, the larger the time step scaling ratio is allowed for the current physical field state. The formula for calculating the stability constraint coefficient is as follows: ,in, For stability constraint coefficients, Output range , For the local CFL number index, For local flow, the Froude number exponent is used. It is a very small positive number, such as 0.000001. The core idea is that whichever is larger (i.e. closer to an unstable state), either the local CFL number exponent or the local flow Froude number exponent, will become the main factor limiting the time step.

[0051] The following is a specific example to illustrate this. The allowed numerical format can be set to a stability upper limit of 0.8 for the Courant number, a critical Froude number of 1.0, a penalty coefficient of 0.5, and a minimum positive number of 0.000001.

[0052] Suppose that at a certain grid point, the current local CFL number is 0.64 and the local flow Froude number is 0.8.

[0053] First, the local CFL number exponent is calculated to be 0.8.

[0054] Next, the Froude number exponent for the local flow is calculated. Because... It is 0.8. Since it is 1.0, the first part of the formula is used: Calculate It is 0.8.

[0055] Finally, based on Values ​​0.8 and The value of 0.8 was imported into the formula, and the stability constraint coefficient was calculated to be 1.25.

[0056] because Output range Therefore, in this case, the stability constraint coefficient It will be limited to 1.

[0057] To give another example, suppose at another grid point, the local CFL number is 0.72 and the local flow Froude number is 1.2.

[0058] First, the local CFL number exponent is calculated to be 0.9.

[0059] Next, the Froude number exponent for the local flow is calculated. Because... Greater than Therefore, the second part of the formula is used: Calculate It is 1.1.

[0060] Finally, Value 1.1 and The value of 0.9 was imported into the formula, and the stability constraint coefficient was calculated to be 0.909.

[0061] In this case, the stability constraint coefficient is approximately 0.909, indicating that due to the presence of supercritical flow, the time step needs to be reduced to a certain extent.

[0062] Through the above technical solutions, the stability constraint model of this application can more accurately assess the numerical stability requirements of the current physical field state. Specifically, by introducing a penalty mechanism for the local flow Froude number in the supercritical state, the model can effectively identify and respond to complex flow phenomena that may occur during tsunami propagation and inundation, such as shock waves or high-speed water flow near the wet-dry boundary. This solves the problem that traditional methods may lead to coefficient issues when dealing with supercritical flows. This avoids the problem of inaccurate estimation, thus preventing computational instability caused by excessively large time steps or sacrificing computational efficiency due to overly conservative time steps. When this stability constraint coefficient... When input into the aforementioned fit model, it can work in conjunction with other factors (such as time urgency index and target area priority weight) to improve the final output physical-business fit. This more accurately reflects the actual time step requirements of the current region. This precise stability assessment allows the overall algorithm to adjust the time step more flexibly while ensuring computational stability. In critical areas (such as steep wavefronts and inundation fronts), smaller step sizes are used to capture high-precision details, while larger step sizes are used in stable areas to improve computational efficiency. Ultimately, this significantly improves the accuracy and computational efficiency of real-time tsunami propagation and inundation forecasts.

[0063] like Figure 1 As shown, in a preferred embodiment of the present invention, the operation flow of the dynamic accuracy model is as follows: The spatial gradient norm of wave height and the moving velocity of the wet-dry boundary are obtained. The spatial gradient norm of wave height can be obtained by performing spatial difference calculation on the wave height data at simulated grid points and then calculating its second norm; or by using image processing techniques to perform edge detection and gradient analysis on the wave height field. The acquisition method can be by tracking the grid points in the simulated grid where the wet-dry state changes and calculating their displacement per unit time; or by analyzing the position changes of grid points with zero water depth in adjacent time steps.

[0064] The wave height spatial gradient norm (the L2 norm of the wave height spatial gradient, characterizing the steepness of the wave front) and the wet-dry boundary movement velocity (the movement velocity of the wet-dry boundary (the inundation front)) are compared with the historical maximum wave height gradient and the theoretical maximum inundation velocity, respectively, to obtain the wave height spatial gradient norm index and the wet-dry boundary movement velocity index. The historical maximum wave height gradient refers to the maximum wave height spatial gradient value recorded or estimated in past tsunami event simulation data, observation data, or theoretical analysis. Its function is to provide a normalized reference benchmark for the currently calculated wave height spatial gradient norm, enabling effective comparison of wave height steepness under different simulation scenarios. It can be obtained by extracting the maximum value from the numerical simulation database of historical tsunami events; or by pre-calculating extreme tsunami scenarios under specific terrain and focal conditions. The theoretical maximum inundation velocity refers to the maximum inundation velocity that water may reach on land under given terrain slope and gravitational acceleration conditions. Its function is to provide a physically reasonable normalized upper limit for the currently calculated wet-dry boundary movement velocity, ensuring the physical meaning of the inundation velocity index. The data can be obtained through estimation based on shallow water wave theory or empirical formulas, combined with local topographic data (such as slope) and gravitational acceleration; or through high-resolution numerical simulations of extreme inundation scenarios, extracting the maximum velocity. Ratio processing transforms the original physical quantities into dimensionless exponents, achieving normalization. This normalization allows for comparison and weighting of different physical quantities on a uniform scale, avoiding the influence of dimensional differences on subsequent calculations, and enabling the exponent to reflect the "tension" of the current state relative to the extreme or maximum probable state. A higher exponent value indicates a more intense physical process in the region, requiring higher computational accuracy.

[0065] The dynamic accuracy coefficient is obtained by weighted linear combination of the wave height spatial gradient norm exponent and the dry-wet boundary moving velocity exponent. The higher the dynamic accuracy coefficient, the higher the complexity of the current flow field and the stronger the demand for high time resolution. The weighted linear combination can be calculated as follows: Where is the dynamic accuracy coefficient. , Output range 0-1, This reflects the complexity of the current flow field and the need for high temporal resolution. The spatial gradient norm exponent of wave height. The velocity index of the wet-dry boundary. The spatial gradient weight coefficients have values ​​ranging from 0 to 1. The parameters used to balance the contributions of the wave height spatial gradient norm index and the dry / wet boundary moving velocity index in calculating the dynamic accuracy coefficient. Based on actual needs or expert experience, it can also be calculated through AI, or by introducing reinforcement learning or adaptive weighting mechanisms.

[0066] As a specific implementation method, suppose that in a real-time forecast simulation of tsunami propagation and inundation, the flow field state of a certain nearshore area needs to be evaluated. First, the system calculates the wave height spatial gradient norm of the area in real time. For example, by performing finite difference on the wave height data at the grid points of the area, the wave height gradient in the x and y directions is obtained, and then its second norm is calculated. Simultaneously, the system tracks the wet and dry boundaries of the area and calculates their movement distance within the current time step, thus obtaining the wet and dry boundary movement velocity. Suppose that through historical data analysis, we have determined the historical maximum wave height gradient of this sea area to be 0.5 m / m, and estimated the theoretical maximum inundation velocity to be 10 m / s based on the local topographic slope and gravitational acceleration. If at the current moment, the wave height spatial gradient norm of the nearshore area is 0.25 m / m, and the wet and dry boundary movement velocity is 5 m / s, then the wave height spatial gradient norm exponent will be calculated as 0.5. The wet and dry boundary movement velocity exponent will also be calculated as 0.5. Next, these two exponents are imported into the formula for calculation. Assuming the weighting coefficient is set to 0.6 based on experience or a pre-defined strategy, the dynamic precision coefficient will be calculated as 0.5. This calculated dynamic precision coefficient will be used as input to the adaptation model to further calculate the physical-business adaptation, and ultimately affect the adjustment of the target numerical integration time step. For example, if A higher value indicates that the flow field in the region is complex and requires high accuracy. Therefore, in subsequent step size adjustments, the system will tend to use a smaller time step to ensure the accuracy of the simulation.

[0067] Through the above technical solutions, this application can accurately quantify the complexity of the flow field, ensuring that the dynamic accuracy coefficient can precisely reflect the requirement for high temporal resolution. Specifically, by comparing the wave height spatial gradient norm with the historical maximum wave height gradient, the steepness of the wavefront is normalized, allowing for fair comparison of waveform changes in different regions, thus more reliably capturing the accuracy requirements of wavefront changes. Simultaneously, by comparing the dry-wet boundary movement velocity with the theoretical maximum inundation velocity, and estimating the maximum possible inundation velocity based on terrain slope and gravitational acceleration, the scientific rationality of the benchmark is ensured, enabling the index to objectively reflect the dynamic movement of the inundation front and helping to identify the high-resolution requirements of key areas. The methods for obtaining the maximum wave height gradient and the maximum possible inundation velocity provide a reliable reference benchmark, avoiding errors caused by subjective settings and enhancing the universality and accuracy of the index. By importing these two indices into a weighted sum formula to calculate the dynamic accuracy coefficient, and combining weighting coefficients to adjust the contributions of different factors, the influence of wave steepness and inundation velocity is balanced, demonstrating the flexibility of dynamic adjustment and enabling the dynamic accuracy coefficient to comprehensively reflect the complexity of the flow field. The dynamic accuracy coefficient output ranges from 0 to 1, achieving standardized output and facilitating seamless integration with subsequent physics-business adaptability calculations. Ultimately, this dynamic accuracy coefficient directly serves time step optimization decisions, enabling the system to automatically select smaller time steps in regions with complex flow fields and high accuracy requirements. This improves computational stability and accuracy, effectively capturing key physical details and resolving the contradiction between computational efficiency and accuracy in traditional methods, ensuring that simulation accuracy in key areas is guaranteed while meeting forecast timeliness requirements.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI acceleration-based real-time prediction algorithm for tsunami propagation and inundation, characterized in that, The method comprises the following steps: obtaining a stable constraint coefficient through a stable constraint model based on a local CFL number and a local flow Froude number; obtaining a dynamic precision coefficient through a dynamic precision model based on a wave height spatial gradient norm and a dry-wet boundary moving speed; obtaining a physical-service adaptability through an adaptability model based on a grid point to a source distance under the stable constraint coefficient and the dynamic precision coefficient and a global calculation residual time; obtaining a system efficiency coefficient through a system efficiency model based on a calculation load balancing degree, a hardware temperature and a hardware power consumption; outputting a target numerical integral time step through an adjustment model according to the physical-service adaptability and the system efficiency coefficient.

2. The AI acceleration based tsunami propagation and inundation real-time forecasting algorithm according to claim 1, characterized in that, The operation process of the adjustment model is as follows: calculating the target numerical integral time step through a weighted comprehensive function based on the physical-service adaptability and the system efficiency coefficient and combining minimum and maximum time steps; wherein the physical-service adaptability participates in the calculation in a positive weighted manner, the system efficiency coefficient participates in the calculation in a reverse weighted manner, and the weight distribution between the physical-service adaptability and the system efficiency coefficient is realized through an adjustable weight coefficient.

3. The AI acceleration based tsunami propagation and inundation real-time forecasting algorithm according to claim 2, characterized in that, The operation process of the system efficiency model is as follows: obtaining a hardware health state based on the hardware temperature and the hardware power consumption; obtaining a calculation load balancing degree reflecting the load balancing degree through a calculation of a relative deviation complement of a maximum calculation time and an average calculation time based on the calculation time distribution difference of all parallel processes; obtaining the system efficiency coefficient through a product operation of the calculation load balancing degree and the hardware health state.

4. The AI acceleration based tsunami propagation and inundation real-time forecasting algorithm according to claim 2, characterized in that, The hardware health state obtaining method is as follows: obtaining a hardware temperature index through a ratio processing of a difference between a current hardware temperature and a hardware nominal working temperature and a difference between a maximum allowed hardware working temperature and the hardware nominal working temperature, limiting the ratio lower limit to 0 through a max function and then limiting the amplitude; obtaining a hardware power consumption index through a ratio processing of a difference between a current hardware power consumption and a normal power consumption and a difference between a maximum safe power consumption and the normal power consumption, limiting the ratio lower limit to 0 through a max function and then limiting the amplitude; obtaining a hardware health state coefficient reflecting the overall hardware health state through taking the larger value between the hardware temperature index and the hardware power consumption index and calculating the complement thereof.

5. The AI acceleration based tsunami propagation and inundation real-time forecasting algorithm according to claim 4, characterized in that, The operation process of the adaptability model is as follows: obtaining a time urgency index reflecting the time urgency degree through a complement calculation of a normalized processing and an upper limit constraint of a comparison between a global calculation residual time and a maximum allowed calculation time set by the system; obtaining a target region priority weight reflecting the region importance degree through a complement calculation of a normalized processing and an upper limit constraint of a comparison between a grid point to a source distance and a maximum influence distance; obtaining a physical constraint factor through a multiplication of the complement of the stable constraint coefficient and the dynamic precision coefficient; obtaining a business urgency factor through a multiplication of the time urgency index and the target region priority weight; The physical constraint factor is divided by the sum of the physical constraint factor and the business urgency factor to obtain a physical-business adaptation degree; when the physical condition allows and the business urgency is low, the adaptation degree value increases, and a larger time step is inclined to be adopted to improve the calculation efficiency; when the business urgency is high or the physical condition is limited, the adaptation degree value decreases, and a smaller time step is inclined to be adopted to ensure the calculation accuracy.

6. The AI acceleration based tsunami propagation and inundation real-time forecasting algorithm according to claim 4, wherein, The operation flow of the stable constraint model is: The local CFL number is divided by the stable upper limit of the numerical format allowed by the Courant number to obtain a local CFL number index; When the local flow Froude number does not exceed the critical Froude number, the local flow Froude number index is obtained through normalization calculation; when the local flow Froude number exceeds the critical Froude number, the local flow Froude number index is calculated through a linear penalty function, and the penalty coefficient is proportional to the amplitude of exceeding the critical value; The reciprocal of the local CFL number index and the local flow Froude number index is calculated respectively, and the smaller one of the two is taken as the stable constraint coefficient. The calculation logic of the stable constraint coefficient ensures that the CFL stability condition and the flow state stability condition are satisfied at the same time. The higher the stable constraint coefficient, the higher the current physical field state allows the larger time step scaling ratio.

7. The AI acceleration based tsunami propagation and inundation real-time forecasting algorithm according to claim 4, wherein, The operation flow of the dynamic precision model is: The wave height spatial gradient norm and the dry-wet boundary moving speed are respectively divided by the historical maximum wave height gradient and the theoretical maximum inundation speed to obtain a wave height spatial gradient norm index and a dry-wet boundary moving speed index; The wave height spatial gradient norm index and the dry-wet boundary moving speed index are linearly combined to obtain a dynamic precision coefficient. The higher the dynamic precision coefficient, the higher the current flow field complexity, and the stronger the demand for high time resolution.

Images