A gpu-accelerated discrete element simulation method for ice rubble accumulation evolution

By dividing tasks between a multi-core CPU and a high-concurrency GPU, and employing adaptive time steps and cross-domain boundary synchronization methods, the problems of unbalanced computing power distribution and computational stability in the simulation of interaction between marine platforms and sea ice were solved, achieving efficient simulation of ice debris accumulation evolution.

CN123116750APending Publication Date: 2026-11-17HUANENG CLEAN ENERGY RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611249368.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing coupled discrete element and finite element simulation methods suffer from problems such as unbalanced computational power distribution, delayed data interaction, computational overload and numerical integration divergence caused by fixed time steps, and computational crashes due to the lack of parameter decision boundaries in the simulation of interaction between marine platforms and sea ice.

Method used

A GPU-accelerated approach is adopted to divide the computational tasks between a multi-core central processing unit and a high-concurrency graphics processing unit. By adjusting the adaptive time step and synchronizing data across the domain boundary nodes, combined with the Rayleigh time step calculation model and the boundary comparison of the safety margin coefficient, efficient coupling computation between the marine platform structure model and the sea ice particle array is achieved.

Benefits of technology

It solves the problems of unbalanced computing power distribution and computational overload, ensures the stability of numerical integration, prevents computational crashes, and generates accurate three-dimensional ice fragmentation evolution output.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116750A_ABST
    Figure CN123116750A_ABST
Patent Text Reader

Abstract

This invention relates to the field of numerical simulation technology in marine engineering, and discloses a GPU-accelerated discrete element method for simulating the evolution of ice fragmentation. The method includes: deploying a marine platform structural model on a multi-core CPU to generate a structural node array; deploying spherical discrete elements on a high-concurrency GPU to generate a sea ice particle array; executing a cohesion and breakup algorithm on the high-concurrency GPU to output local contact forces; calculating the global ice load and solving the motion equations using the multi-core CPU; bidirectionally importing real-time structural displacements, real-time structural velocities, and local contact forces within the execution boundary node range to update the global ice load; and calling a time-step adaptive algorithm to generate three-dimensional evolution output results. This invention achieves cross-domain data interaction by defining the execution boundary node range, adjusting the time step using an adaptive algorithm, and setting a decision boundary triggering anomaly rollback mechanism to avoid computational overload and integral divergence, thus maintaining computational stability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of numerical simulation technology for marine engineering, specifically to a discrete element simulation method for ice fragmentation evolution based on GPU acceleration. Background Technology

[0002] Polar offshore platforms are subjected to impacts and compression from sea ice during operation. When sea ice comes into contact with the platform, it breaks apart, leading to the accumulation and evolution of ice fragments. Coupled calculations using the discrete element method (DEM) and the finite element method (FEM) are commonly used to simulate the interaction between sea ice particles and the offshore platform.

[0003] Existing coupled discrete element and finite element method (EMI) simulations typically rely on a single hardware processing unit to complete all data processing operations, failing to differentiate between the data characteristics of the structural computational domain and the discrete particle computational domain. Discrete particle computation involves determining the collision and contact of numerous sea ice particles, while finite element computation involves solving the structural dynamic response equations. A single hardware processing unit faces the problem of unbalanced computational power distribution. Conventional data interaction mechanisms traverse all nodes and sea ice particles globally to find contact points. As the number of sea ice particles increases, this global traversal computation consumes a significant amount of memory from the hardware processing unit, causing data interaction channel congestion and computational overload.

[0004] In the time integration phase of discrete element simulations, conventional simulation programs use a fixed time step parameter. When sea ice particles collide and break apart, the local stiffness and particle velocity change rapidly. Setting a large time step parameter can cause the numerical integration to fail to converge in regions of rapid stiffness change, resulting in numerical integration divergence; setting a small time step parameter will increase the overall computation cycle and consume excessive hardware computing resources. A fixed time step cannot adapt to the changing motion states during sea ice evolution.

[0005] Coupled discrete element and finite element simulations are nonlinear computational processes. When large-scale compression and accumulation of ice fragments occur, the contact force parameters received by structural nodes undergo abrupt changes, producing abnormal values ​​that exceed conventional extremes. Conventional simulation programs lack decision boundaries based on constraint extreme values ​​and lack intervention callback mechanisms. When the overall environmental ice load parameters exceed reasonable limits, conventional simulation programs do not perform data verification and truncation operations, causing unverified erroneous values ​​to continue participating in subsequent time-iteration calculations. The computational errors accumulate over time, leading to the collapse of the numerical computation process and ultimately distorting the output three-dimensional evolutionary distribution data. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a GPU-accelerated discrete element method for simulating the evolution of ice accumulation, which solves the problems of unbalanced computing power distribution and delayed cross-domain data interaction when a single computing platform performs coupled operations in the discrete element domain and finite element domain.

[0007] To achieve the above objectives, this invention provides a GPU-accelerated discrete element method for simulating the evolution of ice debris accumulation, comprising the following steps: Deploy the marine platform structure model in the memory of a multi-core central processing unit to generate a structure node array, and deploy discrete spherical units in the video memory of a high-concurrency graphics processor to generate a sea ice particle array. A high-concurrency graphics processor is used to perform a bonding and breaking algorithm on discrete elements of a sphere. When the critical deformation condition is met, ice fragments are formed, and the local contact force is output and recorded into the sea ice particle array. A multi-core central processing unit is used to read local contact forces to calculate global ice load, solve motion equations to obtain real-time structural displacement and real-time structural velocity, and input them into the structural node array. By comparing the spatial overlap areas of the structural node array and the sea ice particle array, the execution boundary node range is delineated. Within the execution boundary node range, the real-time structural displacement and real-time structural velocity are imported into the sea ice particle array, and the local contact force is imported back into the structural node array to update the global ice load. Based on the updated sea ice particle array and structural node array, the time step adaptive algorithm is invoked to perform time integration, generating a three-dimensional evolution output of the ice debris accumulation evolution in front of the marine platform structural model.

[0008] The underlying binding operations during the data initialization phase are as follows: The initial spatial node coordinates of finite element beam elements, flat shell elements, and solid elements are written into the structural node array using a multi-core central processing unit, and the initial equivalent structural mass, equivalent structural damping ratio, and stiffness matrix parameters are set. A high-concurrency graphics processor is used to control the connection between the discrete spherical elements through elastic bonds to form a bonded particle model. Each discrete spherical element is given a first degree of freedom, a second degree of freedom, and a third degree of freedom representing the spatial displacement in the first, second, and third dimensions. The initial spatial coordinates and motion velocity components are synchronously recorded into the sea ice particle array to establish the underlying data binding relationship.

[0009] The specific details of the bonding and fragmentation processes simulating the evolution characteristics of sea ice are as follows: The deformation of the elastic bond, determined by the difference between the second and first degrees of freedom, is continuously monitored and compared with the predetermined critical deformation conditions. When the critical deformation conditions are met, the elastic bond between adjacent spherical discrete units fails and breaks, resulting in weakened contact stiffness and separation to form ice fragments. Based on the uniform distribution function, the spherical discrete unit corresponding to the ice fragments is repositioned in the generation area of ​​a preset multiple of ice thickness at a distance from the leading edge of the marine platform structure model and written into the sea ice particle array. The preset multiple of ice thickness is set to 3 to 5 times the average thickness of the sea ice.

[0010] The specific calculations for extracting the global ice load numerical values ​​are as follows: The dynamic response is set as a multi-degree-of-freedom or single-degree-of-freedom oscillator model. The Heaviside step function is called inside the structural node array to determine the contact state and output the difference calculation results. Based on the difference calculation results, the spherical discrete elements that have contact are selected, and the local contact force output by the spherical discrete elements is included in the integral accumulation calculation range to obtain the global ice load.

[0011] The specific computational process for obtaining the dynamic response within the structural domain includes: The global ice load, equivalent structural mass, equivalent structural damping ratio, and natural frequency are extracted from the structural node array. The numerical correlation between real-time structural displacement and global ice load is quantified by calling the first modal dynamic equation of the structure. The dynamic response results, including real-time structural displacement and real-time structural velocity and acceleration generated under the global ice load, are obtained and entered into the structural node array.

[0012] The specific execution of cross-domain boundary node spatial overlap determination is as follows: The system performs boundary node spatial overlap determination between the discrete element domain and the finite element domain; extracts the three-dimensional spatial coordinates of the spherical discrete element and ice fragments from the sea ice particle array, and extracts the spatial node coordinates of the finite element beam element and flat shell element from the structural node array; calculates the spatial distance and performs threshold screening calculation, delineates the three-dimensional spatial range with spatial distance less than the preset collision tolerance as the spatial overlap region, dynamically locks the execution boundary node range, and sets the preset collision tolerance to 0.05 to 0.1 times the average radius of the spherical discrete element.

[0013] The specific process of bidirectional data synchronization between underlying hardware includes: Real-time structural displacement and velocity are extracted from the structural node array, transmitted to the video memory of the high-concurrency graphics processor, and imported into the sea ice particle array to update the input conditions of the contact algorithm. Local contact forces are extracted from the sea ice particle array, transmitted in reverse to the memory of the multi-core central processing unit, and imported into the structural node array to update the global ice load, thus completing bidirectional data synchronization.

[0014] The adaptive time step adjustment to ensure numerical integration stability is as follows: A Rayleigh time step calculation model is introduced to obtain the baseline critical step size. If the current shear modulus is lower than the preset safety lower limit, the baseline critical step size is forcibly reset to the minimum integration period. The preset safety lower limit is set to 0.01 GPa to 0.1 GPa. When the relative motion rate of the spherical discrete element and the marine platform structure model is higher than the preset rate threshold and the number of fracture events is higher than the preset fracture threshold, the time step size is reduced. The preset rate threshold is set to 1.5 to 2.0 times the initially set sea ice drift speed, and the preset fracture threshold is set to 5% to 10% of the total number of elastic bonds in the current discrete element domain within a single time step. When the relative motion rate is lower than the preset rate threshold and the number of fracture events is lower than the preset fracture threshold, the time step size is increased, and the time step size is generated to perform time integration calculation.

[0015] The reconstruction of a 3D evolution scene under multiple working conditions specifically includes: When the marine platform structure model is set as an upright structure, continuous compression failure distribution data and intermittent fragmentation distribution data are generated. When the marine platform structure model is set as a conical structure, bending failure distribution data, stable ice pile clearing distribution data, and accumulation distribution data are generated. The continuous compression failure distribution data, intermittent fragmentation distribution data, bending failure data, stable ice pile clearing distribution data, and accumulation distribution data are merged into a three-dimensional evolution output result.

[0016] The specific rollback verification to prevent distortion in numerical simulations includes: The static peak value of the global ice load is extracted when the marine platform structural model is set as a cone structure. The extreme static ice load of the two-dimensional cone is obtained by referring to the auxiliary calibration formula of the extreme static ice load of the two-dimensional cone. The numerical comparison operation of the global ice load and the extreme static ice load of the two-dimensional cone is performed. The extreme static ice load of the two-dimensional cone is multiplied by the set safety margin coefficient as the judgment boundary. The set safety margin coefficient is set to a value of 1.1 to 1.5. When the global ice load exceeds the judgment boundary, it is determined that the current simulation parameters are distorted. The abnormal rollback mechanism is triggered to restore the simulation state to the previous time node that meets the benchmark. The damping parameter is adaptively increased and the evolution calculation is re-executed.

[0017] This invention provides a GPU-accelerated discrete element method for simulating the evolution of ice debris accumulation. It offers the following advantages: 1. This invention deploys a structural node array in the memory of a multi-core central processing unit and a sea ice particle array in the video memory of a high-concurrency graphics processor. It delineates the execution boundary node range in the overlapping area of ​​the computation space and imports real-time structural displacement, real-time structural velocity and local contact force bidirectionally within the execution boundary node range. This avoids computational overload caused by global contact determination and prevents imbalance in the computing power distribution of a single hardware platform.

[0018] 2. This invention obtains the baseline critical step size by introducing the Rayleigh time step calculation model, and adjusts the time step size value by combining the relative motion rate of the spherical discrete element and the marine platform structural model and the change in the number of rupture events. The time step size value is reduced or expanded according to the preset threshold conditions, which avoids the numerical integration divergence caused by the fixed time step size in discrete element simulation and maintains the numerical stability of the time integration operation.

[0019] 3. This invention compares the global ice load with the extreme static ice load of a two-dimensional cone with a set safety margin coefficient. When the global ice load exceeds the judgment boundary, the simulation parameters are judged to be distorted and an abnormal rollback mechanism is triggered. The simulation state is restored to the time node that meets the benchmark and the damping parameter is increased and recalculated. This solves the problem of numerical calculation collapse caused by parameter mutation and ensures that the three-dimensional evolution output results conform to objective laws. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the heterogeneous computing platform and underlying data interaction architecture of the present invention; Figure 3 This is a schematic diagram illustrating the cross-domain model construction and initialization of the present invention; Figure 4 This is a comparison chart of the response of global ice load over time during the simulated full-cycle period of the present invention. Figure 5 This diagram illustrates the adaptive scaling control of the underlying time integration mechanism of this invention when the relative motion rate changes drastically. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 and attached Figure 2 This invention provides a discrete element simulation method for ice fragmentation evolution based on GPU acceleration. In this embodiment, the method relies on a heterogeneous computing platform to perform multi-threaded parallel computation.

[0023] The heterogeneous computing platform comprises a multi-core CPU, a high-concurrency graphics processing unit (GPU), and a memory bus. The multi-core CPU and the high-concurrency GPU establish a hardware connection and exchange low-level data via the memory bus. Internally, it deploys coupled computations using the Discrete Element Method (DEM) and the Finite Element Method (FEM). The DEM is executed independently within its own domain, as is the FEM within its own domain. Instruction control and finite element matrix solving tasks are deployed within the multi-core CPU. Within the finite element domain established in memory, finite element beam elements, flat shell elements, and solid elements are used to build structural models of the marine platform, encompassing both upright and conical structures. The structural dynamic responses caused by sea ice are calculated uniformly, and the results are output.

[0024] The parallel contact retrieval task for multiple sea ice particles is assigned to a high-concurrency graphics processing unit (GPU). The discrete element domain (DEV) is built in the GPU's memory. A multi-threaded parallel computing architecture is used within the DEV to determine the contact between multiple sea ice particles. The multiple sea ice particles are modeled in 3D using spherical discrete elements exhibiting both bonding and breakage effects. The bonding and breakage contact algorithm between these spherical discrete elements relies on the GPU's computational cores for concurrent computation.

[0025] The memory of the multi-core CPU and the video memory of the high-concurrency graphics processor are connected via a memory bus, enabling bidirectional interaction for cross-domain boundary search and parameter transfer tasks. The cross-domain boundary search task is limited to the boundary nodes between the discrete element domain and the finite element domain. The parameter transfer includes kinematic and dynamic parameters generated by the contact between the spherical discrete element and the marine platform structural model element. The transmission of kinematic and dynamic parameters dynamically adjusts the synchronization frequency according to a time-step adaptive algorithm, generating a three-dimensional simulation model of the entire evolution of sea ice accumulation in front of the upright and conical structures.

[0026] The data communication between the memory of the multi-core CPU and the video memory of the high-concurrency graphics processor is achieved through a bidirectional data communication link. The video memory of the high-concurrency graphics processor is divided into a first contiguous storage block to form a sea ice particle array. This sea ice particle array carries the basic data of multiple sea ice particles within the discrete element domain. The basic data of these multiple sea ice particles includes the three-dimensional spatial coordinates, velocity components, and bonding / fracture state labels of multiple spherical discrete units (hereinafter collectively referred to as spherical discrete units) exhibiting bonding and fracturing effects. The position changes and state transitions of the spherical discrete units within the discrete element domain are recorded into the sea ice particle array in real time.

[0027] A second independent storage block is partitioned within the memory of the multi-core central processing unit to form a structural node array. This structural node array carries the fundamental data of the offshore platform structural model within the finite element domain. The fundamental data of the offshore platform structural model includes the spatial node coordinates, displacement vectors, and finite element mass stiffness matrices of finite element beam elements, flat shell elements, and solid elements. The structural dynamic response results of the offshore platform structural model under sea ice influence within the finite element domain are recorded in real-time into the structural node array.

[0028] The sea ice particle array and the structure node array establish a low-level data binding relationship via a memory bus. The cross-domain boundary search execution range is limited to the region where the data coordinates of the two arrays overlap. The dynamic parameters output from the discrete element domain are extracted from the sea ice particle array and transmitted into the structure node array; the kinematic parameters output from the finite element domain are extracted from the structure node array and transmitted in reverse into the sea ice particle array. The parameters in both arrays are synchronously refreshed according to a time-step adaptive algorithm, ensuring that the three-dimensional simulation data of the entire evolution of sea ice accumulation in front of the upright and conical structures are in a time-aligned state.

[0029] See attached document Figure 1 A GPU-accelerated discrete element method for simulating the evolution of ice fragmentation includes: S1, the heterogeneous computing platform performs cross-domain model construction and initialization when establishing the underlying data binding relationship between the sea ice particle array and the structural node array, respectively establishing a sea ice discrete model and an offshore platform structural model. The multi-core CPU deploys the offshore platform structural model in memory, writes the initial spatial node coordinates of finite element beam elements, flat shell elements, and solid elements into the structural node array, and sets the initial equivalent structural mass, equivalent structural damping ratio, and stiffness matrix parameters for computation. The high-concurrency graphics processor deploys the sea ice discrete model in video memory and writes multiple spherical discrete elements into the sea ice particle array.

[0030] A high-concurrency graphics processor controls discrete spherical units to form a bonded particle model through elastic bonds, determines the initial position relative to the marine platform structure model, and assigns a first, second, and third degree of freedom to each discrete spherical unit to represent spatial displacement.

[0031] The initial spatial coordinates and velocity components of the discrete spherical elements are synchronously entered into the sea ice particle array. The sea ice particle array and the structural node array are in time-aligned and provide underlying data support.

[0032] S2, a high-concurrency graphics processor, mobilizes multiple thread blocks to traverse the sea ice particle array, simultaneously performing bonding and breaking algorithm calculations on the discrete spherical units within the sea ice particle array; determining the compressive and tensile contact states of the discrete spherical units. Under compressive conditions, adjacent discrete spherical units follow a hard contact algorithm to limit spatial overlap and penetration; under tensile conditions, loads are transferred via elastic bonds.

[0033] The computational core of the high-concurrency graphics processor solves the motion equations for each discrete element of the sphere, quantifying the motion rate. It continuously monitors the elastic bond deformation determined by the difference between the second and first degrees of freedom, comparing it to a pre-determined critical deformation condition. When the critical deformation condition is met, the elastic bonds between adjacent elements fail and break, leading to a weakening of contact stiffness, separation, and the formation of ice fragments that participate in the accumulation and evolution.

[0034] A high-concurrency graphics processor (GPU) repositions the discrete spherical units that form ice fragments within a generation region that is a preset multiple of the ice thickness at the leading edge of the marine platform structure model, based on a uniform distribution function, and writes them into the sea ice particle array. As a preferred approach, the preset ice thickness is set to 3 to 5 times the average sea ice thickness. This range ensures that while the ice fragments detach from the current high-frequency collision zone, they remain within the hydrodynamic influence domain at the leading edge of the structure, simulating realistic ice fragment flow and accumulation. The GPU outputs local contact forces in real time and records them into the sea ice particle array for later use.

[0035] S3, the multi-core central processing unit (CPU) reads the local contact forces from the sea ice particle array, transmits them to the structural node array, and performs dynamic response calculations for the offshore platform structural model within it. The CPU simplifies the dynamic response into a multi-degree-of-freedom or single-degree-of-freedom oscillator model, accumulates the local contact forces output by multiple discrete spherical elements, and calculates the global ice load as a function of relative displacement.

[0036] The multi-core central processing unit solves the motion equations under the first mode to quantify the numerical correlation between structural displacement and global ice load, and obtains the dynamic response results including real-time structural displacement and real-time structural velocity and acceleration generated under global ice load, which are then recorded and stored in the structural node array in real time.

[0037] S4. The heterogeneous computing platform performs cross-domain boundary search and bidirectional parameter transfer tasks. A multi-core CPU and a high-concurrency GPU determine the spatial overlap of boundary nodes between the discrete element domain and the finite element domain. It extracts the 3D spatial coordinates of spherical discrete elements and ice fragments from the sea ice particle array, and the spatial node coordinates of finite element beam elements and flat shell elements from the structural node array. The overlapping areas are then defined as the execution boundary node range for the cross-domain boundary search. The execution boundary node range dynamically changes with the evolution of ice fragment accumulation. Local contact failure states within the changing range are identified in real time to trigger parameter updates at the contact interface.

[0038] The multi-core CPU and high-concurrency GPU perform bidirectional parameter transfer within the locked execution boundary node range. The multi-core CPU extracts real-time structural displacement and velocity from the structural node array, transmits them into GPU memory via the memory bus, and imports them into the sea ice particle array. The discrete spherical elements and ice fragments update the contact algorithm input conditions according to the input parameters. The high-concurrency GPU extracts the local contact force output by the ice fragments from the sea ice particle array, transmits it in reverse via the memory bus into GPU memory, and imports it into the structural node array to update the global ice load, completing bidirectional data synchronization.

[0039] S5. The heterogeneous computing platform performs adaptive time-step calculations and updates the ice accumulation evolution state. The motion equations of the discrete sphere units are solved in the time domain using the fourth-order Runge-Kutta scheme. Since independent fragmentation events of multiple discrete sphere units within the discrete element domain need to be handled under high-speed collisions, the heterogeneous computing platform employs an adaptive time-step algorithm to optimize the integration step size. A multi-core CPU and a high-concurrency GPU monitor the relative motion rate between the discrete sphere units and the marine platform structural model in real time, as well as the number of fragmentation events triggered by critical deformation. A Rayleigh time-step calculation model is introduced to obtain the baseline critical step size.

[0040] To prevent time integration divergence caused by abnormal parameters or the extreme softening of ice materials leading to a near-zero shear modulus, a truncation logic is implemented during computation: if the current shear modulus is lower than a preset safety lower limit (0.01 GPa to 0.1 GPa), the reference critical step size is forcibly reset to the minimum integration period allowed by the underlying hardware. Under normal conditions, the reference critical step size formula satisfies the following: ; In the formula: This is the baseline critical step size for a single time integration operation; The average radius of the discrete element of the sphere; This represents the current density of sea ice in the simulated sea area. This represents the shear modulus of the sea ice material.

[0041] A baseline critical step size is introduced, along with a dynamic scaling factor, to adjust the step size. When the relative motion rate and the number of breakage events increase on the heterogeneous computing platform, the time step size is reduced based on the decreased dynamic scaling factor (ranging from 0.01 to 0.1). When the ice fragments enter the stable accumulation stage, the time step size is increased (at which point the dynamic scaling factor ranges from 0.2 to 0.3). The varying time step size ensures synchronized data transfer between the memory of the multi-core CPU and the video memory of the high-concurrency graphics processor.

[0042] The heterogeneous computing platform performs time-integration calculations based on time-step numerical steps, synchronously updating the 3D state data of sea ice accumulation evolution in front of the marine platform structural model. When the marine platform structural model is set as an upright structure, the interaction between the spherical discrete units and the upright structure generates 3D evolution data of continuous compression failure and intermittent fracturing. When set as a conical structure, the interaction of the spherical discrete units generates 3D evolution data of bending failure. The fracturing ice fragments are blocked by wind, ocean currents, and reaction forces, forming ice shards in front of the structure. The ice shards output stable 3D evolution data of ice shard clearing and ice shard accumulation cycles at low ice drift velocities.

[0043] After the multi-core central processing unit (CPU) outputs the evolution data, it extracts the global ice load static peak value when the marine platform structural model is set as a cone structure and performs a limit numerical comparison. The theoretical limit of the static peak value is obtained with reference to the two-dimensional cone extreme static ice load calculation model.

[0044] Reference Appendix Figure 3 The cross-domain model construction and initialization performed within the heterogeneous computing platform encompasses the construction of a discrete sea ice model, which utilizes the video memory of a high-concurrency graphics processor to perform low-level data partitioning and parameter definition to generate a sea ice particle array. Multiple sea ice particles are configured as multiple spherical discrete units, each with its own independent data storage space within the sea ice particle array.

[0045] Each discrete spherical unit is independently set with three spatial motion parameters, including the first degree of freedom, the second degree of freedom and the third degree of freedom, which respectively represent the spatial displacement in the first dimension, the second dimension and the third dimension, and are synchronously recorded inside the sea ice particle array.

[0046] The initial coordinates of the discrete spherical element within the sea ice particle array are spatially located according to a uniform distribution function, and the initial relative position with respect to the marine platform structure model is output. The distribution satisfies the following formula: ; In the formula: Let be the initial spatial displacement of the discrete spherical element in the first dimension; The initial spatial displacement of the discrete spherical element in the second dimension; Let be the initial spatial displacement of the discrete spherical element in the third dimension; This represents the initial position of the marine platform structure model in the global coordinate system. It is a uniformly distributed random function; This represents the maximum offset of the sea ice generation region relative to the offshore platform structure model. The initial drift velocity of the sea ice is set in the global coordinate system; It represents the critical time span from the initial contact structure interface to the complete destruction of the elastic bonds in a single discrete sphere at a given drift velocity.

[0047] Adjacent discrete spherical units establish underlying connection relationships through elastic bonds to generate a bonded particle model. Three-dimensional spatial coordinates and spatial motion parameters complete the initial loading, providing data support for concurrent contact retrieval and solving motion equations.

[0048] The multi-core central processing unit synchronously constructs the marine platform structural model in memory, retrieving finite element beam elements, flat shell elements, and solid elements within the finite element domain. When the marine platform structural model is set as an upright structure, the finite element beam elements, flat shell elements, and solid elements are meshed and spatially combined to generate the spatial node coordinates of the upright structure; when set as a conical structure, the spatial node coordinates of the conical structure are generated by spatial combination according to a predetermined tilt angle, and uniformly written into the structural node array for storage.

[0049] The multi-core CPU sets initial dynamic parameters within the structural node array, encompassing the initial equivalent structural mass, equivalent structural damping ratio, and stiffness matrix. Based on spatial distribution and node density data, it calculates the initial equivalent structural mass, sets the equivalent structural damping ratio to a fraction of the critical damping, and generates the stiffness matrix based on the elastic modulus and node geometric characteristics, simultaneously writing it internally. The marine platform structural model completes data loading in memory and maintains time alignment with the discrete sea ice model data in GPU memory, providing the data foundation for structural dynamic response calculations. A high-concurrency GPU allocates computing cores to execute multi-threaded contact state determination using the bonded fracture algorithm.

[0050] The high-concurrency graphics processing unit's computational core is divided into multiple thread blocks to perform concurrent retrieval calculations on the data within the sea ice particle array. It extracts the three-dimensional spatial coordinates and spatial motion parameters of adjacent discrete spherical units from the sea ice particle array, compares and calculates relative displacement values, and classifies them into compression and tension states based on these values. When the relative displacement values ​​indicate proximity, the state is determined to be under compression, triggering a hard contact algorithm to limit spatial overlap and penetration, and generating repulsive contact forces. When the values ​​indicate distance, the state is determined to be under tension, relying on elastic bonds to transfer loads and synchronously calculate deformation. The loads and repulsive contact forces output by the elastic bonds are combined and recorded within the sea ice particle array for data storage, providing the underlying data support for solving the motion equations.

[0051] The computational core of the high-concurrency graphics processor retrieves the spatial motion parameters of the discrete sphere elements from the sea ice particle array and performs calculations to solve the motion equations of the discrete sphere elements. The internal relative motion of the discrete sphere elements is characterized by dynamic velocity. The computational core extracts the given ice drift velocity, linear spring coefficient, nonlinear spring stiffness function, linear viscous damper coefficient, and nonlinear back damper coefficient, and allocates them to independent threads for concurrent numerical solutions.

[0052] The dynamic rate calculation formula satisfies the following: ; ; In the formula: Let be the derivative of the spatial displacement of the discrete element in the second dimension with respect to time. Let be the derivative of the spatial displacement of the discrete element in the third dimension with respect to time. This is the local nonlinear spring stiffness function of the discrete element of the sphere; For the linear viscous damper coefficient inside the discrete element of the sphere; The linear spring constant inside the discrete element of the sphere; The coefficient of the nonlinear post-damper after the elastic bond fails and fractures.

[0053] After the second and third degrees of freedom of the discrete element on the sphere are calculated, they are input into the sea ice particle array by the computational core of the high-concurrency graphics processor. The spatial displacement of the discrete element on the sphere changes according to the output result of the internal relative motion rate calculation formula. The high-concurrency graphics processor synchronizes the spatial displacement change data in real time in the video memory, completes the dynamic rate update of multiple sea ice particles in the discrete element domain, and provides the computational input conditions for elastic bond failure and ice fragmentation.

[0054] The high-concurrency graphics processing unit's computing core extracts spatial displacement variation data of adjacent discrete spherical elements to determine critical deformation, comparing the deformation of the elastic bond with a pre-set threshold. The deformation of the elastic bond is determined by the difference between the second and first degrees of freedom of the discrete spherical elements. When the difference reaches a pre-determined critical deformation, the elastic bond failure and fracture condition is satisfied. The elastic bond failure and fracture condition follows the formula below: ; In the formula: The initial spatial displacement of the discrete spherical element in the second dimension; Let be the initial spatial displacement of the discrete spherical element in the first dimension; The maximum critical deformation threshold allowed for fracture failure caused by elastic bonds is not randomly selected in this embodiment. Instead, it is calculated by calibration tests based on the macroscopic tensile and shear strength of the simulated sea ice material and the average radius of the discrete sphere unit, so as to truly reflect the brittle fracture characteristics of sea ice.

[0055] When the elastic bond failure condition is met, the elastic bonds between adjacent discrete spherical units fail and break, detaching the connection and forming ice fragments. The computational core of the high-concurrency graphics processor performs a spatial relocation operation on the discrete spherical units that form ice fragments. The spatial relocation operation calls a uniform distribution function to generate updated 3D spatial coordinates for the discrete spherical units that form ice fragments. The high-concurrency graphics processor writes the updated 3D spatial coordinates into the sea ice particle array. The data of the discrete spherical units stored inside the sea ice particle array changes synchronously, providing 3D state data of ice fragment evolution within the discrete element domain.

[0056] The multi-core central processing unit extracts the local contact force generated by the discrete units of the sphere during contact evolution from the video memory of the high-concurrency graphics processor, and imports it into the internal structure node array via the memory bus for later use.

[0057] The multi-core CPU performs a summation calculation of local contact forces within the structural node array and calls the Heaviside step function to determine the contact state. The Heaviside step function performs a difference calculation based on the first degree of freedom and the displacement of the offshore platform structural model. The multi-core CPU filters out the spherical discrete elements in contact based on the difference calculation results, and incorporates the local contact forces output by these elements into the integral summation calculation, integrating them into a global ice load acting on the offshore platform structural model. The global ice load is calculated according to the global ice load summation formula, which satisfies the following: ; In the formula: This represents the global ice load acting on the surface of the marine platform structural model at the current computation moment. This represents the total number of discrete spherical units participating in the interaction at the current moment. This is the Heaviside step function (used to determine whether the element and the structure have effective contact). This represents the displacement of the marine platform structural model in the global coordinate system. The current simulation progresses the integral time.

[0058] The calculated global ice load is stored within the structural node array. The multi-core CPU updates the global ice load data within the finite element domain. The global ice load transforms the local contact forces of discrete particles into external force input conditions for the offshore platform structural model, providing input parameters for subsequent finite element dynamic response calculations.

[0059] A multi-core CPU performs structural dynamics equation solving within the structural node array. The dynamic response of the offshore platform structural model is set as a multi-degree-of-freedom or single-degree-of-freedom oscillator model. Global ice load, equivalent structural mass, the equivalent structural damping ratio (a fraction of the critical damping), and natural frequencies are extracted from the structural node array. Quantitative calculations are then performed on the dynamic state of the offshore platform structural model under ice-induced vibration. The dynamic state is characterized by real-time structural displacement, real-time structural velocity, and real-time structural acceleration. The multi-core CPU calls the first-mode dynamics equation to calculate the real-time structural displacement, real-time structural velocity, and real-time structural acceleration. The first-mode dynamics equation satisfies the following formula: ; In the formula: The equivalent structural mass of the marine platform structural model; Real-time structural acceleration of an offshore platform structure model under ice-induced vibration; The equivalent structural damping ratio serves as the critical damping fraction. The natural frequency of the first mode of the marine platform structural model; Real-time structural velocity of an offshore platform structure model under ice-induced vibration; Real-time structural displacement of an offshore platform structure model under ice load; For global ice load; This is the current load integration time.

[0060] Based on the general principles of finite element structural dynamics, a multi-core CPU is used to solve the first-mode dynamic equation of the structure by inputting the extracted global ice load, equivalent structural mass, equivalent structural damping ratio (a fraction of the critical damping), and the structure's natural frequencies. To prevent matrix singularity problems caused by structural stiffness degradation under extreme loading conditions, a minimum diagonal perturbation term is introduced into the stiffness matrix before solving to ensure the matrix is ​​positive definite. The solution outputs the real-time structural displacement and velocity under ice-induced vibration, thereby quantifying the transient dynamic characteristics of the structure under ice load impact.

[0061] The multi-core central processing unit writes real-time structural displacements and velocities into the structural node array for storage. The dynamic response state data of the marine platform structural model stored in the structural node array is updated in real time. The real-time structural displacements and velocities of the marine platform structural model are synchronously fed back to the GPU memory and RAM of the heterogeneous computing platform, providing kinematic input parameters for cross-domain boundary search and bidirectional parameter transfer tasks.

[0062] The real-time structural displacement and velocity of the marine platform structural model are synchronously fed back to the video memory and main memory of the heterogeneous computing platform to provide kinematic input parameters for cross-domain boundary search and bidirectional parameter transfer tasks. Then, the multi-core central processing unit and high-concurrency graphics processing unit perform cross-domain boundary search calculations.

[0063] In the dynamic changes of sea ice accumulation and ice fragmentation, a multi-core CPU and a high-concurrency GPU continuously locate the spatial overlap regions between the structural interface and the ice fragments. The high-concurrency GPU extracts the 3D spatial coordinates of the discrete spherical elements and the updated 3D spatial coordinates of the ice fragments in real time from the sea ice particle array. The multi-core CPU extracts the spatial node coordinates of the finite element beam elements and the spatial node coordinates of the flat shell elements in real time from the structural node array. The multi-core CPU and the high-concurrency GPU uniformly map the 3D spatial coordinates of the discrete spherical elements, the updated 3D spatial coordinates of the ice fragments, the spatial node coordinates of the finite element beam elements, and the spatial node coordinates of the flat shell elements into a global coordinate system constructed within the heterogeneous computing platform.

[0064] After mapping, the multi-core CPU and high-concurrency GPU perform coordinate distance traversal calculations within the global coordinate system. The heterogeneous computing platform calculates the spatial distance between the structural interface composed of finite element beam elements and flat shell elements, the spherical discrete elements, and the ice fragments. Based on the spatial distance, a threshold screening calculation is performed, defining the three-dimensional spatial range where the spatial distance is less than the preset collision tolerance as the spatial overlap region. As a preferred approach, the preset collision tolerance is set to 0.05 to 0.1 times the average radius of the spherical discrete elements. This range is determined based on the truncation error of the single-precision floating-point operation of the heterogeneous computing platform and the protrusion size of the sea ice surface, avoiding unnecessary memory usage while ensuring that no contact retrieval is missed.

[0065] The spatial overlap region indicates the spatial boundary where the structural interface, discrete spherical units, and ice fragments directly contact each other. Accompanying the real-time structural displacement changes and ice fragment updates of the offshore platform structural model, the multi-core CPU and high-concurrency GPU repeatedly execute coordinate distance traversal calculations and threshold screening calculations within a time step, achieving dynamic locking of the spatial overlap region. The dynamically locked spatial overlap region stores the distribution data of the structural interface and ice fragments, directly providing the data interaction boundary range for subsequent bidirectional parameter transfer tasks.

[0066] The dynamically locked spatial overlapping region stores the distribution data of finite element beam elements, flat shell elements, spherical discrete elements, and ice fragments. After locking the data interaction boundary range, the multi-core CPU and high-concurrency graphics processor rely on the memory bus to perform synchronous data transfer between memory and video memory.

[0067] Multiple multi-core CPUs extract real-time structural displacements and velocities from the marine platform structural model via a self-structured node array. These extracted values ​​are then transferred via memory bus to the video memory of a high-concurrency graphics processor (GPU) and synchronously imported into the sea ice particle array. The GPU's computational cores perform concurrent calculations on the data within the spatially overlapping region, extracting the real-time structural displacements and velocities of the marine platform structural model as the underlying input conditions for contact state determination. Within the spatially overlapping region, the spherical discrete units and ice fragments calculate their internal relative motion rates based on the real-time structural displacements and velocities of the marine platform structural model, updating their relative displacement values ​​and dynamic rates.

[0068] The high-concurrency graphics processing unit (GPU) core performs contact state determination and solves the motion equations of the discrete sphere elements, extracting local contact forces from the overlapping areas of the discrete sphere elements and ice fragments within the sea ice particle array. The GPU packages these local contact forces and sends them to the memory bus. The memory bus then transmits these local contact forces in reverse to the memory of the multi-core CPU. The multi-core CPU receives these local contact forces and synchronously imports them into the structural node array. Within the structural node array, the multi-core CPU extracts these local contact forces and performs summation calculations, updating the global ice load acting on the marine platform structural model in real time. Kinematic and dynamic parameters are continuously exchanged bidirectionally between the multi-core CPU and the GPU via the memory bus, maintaining data synchronization between the finite element structural dynamic equation solving and discrete element contact state determination calculations.

[0069] After kinematic and dynamic parameters are continuously exchanged bidirectionally between the multi-core CPU and the high-concurrency graphics processor via the memory bus, the multi-core CPU and the high-concurrency graphics processor jointly execute the time-step adaptive algorithm.

[0070] The heterogeneous computing platform monitors the relative motion rate and the number of fracture events between the discrete spherical unit and the marine platform structural model in real time within video memory and system memory. The number of fracture events is determined by the number of elastic bonds that satisfy the elastic bond failure fracture condition. The heterogeneous computing platform performs numerical comparison calculations on the relative motion rate and the number of fracture events.

[0071] When the relative motion rate between the discrete element and the marine platform structure model exceeds a preset rate threshold and the number of fracturing events exceeds a preset fracturing threshold, the multi-core CPU and high-concurrency GPU determine that the sea ice and the marine platform structure model are in a state of high-speed impact and complex fragmentation. Specifically, the preset rate threshold is calibrated to 1.5 to 2.0 times the initially set sea ice drift speed, and the preset fracturing threshold is set to 5% to 10% of the total number of elastic bonds in the current discrete element domain within a single time step.

[0072] When subjected to high-speed impact and complex fragmentation, the multi-core CPU and high-concurrency GPU trigger an adaptive time-step algorithm to reduce the time-step value. The reduced time-step value increases the computational density of solving the motion equations of the discrete sphere elements in the time domain using the fourth-order Runge-Kutta scheme, capturing the high-frequency contact response of the discrete sphere elements under high-speed collision.

[0073] When the relative motion rate between the discrete sphere unit and the marine platform structural model is below a preset rate threshold and the number of fracturing events is below a preset fracturing threshold, the multi-core CPU and high-concurrency GPU determine that the ice fragments formed by the fracturing of the discrete sphere unit have entered a stable accumulation stage. Upon entering the stable accumulation stage, the multi-core CPU and high-concurrency GPU trigger an adaptive time step algorithm to increase the time step value. The increased time step value reduces the frequency of numerical computation performed by the multi-core CPU and high-concurrency GPU.

[0074] The varying time step values ​​are synchronously written to the internal storage domain of the heterogeneous computing platform by the multi-core CPU and the high-concurrency GPU. The CPU's memory and the GPU's video memory interact based on the varying time step values. The varying time step values ​​ensure strict alignment of the data from the sea ice discrete model and the ocean platform structure model in the time dimension, providing the underlying time reference for subsequent evolutionary state updates.

[0075] The variable time step ensures that the data from the discrete sea ice model and the marine platform structure model are strictly aligned in the time dimension. Then, the multi-core CPU and high-concurrency GPU execute the three-dimensional evolution output calculation of the entire stacking evolution scenario according to the time integral step.

[0076] The multi-core CPU and high-concurrency GPU perform spatial distribution calculations of the marine platform structural model and spherical discrete elements within the global coordinate system. When the marine platform structural model is set as an upright structure, the multi-core CPU and high-concurrency GPU generate continuous compression failure distribution data and intermittent fragmentation distribution data of the spherical discrete elements in front of the upright structure. When the marine platform structural model is set as a conical structure, the multi-core CPU and high-concurrency GPU generate bending failure distribution data, stable ice debris removal distribution data, and accumulation distribution data of the spherical discrete elements in front of the conical structure. The multi-core CPU and high-concurrency GPU merge the continuous compression failure distribution data, intermittent fragmentation distribution data, bending failure distribution data, stable ice debris removal distribution data, and accumulation distribution data into a three-dimensional evolution output result.

[0077] After the 3D evolution output is generated in memory and video memory, the multi-core CPU simultaneously extracts the global ice load acting on the marine platform structural model. The multi-core CPU and high-concurrency graphics processor call the static peak reference benchmark to perform limit state verification on the 3D evolution output and the global ice load. The static peak reference benchmark is calculated and generated by the two-dimensional cone extreme static ice load auxiliary calibration formula. The two-dimensional cone extreme static ice load auxiliary calibration formula satisfies the following: ; In the formula: The extreme static ice load of a two-dimensional cone; This is the first geometric and friction correction coefficient; The bending strength of sea ice; This represents the average thickness of the sea ice. The density of the local seawater; It is the acceleration due to gravity; The effective width of the cone-shaped structure of the offshore platform at the waterline; The effective elastic modulus of sea ice; This is the second geometric and friction correction coefficient; The maximum height to which ice fragments accumulate along the surface of the cone is taken as twice the thickness of the sea ice. Sea ice density; The angle between the ice-facing surface of the cone and the horizontal sea level; It represents the coefficient of kinetic friction between sea ice and the surface of the structure.

[0078] The multi-core central processing unit outputs the two-dimensional cone extreme static ice load according to the auxiliary calibration formula of the two-dimensional cone extreme static ice load; it performs a numerical comparison operation between the global ice load and the two-dimensional cone extreme static ice load, and multiplies the two-dimensional cone extreme static ice load by the set safety margin coefficient (the set safety margin coefficient, for example, is 1.1 to 1.5) as the judgment boundary.

[0079] If the global ice load does not exceed the numerical boundary, the 3D evolution output is determined to meet the objective reality benchmark. The 3D evolution output that meets the benchmark is synchronously written to the persistent storage space of the heterogeneous computing platform's hard disk, completing the data update cycle of cross-domain evolution and adaptive solution. Conversely, if the global ice load exceeds the numerical boundary, the current simulation parameters are determined to be distorted. The multi-core CPU triggers an abnormal rollback mechanism, restoring the simulation state to the previous time point that met the benchmark. The damping parameter is adaptively increased, and the evolution calculation is re-executed, thereby ensuring the logical closed loop and reliability of the entire simulation results.

[0080] Specific application examples: Based on the on-site simulation verification of the anti-icing structure design of a cone-shaped offshore platform in a typical Bohai Sea area, the constants and initial parameters set by the method of this invention within the heterogeneous computing platform are as follows: The lower limit for preventing divergence in sea ice materials is set at a minimum constant of 1.0 × 10⁻⁶. 5s; the preset dynamic scaling factor is set to 0.2 for stable stacking and 0.05 for high-speed impact; the average radius of the discrete spherical element is set to 0.5m; the maximum critical deformation threshold allowed for fracture failure caused by elastic bond is set to 0.02m.

[0081] During the adaptive time-step computation on the heterogeneous computing platform, the multi-core CPU extracts the density of sea ice in the current simulated sea area as 900 kg / m³, and the shear modulus of the sea ice material as 1.0 × 10⁹ Pa. Based on the current environment and model state, the multi-core CPU substitutes the baseline critical step size formula to calculate the baseline critical step size for a single time integration operation: ; The multi-core CPU will calculate 1.49 × 10 3s was used as the theoretical baseline. In the initial stage, since no high-frequency breakage events occurred, the system was determined to have entered a stable accumulation phase. Based on the extracted dynamic scaling factor of 0.2, the heterogeneous computing platform expanded the actual execution time step to 2.98 × 10⁻⁶. The 4s time eliminates redundant computational burden caused by tiny high-frequency oscillations, preventing excessive consumption of computing power.

[0082] When the heterogeneous computing platform actively applies ice drift velocity to simulate actual sea ice collisions, the high-concurrency graphics processor dynamically quantizes the contact state. The computing core continuously monitors the elastic bond deformation determined by the difference between the second and first degrees of freedom, extracting the spatial displacement of adjacent spherical discrete units at a certain moment as 1.52m in the second dimension and 1.50m in the first dimension. The computing core substitutes the judgment condition: ; The calculated deformation of 0.02m has reached the predetermined maximum critical deformation threshold, triggering the judgment condition. The elastic bond between adjacent elements fails and breaks, leading to a weakening of the contact stiffness and separation to form ice fragments. This fracture event is recorded as an increment in the number of fracture events, used to trigger the time step adaptive algorithm in the future, ensuring that no destructive integral divergence occurs during the high-speed fragmentation process.

[0083] As local contact forces converge and are transmitted across domains, the multi-core CPU completes multiple rounds of dynamic response calculations. In the latest round of integral calculations, the multi-core CPU uses the Heaviside step function to filter out locally stressed particles that are in effective contact, extracting the nonlinear spring force components of the three main load-bearing sphere discrete elements as 250000.0 N, 260000.0 N, and 240000.0 N, respectively. The multi-core CPU then substitutes these components into the global ice load accumulation formula to calculate the existing forces: ; Subsequently, the heterogeneous computing platform extracted the verification parameters when the marine platform was set as a cone structure: sea ice bending strength of 6.0 × 10⁵ Pa, average thickness of 0.5 m, effective width of 5.0 m, and maximum climbing height of 1.0 m. The multi-core CPU substituted these parameters into the two-dimensional cone extreme static ice load auxiliary calibration formula to calculate the theoretical extreme value of approximately 0.85 MN. Since the currently extracted global ice load static peak value (0.93 MN) is less than the judgment boundary after the safety margin coefficient is amplified (0.85 × 1.2 = 1.02 MN), the verification condition is passed. The multi-core CPU determines that the current three-dimensional evolution output result meets the objective reality benchmark and writes the result into the persistent storage space of the heterogeneous computing platform's hard disk.

[0084] Actual operation tests were conducted and data were compared. Specific verification results are detailed in the appendix. Figure 3 and attached Figure 4 The experiment compared the computational state data of the traditional single-core CPU fixed-step discrete element simulation method with the GPU-accelerated discrete element simulation method of this invention under the same particle size and the same initial distribution of ice drift. The experimental data are recorded in Table 1.

[0085] Table 1. Comparison of Operating Data between Traditional Simulation Method and Simulation Method of the Present Invention

[0086] See attached document Figure 3 And Table 1, Appendix Figure 3 This diagram compares the response of global ice load over time during the entire simulation period. The horizontal axis represents the simulation integration time in seconds (s), ranging from 0 to 100. The vertical axis represents the global ice load in MN (meters), ranging from 0 to 1.5. The diagram includes load evolution curves from both the traditional simulation method and the method described in this invention.

[0087] During the high-speed collision phase with an integration time ranging from 20s to 40s, traditional simulation methods fail to adaptively adjust the transient high-frequency response caused by local contact failure, resulting in micro-divergence of the time integral locally. This leads to the ice load response failing to recover in time and generating peak distortion, ultimately resulting in a distorted state with a false extremum of 1.35MN. The method of this invention extracts the local contact state from the discrete element domain within video memory, updates the dynamic parameters across domains, and performs calculations based on the fourth-order Runge-Kutta scheme. The load evolution curve of this method exhibits obvious periodic continuous compression failure and intermittent fracture drop characteristics. It can rapidly unload when multiple fracture events occur, and the peak value is strictly controlled within the extreme reference boundary, completely eliminating load stiffness singularities and numerical accumulation caused by fixed step sizes.

[0088] See attached document Figure 4 And Table 1, Appendix Figure 4This demonstrates the adaptive scaling control of the underlying time integration mechanism under rapid changes in relative motion rate. The horizontal axis represents the simulation execution time in seconds, ranging from 0 to 100. The left vertical axis represents the time step size in seconds, ranging from 0 to 3.5 × 10⁻⁶. 4; The right-hand vertical axis represents the joint triggering ratio of rupture events and rates, dimensionless, with a value range of 0 to 1.2. The figure includes a time step state line of the method of the present invention, a triggering ratio line of the method of the present invention, and a preset anti-divergence triggering threshold baseline.

[0089] Around 30 seconds into the simulation, the number of breakup events increased dramatically due to widespread sea ice contact disruption. Traditional simulation methods, lacking bidirectional data communication and adaptive computation mechanisms, directly overflowed within the original computation cycle, causing a total of 12 time integral divergences.

[0090] This invention's method utilizes a multi-core CPU and a high-concurrency graphics processor to jointly extract the ratio of relative motion rate to the number of rupture events, dynamically updates the changing time step value, and synchronizes execution parameters across domains on a heterogeneous computing platform. As can be seen from the trigger ratio curve of this invention, when the joint trigger ratio approaches the preset anti-divergence trigger threshold baseline of 1.0, the heterogeneous computing platform automatically and drastically reduces the time step value to 7.4 × 10⁻⁵ based on a reduced dynamic scaling factor of 0.05. The computation frequency is strictly limited to a safe range that can capture high-frequency fluctuations within 5 seconds. This ensures that the number of time integral divergences in the method of this invention remains zero throughout the entire evolution cycle, guaranteeing the evolutionary stability of the multidimensional simulation model and the safety of hardware computation.

Claims

1. A GPU-accelerated discrete element method for simulating the evolution of ice fragmentation, characterized in that, Includes the following steps: Deploy the marine platform structure model in the memory of a multi-core central processing unit to generate a structure node array, and deploy discrete spherical units in the video memory of a high-concurrency graphics processor to generate a sea ice particle array. The high-concurrency graphics processor is used to execute the bonding and breaking algorithm on the discrete unit of the sphere. When the critical deformation condition is met, ice fragments are formed, and the local contact force is output and recorded in the sea ice particle array. The multi-core central processing unit is used to read the local contact force to calculate the global ice load, solve the motion equation to obtain the real-time structural displacement and real-time structural velocity, and input them into the structural node array. By comparing the spatial overlap areas of the structural node array and the sea ice particle array, the execution boundary node range is delineated. Within the execution boundary node range, the real-time structural displacement and the real-time structural velocity are imported into the sea ice particle array, and the local contact force is reversed and imported into the structural node array to update the global ice load. Based on the updated sea ice particle array and the structural node array, a time step adaptive algorithm is invoked to perform time integration, generating a three-dimensional evolution output result of the ice debris accumulation evolution in front of the marine platform structural model.

2. The discrete element method for simulating the evolution of ice fragmentation based on GPU acceleration as described in claim 1, characterized in that, The generated structure node array and the generated sea ice particle array specifically include: The initial spatial node coordinates of finite element beam elements, flat shell elements, and solid elements are written into the structural node array using the multi-core central processing unit to establish the marine platform structural model, and the initial equivalent structural mass, equivalent structural damping ratio, and stiffness matrix parameters are set. The high-concurrency graphics processor is used to control the discrete sphere units to form a bonded particle model by connecting them with elastic bonds. Each discrete sphere unit is given a first degree of freedom, a second degree of freedom, and a third degree of freedom representing spatial displacement in the first, second, and third dimensions. The initial spatial coordinates and motion velocity components are synchronously recorded into the sea ice particle array to establish a low-level data binding relationship, thus completing the generation of the structural node array and the sea ice particle array.

3. The discrete element method for simulating the evolution of ice fragmentation based on GPU acceleration according to claim 2, characterized in that, The specific steps of performing the bonding and breaking algorithm on the discrete spherical units include: The high-concurrency graphics processor continuously monitors the deformation of the elastic bond, which is determined by the difference between the second degree of freedom and the first degree of freedom, and compares it with the predetermined critical deformation condition. When the critical deformation condition is met, the elastic bond between adjacent discrete spherical units fails and breaks, resulting in weakened contact stiffness and separation to form the ice fragments. Based on the uniform distribution function, the spherical discrete unit corresponding to the broken ice is repositioned in the generation area of ​​a preset multiple of ice thickness at a distance from the leading edge of the marine platform structure model, written into the sea ice particle array and outputting the local contact force. The preset multiple of ice thickness is 3 to 5 times the average thickness of sea ice.

4. The discrete element method for simulating the evolution of ice fragmentation based on GPU acceleration as described in claim 1, characterized in that, The calculation of the global ice load by reading the local contact force specifically includes: The dynamic response is set to a multi-degree-of-freedom or single-degree-of-freedom oscillator model using the multi-core central processing unit. The Heaviside step function is called inside the structural node array to determine the contact state and output the difference calculation result. Based on the difference calculation results, the discrete sphere elements that have made contact are selected, and the local contact force output by the discrete sphere elements is included in the integral accumulation calculation range to obtain the global ice load.

5. The discrete element method for simulating the evolution of ice fragmentation based on GPU acceleration as described in claim 1, characterized in that, The process of solving the motion equations to obtain real-time structural displacement and real-time structural velocity, and then inputting these values ​​into the structural node array, specifically includes: The global ice load, equivalent structural mass, equivalent structural damping ratio, and natural frequency are extracted from the structural node array using the multi-core central processing unit. The multi-core central processing unit is used to call the first modal dynamic equation of the structure to quantify the numerical correlation between the real-time structural displacement and the global ice load, and obtain the dynamic response results including the real-time structural displacement and the real-time structural velocity and real-time structural acceleration generated under the global ice load, and record them into the structural node array.

6. The discrete element method for simulating the evolution of ice debris accumulation based on GPU acceleration according to claim 2, characterized in that, The definition of the execution boundary node range specifically includes: Determine the spatial overlap of boundary nodes between the discrete element domain and the finite element domain; The three-dimensional spatial coordinates of the spherical discrete unit and the ice fragments are extracted from the sea ice particle array, and the spatial node coordinates of the finite element beam unit and the flat shell unit are extracted from the structural node array. Calculate the spatial distance and perform threshold screening calculations. Define the three-dimensional spatial range where the spatial distance is less than the preset collision tolerance as the spatial overlapping region. Dynamically lock the execution boundary node range. The preset collision tolerance is 0.05 to 0.1 times the average radius of the spherical discrete unit.

7. The discrete element method for simulating the evolution of ice fragmentation based on GPU acceleration as described in claim 1, characterized in that, The step of importing the real-time structural displacement and the real-time structural velocity into the sea ice particle array, and importing the local contact force in reverse into the structural node array, specifically includes: The real-time structural displacement and real-time structural velocity are extracted from the structural node array, transmitted into the video memory of the high-concurrency graphics processor, and imported into the sea ice particle array to update the contact algorithm input conditions. The local contact force is extracted from the sea ice particle array, transmitted in reverse to the memory of the multi-core central processing unit, and imported into the structural node array to update the global ice load, thus completing bidirectional data synchronization.

8. A GPU-accelerated discrete element method for simulating the evolution of ice debris accumulation according to claim 2 or 3, characterized in that, The invocation of the time step adaptive algorithm to perform time integration specifically includes: By using the updated sea ice particle array and the structural node array, a Rayleigh time step calculation model is introduced to obtain the baseline critical step size. If the current shear modulus is lower than the preset safety lower limit, the baseline critical step size is forcibly reset to the minimum integration period. The preset safety lower limit is between 0.01 GPa and 0.1 GPa. When the relative motion rate between the discrete sphere unit and the marine platform structure model exceeds a preset rate threshold and the number of fracture events exceeds a preset fracture threshold, the time step value is reduced. The preset rate threshold is 1.5 to 2.0 times the initially set sea ice drift speed, and the preset fracture threshold is 5% to 10% of the total number of elastic bonds in the current discrete element domain within a single time step. When the relative motion rate is lower than the preset rate threshold and the number of rupture events is lower than the preset rupture threshold, the time step value is increased to generate a variable time step value and the time integration operation is performed.

9. The discrete element method for simulating the evolution of ice debris accumulation based on GPU acceleration as described in claim 1, characterized in that, The three-dimensional evolution output of the ice debris accumulation evolution in front of the marine platform structural model specifically includes: When the marine platform structure model is set as a vertical structure, continuous compression failure distribution data and intermittent breakage distribution data are generated. When the marine platform structural model is set as a cone structure, bending failure distribution data, stable ice debris removal distribution data, and accumulation distribution data are generated. The continuous compression failure distribution data, the intermittent breakage distribution data, the bending failure distribution data, the stable ice pile clearing distribution data, and the accumulation distribution data are combined into the three-dimensional evolution output result.

10. The discrete element method for simulating the evolution of ice debris accumulation based on GPU acceleration according to claim 9, characterized in that, The result, after being merged into the three-dimensional evolution output, also includes: The global static peak value of ice load was extracted when the marine platform structure model was set as a cone structure, and the extreme static ice load of the two-dimensional cone was obtained by referring to the auxiliary calibration formula of the extreme static ice load of the two-dimensional cone. Perform a numerical comparison operation between the global ice load and the two-dimensional cone extreme static ice load, and multiply the two-dimensional cone extreme static ice load by a set safety margin factor as the judgment boundary. The set safety margin factor is between 1.1 and 1.

5. When the global ice load exceeds the judgment boundary, it is determined that the current simulation parameters are distorted, triggering an abnormal rollback mechanism to restore the simulation state to the previous time node that meets the benchmark, and then adaptively increasing the damping parameter before re-executing the evolution calculation.