Island engineering material performance simulation method and system

Abstract

The invention discloses an island engineering material performance simulation method and system, and relates to the technical field of island engineering, and the method comprises the steps: S1, obtaining the tide historical data and field monitoring data of an island engineering structure; s2, initializing and maintaining a shared state space, wherein characterization parameters of the shared state space are a microenvironment activity index, a hysteresis memory state and a virtual aggregate field; s3, based on tide historical data and on-site monitoring data, utilizing a reinforcement learning framework to carry out structured editing on a hysteresis memory state so as to update a microenvironment activity index; s4, taking the updated microenvironment activity index as a condition, inputting the updated microenvironment activity index into a physical constraint generation model so as to update a virtual aggregate field and calculate a corresponding pumping dynamic response; s5, on the basis of the pumping dynamic response and the updated microenvironment activity index, in combination with tide-driven boundary conditions, a shared state space is updated; and S6, according to the updated shared state space, outputting a material performance prediction result of the island engineering structure.

Description

Technical Field

[0001] This application relates to the field of island and reef engineering technology, and more specifically, to a method and system for simulating the performance of island and reef engineering materials. Background Technology

[0002] Island and reef engineering structures are subjected to harsh marine environments characterized by high salinity, high humidity, and intense sunlight for extended periods. The degradation of their material properties directly impacts the safety and durability of these projects. Therefore, accurate simulation and prediction of the material properties in island and reef engineering are of significant engineering importance and theoretical research value for guiding material selection, optimizing structural design, developing scientific maintenance strategies, and reducing life-cycle costs.

[0003] However, existing material performance simulation methods still have certain limitations when facing the unique, long-term cyclical tidal effects of island and reef environments. On the one hand, traditional models often fail to fully characterize the complex relationship between tidal dynamic cycles and the response of the material's internal microenvironment, leading to biases in the prediction of long-term performance evolution. On the other hand, existing methods are still insufficient in their ability to learn and characterize the time-varying evolution of material microstructures from limited field monitoring data, which restricts the accuracy of simulation models in reproducing real physical processes and the reliability of predictions.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method and system for simulating the performance of materials used in island and reef engineering to solve the aforementioned technical problems.

[0006] This application provides a method for simulating the material performance of island and reef engineering structures, comprising: S1, acquiring historical tidal data and on-site monitoring data of the island and reef engineering structure; S2, initializing and maintaining a shared state space, wherein the characteristic parameters of the shared state space are a microenvironment activity index, a hysteresis memory state, and a virtual aggregate field; S3, based on the historical tidal data and the on-site monitoring data, using a reinforcement learning framework to perform structured editing of the hysteresis memory state to update the microenvironment activity index; S4, using the updated microenvironment activity index as a condition, inputting it into a physical constraint generation model to update the virtual aggregate field and calculate its corresponding pumping dynamics response; S5, based on the pumping dynamics response and the updated microenvironment activity index, combined with tidal-driven boundary conditions, updating the shared state space; S6, outputting the predicted material performance results of the island and reef engineering structure according to the updated shared state space.

[0007] This application provides a simulation system for the material performance of island and reef engineering structures, comprising: a data acquisition unit for acquiring historical tidal data and on-site monitoring data of the island and reef engineering structure; an initialization unit for initializing and maintaining a shared state space, wherein the characteristic parameters of the shared state space are a microenvironment activity index, a hysteresis memory state, and a virtual aggregate field; a structured editing unit for using a reinforcement learning framework to structure-edit the hysteresis memory state based on the historical tidal data and the on-site monitoring data to update the microenvironment activity index; a first update unit for inputting the updated microenvironment activity index as a condition into a physical constraint generation model to update the virtual aggregate field and calculate its corresponding pumping dynamics response; a second update unit for updating the shared state space based on the pumping dynamics response and the updated microenvironment activity index, combined with tidal-driven boundary conditions; and an output unit for outputting the predicted material performance of the island and reef engineering structure based on the updated shared state space.

[0008] Based on the embodiments provided in this application, by maintaining a shared state space that uniformly characterizes the microenvironment activity index, hysteresis memory state, and virtual aggregate field, key states in the simulation process can co-evolve and be consistently updated. This overcomes the limitations of traditional simulations, such as loose coupling between different physical fields or state variables and delayed information transmission, and provides a unified and dynamic digital carrier for characterizing the complex nonlinear interaction between tidal cycles and material properties.

[0009] By introducing a reinforcement learning framework to structurally edit the hysteresis memory state and using this to drive the update of the microenvironment activity index, the simulation system is endowed with the ability to autonomously learn from historical tidal data and on-site monitoring data and adapt to dynamic changes in the environment. This enables the simulation model to more sensitively capture the cumulative effect and historical dependence of tidal loads, thereby more accurately reflecting the time-varying behavior of material properties in real marine environments and improving the rationality of long-term predictions.

[0010] By constructing a closed-loop simulation process that goes from updating the microenvironment activity index to generating a virtual aggregate field and calculating the pumping dynamics response, and then feeding back to update the shared state space, an integrated simulation of the entire chain of environmental excitation, internal structural response, and macroscopic performance evolution is achieved. This ensures that the evolution of the material's microstructure is always driven by both physical constraints and the current environmental state, significantly enhancing the fidelity of the simulation results to the actual engineering scenario, and ultimately outputting more reliable material performance prediction results. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 A flowchart of an optional simulation method for the properties of island and reef engineering materials according to an embodiment of this application; Figure 2 This is a flowchart of another optional simulation method for the properties of island and reef engineering materials according to an embodiment of this application; Figure 3 A timing diagram of an optional simulation method for the performance of island and reef engineering materials according to an embodiment of this application; Figure 4 This is a structural diagram of an optional island and reef engineering material performance simulation system according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application.

[0012] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0013] The technical solutions of 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0014] In the durability assessment of island and reef engineering structures, traditional chloride ion diffusion models based on Fick's second law exhibit significant prediction bias. These models assume a constant diffusion coefficient, homogeneous media, and stable boundary conditions. However, island and reef engineering structures operate in tropical marine tidal environments, facing the combined effects of high-frequency wet-dry cycles, heterogeneous aggregate distribution, and multi-scale pore structures. Actual observational data shows that due to the pumping effect (capillary suction-drainage) and the evolution of interfacial microstructure, the actual chloride ion transport rate within concrete can deviate from the Fick model's prediction by more than three times, leading to prediction errors in corrosion initiation time often exceeding 200%. This systematic bias is not a random error but stems from the traditional model's insufficient ability to characterize the microstructural features such as aggregate geometry and pore network connectivity, as well as its neglect of the dynamic synchronicity between tidal phase and material response.

[0015] Some implementation methods include: empirical models based on equivalent diffusion coefficients lack explicit descriptions of mesostructures, resulting in significant extrapolation errors; CFD-based multiphysics numerical simulations suffer from the "curse of dimensionality," are computationally expensive, and are difficult to embed with data learning mechanisms; and purely data-driven machine learning methods lack physical consistency constraints, often producing non-physical interpretations that violate conservation laws and fail to provide interpretable information about microstructure evolution. The limitations of these implementation methods stem from a fundamental technical contradiction: materials are highly heterogeneous (random aggregate distribution, localized damage), service environments exhibit strong periodicity (tides, diurnal cycles), while available field monitoring data is extremely sparse (limited sensors, scarce maintenance windows). This interplay of characteristics means that no single methodology—pure mechanistic models, pure data models, or a simple combination of both—can simultaneously meet the demands for prediction accuracy, computational efficiency, and physical interpretability.

[0016] To achieve high-fidelity prediction of the properties of materials used in island and reef engineering, this invention includes three core concepts that together form a bridge connecting macroscopic environmental input, microscopic structural evolution, and microscopic performance output, as detailed below: Microenvironment activity index: Corrosion degradation on the surface of island and reef engineering materials is concentrated in specific active micro-regions, whose activity is controlled by the multiple couplings of humidity accumulation, chloride ion concentration, and previous damage history. Traditional parameters can only describe a single dimension and cannot comprehensively reflect the complex competition between the thermodynamic tendency and kinetic barriers of localized corrosion. The microenvironment activity index, as a dimensionless scalar field, is defined as a comprehensive measure of the material's comprehensive potential for significant performance degradation at a specific spatial location. In the data stream, the microenvironment activity index acts as a "modulation signal" that transmits information from the macro-environment to the meso-structure—driving the physical constraint generation model to adjust the aggregate generation strategy. Its historical sequence also constitutes part of the reinforcement learning framework state, used to learn the long-term dependence of tidal phase and material response.

[0017] Hysteresis Memory State: The response of island and reef materials in tidal environments exhibits significant memory effects and frequency dependence. Under the same boundary conditions, different historical paths induce different internal transport rates, and the material's response characteristics differ between diurnal and semi-diurnal tides. Traditional state variables cannot characterize this hysteresis behavior and multi-scale periodic effects related to historical paths. Hysteresis memory states are structured collections composed of multiple hysteresis memory units, each corresponding to a specific tidal frequency. In the data stream, hysteresis memory states act as "filters" for transmitting information from environmental history to the current state—receiving historical tidal data, refining the most predictive features of current performance through structured editing, and aggregating to generate an updated microenvironment activity index, thus achieving the "memory" of key extreme events and the "forgetting" of redundant fluctuations.

[0018] Virtual Aggregate Field: The essential determinant of material properties is its microstructure morphology. However, the actual aggregate distribution in island and reef engineering is highly random and difficult to observe across the entire field. Traditional representative volume element approximations are severely distorted in tidal environments with significant pumping effects. The virtual aggregate field is a parameterized microstructure generation result, a probabilistic sampling of the most likely structural morphology that satisfies physical consistency constraints under a given microenvironment activity index. In the data stream, the virtual aggregate field acts as a carrier for transmitting information from the microstructure to the macroscopic performance—receiving the conditional constraints of the microenvironment activity index. It feeds back the transmission characteristics of the current structure by calculating the pumping dynamic response function, and its key physical features (mean porosity, characteristic transmission path length, response time scale) inversely update the state representation of the reinforcement learning framework.

[0019] The environmental input layer drives the updating of the hysteresis memory state and the microenvironment activity index; the state decision layer optimizes the hysteresis memory state and generates the microenvironment activity index through reinforcement learning; the structure generation layer constructs a virtual aggregate field and calculates its pumping response based on the microenvironment activity index; the feedback update layer feeds back the characteristics of the virtual aggregate field to the state decision layer and adjusts the permeability tensor of the physical calculation layer with the spatial change of the microenvironment activity index; finally, the consensus verification layer confirms the physical consistency of the pairing between the microenvironment activity index and the virtual aggregate field, forming a closed-loop evolutionary cognition of environmental stimulus, historical memory, and structural response.

[0020] According to one aspect of the embodiments of this application, such as Figure 1 As shown, this application provides a method for simulating the performance of materials used in island and reef engineering, including: S1, acquire historical tidal data and on-site monitoring data of island and reef engineering structures; S2, initialize and maintain the shared state space, whose characterization parameters are the microenvironment activity index, hysteresis memory state, and virtual aggregate field; S3, based on historical tidal data and on-site monitoring data, uses a reinforcement learning framework to structurally edit the hysteresis memory state in order to update the microenvironment activity index; In some embodiments, structured editing is visualized in this application as three types of operations on a set of memory cells: The first type is threshold drift, which numerically shifts the flip threshold of selected cells based on the current tidal acceleration, making the cells more sensitive or less sensitive to rapidly changing environmental stimuli. The second type is region reset, which forcibly resets the state of all cells within a preset spatial region to simulate the erosion effect of extreme events (such as storm surges) on historical memory. The third type is cell proliferation, which adds new cells with new flip thresholds to the existing set to adapt to new frequency components revealed by monitoring data.

[0021] The decision-making for editing operations is executed by a reinforcement learning agent. Its state input includes the historical sequence of the activity index, the porosity and connectivity distribution of the aggregate field, and the current state set of all memory units. Its action output is a combination of the above three types of editing operations. Its reward signal is composed of the difference between the simulated and measured electrochemical impedance spectra, the consistency of the output of the physical constraint generation model, and the synchronicity of the system activity change and the tidal drive.

[0022] S4. The updated microenvironment activity index is used as a condition and input into the physical constraint generation model to update the virtual aggregate field and calculate its corresponding pumping dynamics response. In some embodiments, the physical constraint generation model in this application adopts a Generative Flow Network (GFlowNet) architecture, which is a well-known technology and will not be described in detail here. Its working principle in this application is as follows: through a series of reversible transformation steps, a simple prior distribution (such as Gaussian noise) is mapped to a complex virtual aggregate field distribution, and the physical constraints of Darcy's law and convection-diffusion equations are embedded in each transformation step.

[0023] The specific generation process includes: using the updated microenvironment activity index and spatial coordinates as input conditions, in the sequential generation step, based on the pressure field gradient direction and structural growth history within the currently partially generated aggregate structure, selecting the next generation action from a set of actions including pore growth, region closure, branching, and merging. After each generation action is executed, a differentiable physics calculation layer is invoked to perform a single-step fluid pressure and ion concentration field simulation of the current structure based on the pressure Poisson equation and convection-diffusion equation described in the claims. The physics field residuals and the estimated steady-state transport flux jointly adjust the selection strategy for subsequent generation actions until a complete virtual aggregate field is generated, and its pumping kinetic response function is output synchronously.

[0024] S5, based on the pump suction dynamics response and the updated microenvironment activity index, combined with tidal-driven boundary conditions, updates the shared state space; The specific calculation process of S5 is as follows: First, the pumping dynamic response function output in step S4 is analyzed into an effective transport coefficient. Combined with the boundary conditions (surface concentration, pressure) determined by the current tidal phase, the simplified steady-state transport equation recorded in the claims is solved to obtain the updated distribution of the chloride ion concentration field. Second, based on the coupling relationship between the concentration field and the humidity field, the new spatial distribution of the microenvironment activity index is calculated. Finally, the updated activity index, the current hysteresis memory state (updated in step S3), and the newly generated virtual aggregate field and its pumping response function are jointly written into the next time step snapshot of the shared state space.

[0025] In its implementation, the shared state space is a high-performance, memory-oriented data container organized using a hierarchical dictionary structure. The top layer is a main dictionary with a time index, where each time point corresponds to a state snapshot. Each snapshot contains three parallel data members, corresponding to the microenvironment activity index, the hysteresis memory state, and the virtual aggregate field. The relationship between these three is achieved through a shared spatial coordinate index: for any spatial location of the engineering structure, the activity index at the current moment, the set of hysteresis memory units associated with that location, and the structural features of the corresponding region in the virtual aggregate field can be accessed simultaneously using the same coordinate key.

[0026] This organizational approach enables reinforcement learning frameworks and physical constraint generation models to exchange data through a unified interface. For example, when a reinforcement learning framework needs to evaluate the structural state of a region to calculate a reward, it can directly extract the activity index and aggregate field characteristics of that region through coordinate indexing, without the need for complex data transformation.

[0027] The initialization of the shared state space follows a hierarchical strategy from empirical observation to deduction.

[0028] The initial field of the microenvironment activity index is constructed based on the first field monitoring data. If temperature, humidity and electrochemical sensors are deployed on-site, the monitored values ​​are normalized and mapped to the activity index, and inverse distance weighted interpolation is used to fill in areas without sensor coverage. If there is no monitoring data, a uniform initial value characterizing the undamaged matrix state is assigned based on the material mix ratio (water-cement ratio, amount of cementitious material) and the island and reef environmental benchmark parameters.

[0029] The initial set of memory cells for the hysteresis memory state is constructed based on the spectral analysis results of historical tidal data. The system first identifies the dominant tidal components of the target sea area (usually diurnal and semi-diurnal tides) and assigns a corresponding number of memory cells to each frequency component. For cells corresponding to diurnal tides, the flip threshold interval is set to a smaller range to capture long-term trends; for cells corresponding to semi-diurnal tides, the flip threshold interval is set to a larger range to respond to short-term fluctuations. The initial state of each cell is set to unflipped.

[0030] The initial structure of the virtual aggregate yard is generated using a constrained random placement algorithm, which is a well-known technology and will not be described in detail here. Its application in this embodiment is as follows: using the aggregate particle size distribution curve as a constraint, ellipsoidal aggregate particles are randomly placed within the simulation domain to ensure that the aggregate volume fraction meets the mix proportion requirements and does not cause geometric overlap; after placement, the remaining space constitutes the initial pore network. Based on this, the system extracts the porosity distribution and connectivity matrix as the initial topological description of the aggregate yard.

[0031] In the simulation loop, the shared state space is updated using an incremental strategy to ensure computational efficiency and physical continuity.

[0032] In step S3, the reinforcement learning framework performs structured editing on the hysteresis memory state, operating on the parameters (position of the flip threshold) and topology (addition and deletion of units) of the memory cells. The updated hysteresis memory state generates a new microenvironment activity index through aggregation operations. This process covers the activity index field in the updated state space, but retains the virtual aggregate field unchanged.

[0033] In step S4, the physical constraint generation model, conditioned on the updated activity index, reconstructs the pore distribution of the virtual aggregate field through a sequential generation process. The generation process employs an overlay update: the new aggregate field completely replaces the old structure, but retains a summary of the statistical characteristics of the old field (such as mean porosity and characteristic path length) for subsequent verification. The synchronously calculated pumping kinetic response function is written as an auxiliary attribute into the aggregate field data structure.

[0034] In step S5, the updated pumping dynamics response is combined with the tidal-driven boundary conditions to solve the transport equations to correct the spatial distribution of the activity index. The corrected activity index, the current hysteresis memory state, and the newly generated aggregate field are then written into the state snapshot of the next time step. The entire process ensures data consistency through a transaction mechanism: any failure in a sub-step triggers a rollback to prevent logical breaks in the state space.

[0035] S6, based on the updated shared state space, outputs the predicted material properties of the island and reef engineering structures.

[0036] Based on the updated shared state space, the output material property prediction results contain three levels: The first level is the field distribution output, including the spatiotemporal evolution field of chloride ion concentration on the steel bar surface, the three-dimensional distribution cloud map of the activity index, and the pore solution saturation field, used to locate high-risk corrosion areas; the second level is the prediction of key time nodes, including the probability distribution of the steel bar depassivation time, the expected time to reach the critical chloride ion concentration, and the remaining service life range corresponding to different reliability levels; the third level is the structural evolution early warning, which triggers a microstructural deterioration early warning when the connectivity index of the virtual aggregate field exceeds the threshold or the pumping response shows irreversible growth, prompting targeted testing.

[0037] It should be noted that the microenvironment activity index is a dimensionless scalar field defined on a three-dimensional spatial grid. Its physical meaning is a comprehensive measure of the potential for significant performance degradation (such as rebar depassivation and interface peeling) at a specific location. This index comprehensively reflects the coupling state of ion activity, humidity saturation, and electrochemical reaction rate in the local pore solution. The value range is defined by the two ends of a closed interval, with one end representing a completely inert, undamaged matrix and the other end representing a highly active critical corrosion state. The numerical evolution of the index is controlled by two driving forces: first, the immediate excitation of environmental boundary conditions (tidal phase, surface concentration); and second, the modulation effect of the material's internal transport characteristics (pore connectivity determined by the virtual aggregate field).

[0038] The theoretical basis of the hysteresis memory state is borrowed from the mathematical framework of the Preisach hysteresis model, which is a well-known technology and will not be elaborated upon in this embodiment. Its application in this application is as follows: the hysteresis operator in the Preisach plane is abstracted into a bi-state memory unit with a flip threshold, each unit corresponding to a specific tidal frequency component. The basic structure of the memory unit includes two core parameters: one is the flip threshold pair (defined by a lower and upper limit), which determines the boundary conditions for unit state switching; the other is the current state flag (flipped or not flipped), characterizing whether the historical load of this frequency component has ever reached the excitation level. The collection of multiple units constitutes a multi-frequency decomposition representation of tidal history, enabling the system to simultaneously track the differentiated effects of long-term trends and short-term fluctuations on the material state.

[0039] The virtual aggregate field characterizes the spatial distribution and topological properties of the material's internal microstructure, comprising three levels: First, the spatial occupancy of pores and solids (binary segmentation) identifies any location as belonging to aggregate, cement paste, or pores; second, the connectivity matrix of the pore network describes the throat connections and radius distribution between pores; and finally, the transport characteristic parameters of each local region, calculated in real-time by the physically constrained generation model through an embedded differentiable physics computation layer during the generation process. Its data representation employs a hybrid structure of three-dimensional voxel meshes and a graph-based pore network model: the voxel mesh is used for spatial occupancy and physics calculations, while the pore network model is used for topology analysis and transport path search. A one-to-one correspondence is established between the two representations through coordinate mapping, ensuring efficient switching between spatial querying and network analysis.

[0040] The pumping kinetics response describes the kinetic characteristics of fluid and dissolved ions transported through the internal pore network of a material under tidal pressure. Its physical essence stems from the capillary suction-drainage effect caused by wet-dry cycles: during high tide, the external solution is forced into the pores; during low tide, the internal fluid seeps out under pressure gradient and evaporation, forming a periodic convection-diffusion enhancement mechanism. This response is characterized by a parameterized function containing three exponentially decaying terms, corresponding to rapid transport dominated by macropores, medium-rate transport dominated by mesopores, and slow transport dominated by micropores. The weighting coefficients and characteristic time constants of each term are predicted by a neural network based on the pore distribution characteristics of a virtual aggregate field, enabling the response function to adapt to different aggregate morphologies and pore topologies.

[0041] In one specific implementation, such as Figure 2 The diagram illustrates the overall system architecture and data logic flowchart of the simulation method for island and reef engineering materials according to the present invention. It employs a collaborative simulation framework driven by the fusion of data and physical models. The core of this framework lies in maintaining a dynamically updated shared state space, which integrates indices characterizing the activity of the material's microenvironment, hysteresis memory states reflecting historical dependence, and a microscale virtual aggregate field.

[0042] like Figure 2 As shown, the simulation process begins with acquiring historical tidal data and sparse monitoring data (such as electrochemical impedance spectroscopy) from the engineering site. This data collectively initializes and continuously drives the core engine of the system. The system mainly consists of two processing engines: a reinforcement learning engine responsible for adaptively structuring and optimizing the hysteresis memory state based on environmental feedback; and a physical constraint generation engine (built based on a generative flow network) that uses the activity index output by the aforementioned engines as a condition to generate virtual aggregate microstructures that conform to physical laws and calculate their pumping kinetic response. The two engines exchange data in real time through a bidirectional coupling protocol to ensure logical consistency between structure generation and state evolution.

[0043] The updated state from the engine is input into the coupled transport solver, which calculates the evolution of the chloride ion concentration field and humidity field based on the material transport equation. The results are fed back to the shared state space, forming a closed loop. The system uses the tidal cycle as its rhythm, initiating periodic consensus verification at designated nodes. This mechanism performs physical consistency checks on the outputs of the two engines. If the verification passes, the system outputs the final material performance prediction (such as the service life curve); if it fails, parameter adjustments are triggered and a new round of iterations is started until a self-consistent solution is achieved. This architecture realizes a fully integrated simulation across the entire chain, from environment-driven simulation, microstructure generation, multiphysics evolution to service life prediction.

[0044] Furthermore, S3 utilizes a reinforcement learning framework to perform structured editing of delayed memory states, including: S31, construct multiple hysteresis memory units corresponding to the frequencies of the main tidal components, wherein the flipping threshold spacing of the hysteresis memory units corresponding to the diurnal tidal component is set to a first range, and the flipping threshold spacing of the hysteresis memory units corresponding to the semi-diurnal tidal component is set to a second range greater than the first range. In this embodiment, the hysteresis memory state is specifically implemented as a hierarchical set of memory units, which are constructed based on the results of tidal spectrum analysis and configured differently.

[0045] The system first performs harmonic analysis on the historical tidal data of the target sea area to identify the dominant tidal components. For typical tropical island and reef environments, two main types of tidal components are usually identified: diurnal tidal components with a period of approximately 24 hours (such as K1 and O1 tidal components), and semi-diurnal tidal components with a period of approximately 12.4 hours (such as M2 and S2 tidal components). The system is configured with memory units of different parameters for different frequency characteristics.

[0046] For memory cells corresponding to diurnal tides, the flip-threshold spacing is set to a first range, specifically between 0.1 and 0.3 units of dimensionless pressure. This smaller spacing allows the cells to remain sensitive to slowly changing environmental stimuli, capturing the cumulative effects of the diurnal cycle. For memory cells corresponding to semi-diurnal tides, the flip-threshold spacing is set to a second range, specifically between 0.5 and 1.2 units of dimensionless pressure. This larger spacing ensures that the cell state does not flip frequently due to high-frequency fluctuations, maintaining relative stability while retaining the ability to respond to significant stimuli. This differentiated configuration allows the system to track environmental history at different timescales in parallel.

[0047] S32, define the state of the reinforcement learning framework. The state of the reinforcement learning framework includes: the historical sequence of the microenvironment activity index, the porosity and connectivity distribution of the virtual aggregate field, and the current state set of all hysteresis memory units. The state of the reinforcement learning framework is composed of three components. The first component is the historical sequence of the microenvironment activity index, with a time window length set to the longer of the last 10 tidal cycles or 100 simulation steps to ensure coverage of at least one complete diurnal tidal cycle. This sequence is extracted at a fixed sampling rate to form a one-dimensional time series vector.

[0048] The second component represents the porosity and connectivity distribution of the virtual aggregate field, specifically including a local porosity histogram (dividing the porosity range into 10 intervals for statistical frequency) and three eigenvalues ​​of the permeability tensor (characterizing the degree of anisotropy). These indicators are obtained by spatially partitioning the aggregate field, with the partition size determined based on the maximum aggregate particle size, typically 5 to 10 times the particle size.

[0049] The third component is the set of current states of all delayed memory units, with each unit contributing a binary state value (flipped or unflipped), forming a sparse binary vector.

[0050] S33, defines the actions of the reinforcement learning framework, which include performing at least one of the following editing operations on all delayed memory units: The flip threshold of the selected cell is numerically drifted according to the current tidal acceleration; the state of the cells in the preset spatial area is reset; a new cell with the new flip threshold is added to all hysteresis memory cells; The action space contains three types of editing operations, and the specific implementation logic for each type of operation is as follows: The first type involves numerically drifting the flipping threshold of selected cells based on the current tidal acceleration. The selection logic employs a frequency matching algorithm, which includes calculating the Euclidean distance between the nominal frequency of each memory cell and the dominant tidal frequency at the current moment, and selecting the top 20% of cells with the smallest distance as the selected set. The magnitude of the numerical drift is proportional to the tidal acceleration; the greater the acceleration, the greater the drift, making the cells more sensitive to rapidly changing environments.

[0051] The second category involves resetting the state of units within a preset spatial region. The preset spatial region is defined using a combination of static and dynamic methods: the static region is pre-divided according to the structural design drawings, including the splash zone, water level fluctuation zone, and underwater zone; the dynamic region is automatically identified during the simulation process based on the spatial variance of the microenvironment activity index. When the variance of the activity index of a certain region exceeds a preset threshold, the region is marked as an abnormal region and included in the target range of the reset operation.

[0052] The third category involves adding new cells with new flip thresholds to all hysteresis memory cells. When the existing cell set is unable to explain newly emerging monitoring data patterns, new cells are inserted into the gaps within the existing threshold range, with their initial thresholds set to the median of the thresholds of adjacent cells.

[0053] The direction of drift is determined by the direction of the mechanical action represented by the current tidal acceleration, and the magnitude of drift is proportional to the absolute value of the tidal acceleration.

[0054] S34, define the reward function of the reinforcement learning framework, which is a weighted sum of at least two of the following sub-rewards: a first sub-reward based on the difference between simulated and measured electrochemical impedance spectra, a second sub-reward based on the consistency of the output of the physical constraint generative model under predicted and observed conditions, and a third sub-reward based on the synchronicity of system activity changes and tidal driving. The reward function consists of a weighted average of three sub-rewards, and the specific calculation method for each sub-reward is as follows: The first sub-reward is calculated based on the difference between simulated and measured electrochemical impedance spectroscopy (EIS). This difference is measured using a weighted mean square error (MSE). The weighting function is designed so that during high tide, impedance errors corresponding to charge transfer and diffusion processes receive higher weights; while during low tide, impedance errors corresponding to the electrode-solution double-layer processes receive higher weights. This phase-dependent weighting mechanism ensures that the system focuses on different electrochemical processes under different tidal conditions.

[0055] The second sub-reward is calculated based on the difference in consistency between the output of the physical constraint generation model under predicted and observed conditions. Specifically, the physical constraint generation model is driven to generate a virtual aggregate field under conditions based on both the predicted activity index and the activity index derived from measured data. The difference in the statistical distribution of the aggregate field under the two conditions is calculated, and the negative value of this difference is used as the sub-reward. The distribution difference is measured using either the Kuhlbek-Leibler divergence or the Jensen-Shannon divergence, both of which are well-known techniques and will not be elaborated upon in this embodiment.

[0056] The third sub-reward is calculated based on the synchronicity between changes in system activity and tidal forces. Specifically, it calculates the correlation coefficient between the rate of change of the microenvironment activity index and the rate of change of tidal level over the entire time window; the higher the correlation coefficient, the greater the reward. This sub-reward incentivizes the system to learn the phase-locked relationship between the evolution of the activity index and the tidal cycle.

[0057] S35, based on the state, action, and reward function of the reinforcement learning framework, executes reinforcement learning decisions, updates the delayed memory state, and aggregates to obtain the updated microenvironment activity index.

[0058] This embodiment uses a reinforcement learning decision-maker. This algorithm is a well-known technology and will not be described in detail here. Its key hyperparameters are set as follows: the learning rate is set to 0.003%, and the discount factor is set to 0.99 to balance immediate rewards and long-term returns. Both the policy network and the value network adopt a multilayer perceptron architecture, with the input layer dimension matching the state vector dimension and the output layer dimension matching the action space dimension.

[0059] The specific process of generating a new microenvironment activity index from the updated hysteresis memory states is as follows: The state vector of the memory unit is input into a fully connected neural network mapper, which contains two hidden layers, each with 64 neurons, and the output layer dimension is equal to the number of spatial grid nodes. The mapper is trained using supervised learning, with training samples consisting of pairs of memory states and corresponding activity indices from historical data. During inference, the current memory state is converted into a spatial distribution field by the mapper and then fused with boundary condition constraints to obtain the updated microenvironment activity index.

[0060] Based on the embodiments provided in this application, by constructing multiple hysteresis memory units directly corresponding to the frequencies of major tidal components such as diurnal and semi-diurnal tides, and setting differentiated flip threshold intervals, the simulation system can physically distinguish and memorize the historical effects of tidal loads of different periods. A state space is defined, containing the historical sequence of the microenvironment activity index, the virtual aggregate field state, and the states of all memory units. An action space is also defined, containing editing actions such as numerical drifting of unit thresholds, state reset, and adding new units. Together, these constitute a complete and operable reinforcement learning environment. This transforms the editing of hysteresis memory states into a learnable and optimizable process, ensuring that the update of the microenvironment activity index is based on a dynamic memory mechanism that clearly maps to the physical process, significantly improving the modeling accuracy of complex tidal historical impacts.

[0061] Furthermore, the calculation of the first sub-reward in S34 includes: Calculate the weighted mean square error of simulated and measured electrochemical impedance spectroscopy in logarithmic coordinates, where the weighting function is... ; The reference weight constant is dimensionless. The modulation amplitude is dimensionless. For normalized tidal phase, This is the impedance frequency, with dimensions of the reciprocal of time. pass The weighted mean square error is calculated so that the low-frequency impedance error corresponding to the charge transfer and diffusion process is given a higher weight when simulating the full tide phase, and the high-frequency impedance error corresponding to the double layer process at the electrode / solution interface is given a higher weight when simulating the dry tide phase.

[0062] The weighting function is designed based on the physical characteristics of electrochemical impedance spectroscopy. In the low-frequency range (e.g., on the order of one hundredth of a Hz), the impedance response mainly reflects the charge transfer process on the steel surface and the chloride ion diffusion process in the pores. During the full tidal phase, the material is saturated, the pore solution has good connectivity, and the charge transfer and diffusion processes dominate the corrosion kinetics; therefore, the error in this frequency range is given a higher weight.

[0063] At high frequencies (e.g., in the 100 kHz range), the impedance response primarily reflects the double-layer capacitance effect at the electrode-solution interface. During the dry / tidal phase, the material surface is exposed to the atmosphere, and the double-layer structure changes significantly, giving higher weight to errors in this frequency range.

[0064] In practice, the baseline weight constant in the weighting function is set to a dimensionless value of 1.0, and the modulation amplitude is set to a value within the range of 0.2 to 0.8. The specific values ​​are determined through grid search on the validation set, with the goal of maximizing the prediction accuracy on the validation set. The phase shift in the modulation term ensures that the weights of different frequency bands are enhanced during high and low tides, respectively.

[0065] Based on the embodiments provided in this application, a weighting function dynamically coupled with the tidal phase is employed. The reward is calculated based on the difference between simulated and measured electrochemical impedance spectra, allowing the reward mechanism to change synchronously with the simulated tidal environment. During the simulation of high tide, the reward function automatically emphasizes low-frequency impedance errors reflecting charge transfer and diffusion processes within the material; during the simulation of low tide, it emphasizes high-frequency impedance errors reflecting interfacial processes on the material surface. This design encodes the physical fact that the material's main conductive chemical response mechanism differs under different tidal states (immersion or exposure) into the learning objective, thereby guiding the microenvironment state output by the reinforcement learning process to accurately match the key physicochemical response characteristics of each environmental stage, improving the physical consistency and accuracy of the simulation results throughout the entire tidal cycle.

[0066] Furthermore, the calculation of the second sub-reward in S34 includes: The physical constraint generation model is driven by the first microenvironment activity index based on prediction and the second microenvironment activity index based on measured data, respectively. The difference in statistical distribution of the virtual aggregate field generated under the two conditions is calculated, and the negative value of the difference is used as the second sub-reward. The difference is either the Kolbec-Leibler divergence or the Jansen-Shannon divergence.

[0067] In some embodiments, the specific method for inverting the second microenvironment activity index based on measured data is as follows: A physical inverse problem solver is established, using the measured chloride ion concentration profile and humidity distribution as input, and an iterative optimization algorithm is used to minimize the difference between the predicted and measured data, thereby inverting the activity index field. This inversion process is a well-known technique and will not be elaborated upon in this embodiment.

[0068] The calculation of the statistical distribution difference between the virtual aggregate field generated under predicted and observed conditions specifically targets the one-dimensional pore size distribution and the joint probability distribution of two-dimensional pore size and throat radius. Taking the Kolb-Leibler divergence as an example, it is calculated as the sum of the relative entropies of the predicted and observed distributions in each discrete interval, measuring the degree of difference between the two probability distributions. The greater the difference, the worse the consistency of the output of the generative model under the two conditions, and the smaller the corresponding sub-reward.

[0069] Based on the embodiments provided in this application, a strong physical credibility constraint is established between the reinforcement learning process and the physical generation process by setting up a second sub-reward based on the consistency of the model output under physical constraints. This reward is constructed by comparing the statistical distribution differences between virtual aggregate fields generated by the same physical generation model, driven by the predicted microenvironment activity index and the microenvironment activity index derived from measured data, respectively. This compels the reinforcement learning agent to output a microenvironment activity index that can not only fit the data but also translate reasonable microstructures through the physical model, ensuring that the predicted state parameters have clear physical generation meaning and structural evolution interpretation capabilities, thereby enhancing the physical interpretability and intrinsic consistency of the entire simulation chain.

[0070] Furthermore, the physical constraint generation model in S4 is a generative flow network. The process of updating the virtual aggregate field and calculating the pump suction dynamic response by the physical constraint generation model includes: S41, with the updated microenvironment activity index and spatial coordinates as conditional inputs; S42, Execute the serialization generation step: In each step, based on the pressure field gradient direction inside the aggregate structure generated in the current part and the structure growth history, the Boltzmann distribution is used to select the next generation action from the action set including pore growth, region closure, branching and merging. The action score of the generation action is determined by the pressure field gradient alignment degree and the historical growth continuity. In voxel mesh characterization, the pore growth operation transforms the selected solid voxel state into a pore voxel; the region closure operation identifies isolated pore clusters and marks them as closed pores, which do not participate in transmission; the branch operation creates a new channel direction at the end of an existing pore channel; and the merge operation connects two adjacent isolated pore clusters to form a connected transmission path.

[0071] The action score is determined jointly by the pressure field gradient alignment and the historical growth continuity. The alignment score is calculated as the cosine of the angle between the proposed growth direction and the local pressure gradient direction; the smaller the angle, the higher the score. The historical growth continuity score is calculated as the similarity between the current proposed direction and the previous growth direction; maintaining directional consistency yields a higher score. The two scores are weighted and summed to obtain the final action score.

[0072] Action selection follows a Boltzmann distribution, with the initial temperature parameter set to 1.0. Higher temperatures result in more random selection and greater exploratory power; lower temperatures lead to more deterministic selection and a preference for high-scoring actions. In scenarios where the sampling temperature is increased, the new temperature is set to 1.5 times the current temperature to enhance exploratory capabilities and find new solutions that satisfy physical consistency.

[0073] S43, after each generation action is executed, calls the differentiable physics calculation layer to perform a single-step fluid pressure and ion concentration field simulation of the current structure based on Darcy's law and the convection-diffusion equation; S44, based on the physical field residuals output by the differentiable physical computation layer, adjust the selection strategy for subsequent generation actions; S44 specifically includes: physical field residuals based on the output of the differentiable physical computing layer, and a neural network for evaluating future transmission potential, adjusting the selection strategy for subsequent generation actions; The neural network takes the graph embedding of the current pore network topology as input and the estimated steady-state transmission flux as output.

[0074] The specific method for integrating spatial coordinates as conditional input is as follows: the three-dimensional coordinates are sinusoidally encoded, and the encoded result is concatenated with the microenvironment activity index to form a conditional input vector. This vector serves as a global condition in each step of the generative flow network, guiding the local generation actions.

[0075] In some embodiments, the neural network used to assess future transmission potential employs a graph neural network architecture, which is a well-known technology and will not be elaborated upon here. Its specific configuration in this application is as follows: the input is a graph embedding of the current porosity network topology; topological features are extracted through three graph convolutional layers; and the final output is the estimated steady-state transmission flux. Network training employs supervised learning, and training data is generated through high-fidelity Darcy flow and diffusion coupling simulations, including pairing porosity network graph embeddings with corresponding steady-state transmission fluxes. During the generation process, the graph embedding of the current partial structure is input into the network to obtain the estimated flux; the difference between this estimated flux and the target flux serves as an additional reward signal to adjust the probability distribution of action selection.

[0076] S45, repeat S42 to S44 to generate a complete virtual aggregate field and simultaneously output a parameterized response function to characterize its pumping dynamics. ; in, These are the weight coefficients obtained through training, with the dimension being the square of length per time. For time variables, The characteristic time constant with the dimension of time, Corresponding to the high-speed transmission dominated by large apertures, Corresponding to slow transmission dominated by micropores, This corresponds to medium-speed transmission dominated by the central aperture.

[0077] It needs to be explained that the pump suction dynamics response function adopts a three-exponential term form to characterize the non-single exponential decay behavior under the synergistic effect of multi-scale pores inside the coral aggregate.

[0078] in, These are weighting coefficients, with dimensions [L]. 2 T -1 Physically, this can be interpreted as the contribution rate of the corresponding pore size to the total pump flux.

[0079] The characteristic time constant, in physical terms, represents the characteristic time required for solution exchange or ion diffusion to reach dynamic equilibrium within pores of different scales. Based on existing pore classification standards, this invention associates i=1, 2, 3 with macropores, mesopores, and micropores, respectively. For macropores (pore size > 50 nm): the pore size is large, the fluid flow resistance is low, ion transport is mainly by convection, and the transport rate is fast. When the value is small (e.g., on the order of 1-10 hours), it exhibits a rapid decay.

[0080] For mesopores (pore size 2-50 nm): the pore size is moderate, capillary action is significant, ion transport is influenced by both diffusion and capillary flow, and the rate is moderate. Values ​​in the middle range (e.g., on the order of 10-100 hours) are characterized by a medium-speed decay term.

[0081] For micropores (pore size < 2 nm): the pore size is extremely small, the surface adsorption effect is strong, ions mainly diffuse and the path is tortuous, resulting in a slow transport rate. Larger values ​​(e.g., on the order of 100-500 hours) indicate a slow decay term.

[0082] This classification reflects the fundamental physical principle that aperture size determines the transmission mechanism and rate, enabling the pump suction dynamic response function to mechanistically describe the dynamic process of aggregate absorbing and releasing water during tidal cycles.

[0083] Based on the embodiments provided in this application, the physical constraint generation model is specifically implemented as a generative flow network that integrates sequential decision-making and online physical simulation. This network uses the microenvironment activity index and spatial coordinates as conditions, and progressively constructs a virtual aggregate field through a series of sequential actions (such as pore growth and branching) guided by the current internal pressure field gradient and growth history. After each generation step, a differentiable physical computation layer is invoked to perform single-step fluid and ion transport simulations, utilizing real-time feedback from the physical field residuals to adjust subsequent generation strategies. This ensures that the generated virtual aggregate field satisfies basic fluid dynamics and transport constraints in its microscopic topology. Finally, a parameterized pumping dynamics response function is output synchronously. Its multi-exponential form can naturally characterize the transport process dominated by pores of different scales and with multiple characteristic time constants, realizing the joint generation and analytical characterization of the material microstructure and its transport dynamics under physical constraints.

[0084] Furthermore, the single-step simulation of the differentiable physics computation layer in S43 includes: Using the current aggregate porosity field as the transmission coefficient, the pressure field is updated by solving the following pressure Poisson equation in a single iteration. in, For divergence operators; Porosity depends on permeability in units of length squared. The fluid dynamic viscosity is expressed in units of mass per length per time. Pressure, measured in units of mass per length per square time. The external mass injection rate is represented by the volume source term with the dimension of the reciprocal of time, and is set to 0 when there is no external injection. The obtained pressure gradient is then used to update the ion concentration field through one explicit time integration step. The gradient is kept backpropagated to the parameters of the generating flow network throughout the calculation process.

[0085] The porosity-dependent permeability in the pressure Poisson equation is calculated using the Kozeny-Carman equation, which is a well-known technique and will not be elaborated upon in this embodiment. The shape factor in the equation is set to 5, and the specific surface area is estimated based on the aggregate particle size distribution. The hydrodynamic viscosity is taken as the dynamic viscosity of seawater at the service temperature, and the volume source term is set to 0 when there is no external mass injection.

[0086] Differentiability is achieved using the automatic differentiation function of the PyTorch deep learning framework, a well-known technology that will not be elaborated upon in this embodiment. Specifically, the solution to the pressure Poisson equation is encapsulated as a custom computation module, internally employing a differentiable conjugate gradient method for iterative solution, ensuring that the gradient of the pressure field on the porosity field can propagate backwards. The ion concentration field is updated using explicit time integration, also maintaining computational graph connectivity, allowing the gradient of the entire physical calculation process to propagate backwards to the parameters of the generating flow network.

[0087] Based on the embodiments provided in this application, a differentiable core solver is provided for the physical simulation of the generation process. This solver updates the pressure field by solving the pressure Poisson equation using the current aggregate porosity field as the conduction coefficient, and subsequently updates the ion concentration field. Crucially, the entire physical calculation process is designed so that gradients can propagate back to the parameters of the generating flow network. This allows the generating model to not only adjust based on the results of the physical simulation, but also to directly perceive the intensity and direction of the influence of structural changes on the physical field through gradient signals. This enables physical laws to be efficiently embedded in and guided in the structural generation process in a continuous and differentiable manner. This gradient-based soft physical constraint, compared to hard rules or post-hoc verification, more naturally and accurately ensures that the generated microstructure conforms to real physical behavior.

[0088] Furthermore, a first coupling step is included between S3 and S4: The updated microenvironment activity index and target area information are sent to the physical constraint generation model as its condition input. The system receives key physical features of the virtual aggregate field generated based on the physical constraint generation model and uses them to update the state representation of the reinforcement learning framework. The key physical features include the average aggregate porosity, the feature transmission path length, and the pumping response time scale.

[0089] The calculation and feedback methods for key physical characteristics are as follows: The average aggregate porosity was obtained by statistically averaging all voxels of the virtual aggregate field. The characteristic transport path length was calculated using the maximum flow and minimum cut algorithm: the pore network was abstracted as a graph, with the exposed surface and the rebar surface as the source and sink points, respectively. The equivalent transport resistance corresponding to the minimum cut set was solved, and its reciprocal is the measure of the characteristic transport path length. The pump suction response time scale was extracted from the response function, taking the median value among the three characteristic time constants (corresponding to medium-rate transport dominated by medium pores). This value characterizes the overall transport dynamics of the aggregate field.

[0090] When these feature vectors are fed back into the reinforcement learning framework, they are concatenated with the original state vectors as additional state channels, expanding the state dimension and enabling the agent to perceive the impact of aggregate field structure on decision-making.

[0091] Based on the embodiments provided in this application, a reverse feature feedback from the physical generation model to the reinforcement learning framework is achieved by establishing a first coupling step between the reinforcement learning framework and the physical constraint generation model. When the physical constraint generation model generates a virtual aggregate field based on the updated microenvironment activity index, it feeds back the key physical features of the aggregate field (such as average porosity and feature transmission path length) to the reinforcement learning framework to update its state representation. This allows the reinforcement learning agent to consider the downstream physical structural consequences of its previous decisions as higher-order state information when making decisions, thereby learning the complex mapping relationship between the microenvironment state and the physical properties of the microstructure. This bidirectional information flow enhances the collaboration between the two core modules, making the overall simulation behavior of the system more coordinated and consistent, and avoiding target deviations that may result from isolated module optimization.

[0092] Furthermore, a second coupling step is included between S3 and S4: Calculate the spatial change of the microenvironment activity index before and after the update; Send the spatial variation to the physical constraint generation model; The physical constraint generation model dynamically adjusts the main diagonal component of the Darcy permeability tensor in its internal differentiable physical calculation layer based on the spatial variation, so as to enhance the pumping intensity in the region with increased microenvironment activity index and weaken the pumping intensity in the region with decreased activity index. The adjustment range of the main diagonal component is proportional to the spatial variation.

[0093] The spatial variation is calculated as the difference between the microenvironment activity index at the current moment and the previous simulation step. This variation is sent to the physical constraint generation model and used to dynamically adjust the principal diagonal component of the Darcy permeability tensor in the differentiable physical computation layer. The adjustment method is as follows: the new permeability equals the original permeability multiplied by an adjustment coefficient, which is 1 plus the product of the gain coefficient and the spatial variation. The gain coefficient is set to a value in the range of 0.1 to 0.5. In regions where the microenvironment activity index increases (positive variation), the pumping intensity is enhanced; in regions where it decreases (negative variation), the pumping intensity is weakened. This dynamic adjustment allows the physical computation layer to respond to the spatial evolution of environmental activity.

[0094] Based on the embodiments provided in this application, a more direct and faster inter-module coupling mechanism is established by introducing a second coupling step. This step directly transmits the spatial variation of the microenvironment activity index to the physical constraint generation model, which then dynamically adjusts the principal diagonal components of the Darcy permeability tensor in its internal physical calculation layer accordingly. The adjustment logic is: in regions where the activity index increases, permeability is increased proportionally (i.e., pumping intensity is increased), and in regions where it decreases, permeability is weakened. This allows the changes in the microenvironment state evaluated by the upper layer to affect the parameters of the lower-level transport simulation in real time and in a physically intuitive way during each simulation cycle, realizing dynamic real-time coupling between the chemical / active environment and the physical transport environment, and significantly improving the sensitivity and realism of the system in simulating complex multi-field interactions within materials.

[0095] Furthermore, the method also includes: After every integer number of tidal cycles, perform the following iterations: With the current virtual aggregate field fixed, S3 and S5 are iteratively run. The termination condition is that the rate of change of the reward function of the reinforcement learning framework is less than the preset change threshold ε or the number of iterations reaches the maximum number of iterations N_max. The microenvironment activity index is optimized, where ε is the unit of reward per time and N_max is a dimensionless positive integer. With a fixed current microenvironment activity index, S4 is iteratively run to sample and generate a high-probability virtual aggregate field; The current paired microenvironment activity index and virtual aggregate field are validated. The validation includes solving the simplified steady-state transport equation under the current paired microenvironment activity index and virtual aggregate field conditions. Calculate whether its residual norm is less than the preset residual threshold δ; in, For divergence operators; The effective diffusion coefficient has the dimension of length squared per time. The chloride ion concentration is expressed in units of mass per volume. For gradient operators, express The gradient; The dimension of the pump suction source term is mass per volume per time, the dimension of the residual is mass per volume per time, and the dimension of δ is mass per volume per time. If the verification passes, the consensus state is output as the initial condition for the next prediction stage; otherwise, the search range of the flip threshold of the hysteresis memory unit is expanded, or the sampling temperature of the generating stream network is increased, and the above steps are repeated.

[0096] The preset change threshold ε is measured in units of reward per second, and its physical meaning is the convergence criterion for the rate of change of the reward function. Specifically, it is set as follows: when optimizing the microenvironment activity index, if the average absolute rate of change of the reward function is less than 0.001 rewards per second over 10 consecutive iterations, it is considered convergent. This threshold is determined based on one percent of the reward fluctuation range in previous experiments.

[0097] The maximum number of iterations is set to 100 to prevent infinite loops.

[0098] The preset residual threshold δ is measured in units of mass per volume per time, and its physical meaning is to verify the numerical error tolerance of the equation solution. The set value is 10 to the power of -6 moles per cubic meter per hour, determined based on numerical calculation error analysis and engineering accuracy requirements.

[0099] The effective diffusion coefficient in the verification equation is calculated using the equivalent diffusion coefficient of the current virtual aggregate field. This calculation employs homogenization theory, which is a well-known technique and will not be elaborated upon in this embodiment. The pumping source term is calculated using the currently paired pumping response function and the microenvironment activity index.

[0100] If the verification fails, two adjustments are made: first, the search range for the flip threshold of the hysteresis memory unit is expanded, increasing the selectable threshold range for each unit by 20 percent from the current baseline value; second, the sampling temperature of the generating stream network is increased, raising the temperature parameter from the current value to 1.5 times. After the adjustments are completed, the optimization and verification loop is re-executed until a consensus state that satisfies physical consistency is found.

[0101] Based on the embodiments provided in this application, a crucial stability calibration mechanism is offered for long-term dynamic simulation by establishing a periodic consensus step. This step is executed after every integer number of tidal cycles. It seeks a consensus state where the two are highly compatible by alternately fixing the virtual aggregate field to optimize the microenvironment activity index and sampling to generate a virtual aggregate field with the fixed microenvironment activity index. A simplified steady-state transmission equation is then used for rapid physical verification. This process simulates periodic calibration within the system, effectively correcting state drift that may result from sequential execution and error accumulation, forcing the system to return to a physically consistent stable point across the cyclic scale. If verification fails, the search is restarted by expanding the parameter search range.

[0102] In some embodiments, to characterize the active transport effect of coral aggregate under tidal wet-dry cycles in the transport equation, this invention introduces a pumping source term into the chloride ion transport equation. The physical meaning of this source term is the net rate at which the aggregate, acting as a distributed ion reservoir, releases or absorbs chloride ions from the surrounding slurry during water absorption and dehydration, driven by the humidity gradient between the pore water inside the aggregate and the external environment. Its dimension is mass per volume per time.

[0103] The value of the pumping source term is determined by the microenvironment activity index, the pumping kinetic response function, and tidal history data, specifically through a multi-step physical calculation process. First, the kinetic response function characterizing the pumping properties of a single aggregate is convolved with the tidal level change rate from the tidal history data to obtain the equivalent volumetric flow rate contribution of each virtual aggregate under specific tidal driving. Then, this volumetric flow rate contribution, the microenvironment activity index at the aggregate's location, and the local chloride ion concentration gradient at that location are multiplied to obtain a vector form of the pumping mass flux. Finally, the spatial divergence of this pumping mass flux is calculated, yielding the pumping source term that needs to be added to the right-hand side of the chloride ion transport equation. Through this series of steps, cross-scale coupling is achieved from macroscopic tidal driving to microscopic aggregate response and then to microscopic ion flux.

[0104] Meanwhile, the boundary condition for the humidity transport equation is set as the surface saturation of the structure. This saturation value is directly derived from the instantaneous tide level in historical tidal data. At the highest tide level, the surface saturation is set to 1.0, representing complete saturation; at the lowest tide level, the surface saturation is set to a partially saturated value based on local climate experience, for example, between 0.2 and 0.4. During the transition period between rising and falling tide levels, the surface saturation value is smoothly interpolated using a cosine function between these extreme values ​​to simulate real natural wetting and drying processes and avoid abrupt numerical changes.

[0105] In one specific implementation, such as Figure 3 As shown, Figure 3 The time-series diagram illustrates the rhythmic coupling between the external tidal environment and internal simulation calculations within a standard tidal cycle. This diagram demonstrates that the simulation method does not proceed in uniform time steps, but rather precisely synchronizes its key computational steps with the physical processes of the tidal phase.

[0106] like Figure 3 As shown, the system's computational activities are closely organized around four key tidal phases: low tide, mid-high tide, high tide, and mid-low tide. During the mid-high and mid-low tide periods, when tidal level changes are most drastic, the activity index of the system's internal microenvironment undergoes significant changes. The physical significance of this design lies in the inherent hysteresis effect of transport and chemical reactions within the material in response to tidal forces; these drastic changes occur after the moment of strongest driving force (i.e., the maximum rate of tidal level change).

[0107] Following changes in the activity index, the system performs reward calculations and model updates at appropriate phases (e.g., near high tide). This aims to utilize the stable or characteristic state exhibited by the system after a complete half-cycle of excitation to assess simulation accuracy and update model parameters, thereby ensuring the reliability of the learned signal. Simultaneously, the active period of the pumping source term covers the main tidal level change periods, ensuring the synchronization between microscopic physical processes and macroscopic environmental drivers. Crucially, high tide is set as a fixed trigger point for the system's periodic consensus verification. This point is chosen because the system state has undergone a complete tidal excitation, placing it at a representative checkpoint where global consistency verification and reconciliation are most efficient. The entire timing design embodies the core principle of environmental rhythm-driven computational timing, making the simulation process isomorphic to the evolution of the real physical world.

[0108] According to another aspect of the embodiments of this application, a simulation system for the performance of island and reef engineering materials is also provided. For example... Figure 4 As shown, the system includes: Data acquisition unit 41 is used to acquire historical tidal data and on-site monitoring data of island and reef engineering structures; Initialization unit 42 is used to initialize and maintain the shared state space, which represents the microenvironment activity index, hysteresis memory state and virtual aggregate field. The structured editing unit 43 is used to perform structured editing of the hysteresis memory state based on historical tidal data and on-site monitoring data using a reinforcement learning framework, so as to update the microenvironment activity index. The first update unit 44 is used to input the updated microenvironment activity index as a condition into the physical constraint generation model in order to update the virtual aggregate field and calculate its corresponding pumping dynamic response. The second update unit 45 is used to update the shared state space based on the pump suction dynamics response and the updated microenvironment activity index, combined with the tidal-driven boundary conditions. Output unit 46 is used to output the predicted material properties of the island and reef engineering structure based on the updated shared state space.

[0109] It should be noted that the embodiments implemented on the island and reef engineering material performance simulation system side in this application can be referenced with the embodiments implemented on the island and reef engineering material performance simulation method side, and will not be described in detail here.

[0110] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described simulation method for the properties of island and reef engineering materials is also provided. This electronic device may be... Figure 5 The terminal device or server shown. This embodiment uses this electronic device as an example of a server. Figure 5As shown, the electronic device includes a memory 402, a processor 404, and a transmission device 406. The memory 402 stores a computer program, and the processor 404 is configured to execute the steps of any of the above method embodiments through the computer program.

[0111] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0112] Optionally, the transmission device 406 is used to receive or send data via a network. Specific examples of the network described above may include wired and wireless networks. In one example, the transmission device 406 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 406 is a Radio Frequency (RF) module used to communicate with the Internet wirelessly. Furthermore, the electronic device also includes a display 408 and a connection bus 410, which connects the various module components within the electronic device.

[0113] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for simulating the performance of materials used in island and reef engineering, characterized in that, include: S1, acquire historical tidal data and on-site monitoring data of island and reef engineering structures; S2, initialize and maintain a shared state space, wherein the characteristic parameters of the shared state space are the microenvironment activity index, hysteresis memory state, and virtual aggregate field; S3, based on the historical tidal data and the on-site monitoring data, the hysteresis memory state is structurally edited using a reinforcement learning framework to update the microenvironment activity index; S4, The updated microenvironment activity index is used as a condition and input into the physical constraint generation model to update the virtual aggregate field and calculate its corresponding pumping dynamics response; S5. Based on the pump suction dynamics response and the updated microenvironment activity index, and combined with the tidal-driven boundary conditions, update the shared state space. S6. Based on the updated shared state space, output the predicted material properties of the island and reef engineering structure.

2. The method for simulating the performance of island and reef engineering materials according to claim 1, characterized in that, S3 utilizes a reinforcement learning framework to perform structured editing of the delayed memory states, including: S31, construct multiple hysteresis memory units corresponding to the frequencies of the main tidal components, wherein the flipping threshold spacing of the hysteresis memory units corresponding to the diurnal tidal component is set to a first range, and the flipping threshold spacing of the hysteresis memory units corresponding to the semi-diurnal tidal component is set to a second range greater than the first range.

3. The method for simulating the performance of island and reef engineering materials according to claim 2, characterized in that, The method further includes: S32, Define the state of the reinforcement learning framework, the state of the reinforcement learning framework includes: the historical sequence of the microenvironment activity index, the porosity and connectivity distribution represented by the virtual aggregate field, and the current state set of all hysteresis memory units. S33, define the actions of the reinforcement learning framework, the actions of the reinforcement learning framework including performing at least one of the following editing operations on all delayed memory units: The flip threshold of the selected cell is numerically drifted according to the current tidal acceleration; the state of the cells in the preset spatial area is reset; and a new cell with a new flip threshold is added.

4. The method for simulating the performance of island and reef engineering materials according to claim 3, characterized in that, The method further includes: S34, define the reward function of the reinforcement learning framework, wherein the reward function is a weighted sum of at least two sub-rewards: a first sub-reward based on the difference between simulated and measured electrochemical impedance spectra, a second sub-reward based on the consistency of the output of the physical constraint generation model under predicted and observed conditions, and a third sub-reward based on the synchronicity of system activity changes and tidal drive. S35, based on the state, action, and reward function of the reinforcement learning framework, execute reinforcement learning decisions, update the lag memory state, and aggregate to obtain the updated microenvironment activity index.

5. The method for simulating the performance of island and reef engineering materials according to claim 1, characterized in that, The physical constraint generation model described in S4 is a generating flow network. This model updates the virtual aggregate field and calculates its corresponding pumping dynamics response, including: S41, with the updated microenvironment activity index and spatial coordinates as input conditions; S42, Execute the serialization generation step: In each step, based on the pressure field gradient direction and structural growth history inside the aggregate structure generated in the current part, the Boltzmann distribution is used to select the next generation action from the action set including pore growth, region closure, branching and merging. The action score of the generation action is determined by the pressure field gradient alignment and historical growth continuity.

6. The method for simulating the performance of island and reef engineering materials according to claim 5, characterized in that, The method further includes: S43, after each generation action is executed, the differentiable physics calculation layer is called to perform a single-step fluid pressure and ion concentration field simulation on the current structure; S44, Based on the physical field residuals output by the differentiable physical computing layer, adjust the selection strategy for subsequent generation actions; S45, repeat S42 to S44 to generate a complete virtual aggregate field and simultaneously output a parameterized response function to characterize pump suction dynamics.

7. The method for simulating the performance of island and reef engineering materials according to claim 1, characterized in that, Between S3 and S4, it also includes: The updated microenvironment activity index and target area information are sent to the physical constraint generation model as its condition input. The system receives key physical features of the virtual aggregate field generated based on the physical constraint generation model and uses them to update the state representation of the reinforcement learning framework. The key physical features include the average aggregate porosity, feature transmission path length, and pump response time scale.

8. The method for simulating the performance of island and reef engineering materials according to claim 1 or 7, characterized in that, Between S3 and S4, it also includes: Calculate the spatial change of the microenvironment activity index before and after the update; The spatial variation is sent to the physical constraint generation model; The physical constraint generation model dynamically adjusts the main diagonal component of the Darcy permeability tensor of the differentiable physical computation layer based on the spatial variation.

9. The method for simulating the performance of island and reef engineering materials according to claim 1, characterized in that, The method further includes: After every integer number of tidal cycles, perform the following iterations: With the current virtual aggregate field fixed, S3 and S5 are iteratively run, and the microenvironment activity index is optimized when the rate of change of the reward function of the reinforcement learning framework is less than a preset change threshold or the number of iterations reaches the maximum number of iterations. Fix the current microenvironment activity index, iteratively run S4, and sample to generate a virtual aggregate field; The activity index of the currently paired microenvironment and the virtual aggregate field were validated.

10. A simulation system for the performance of island and reef engineering materials, wherein the system implements the simulation method for the performance of island and reef engineering materials as described in claim 1, characterized in that, include: The data acquisition unit is used to acquire historical tidal data and on-site monitoring data of island and reef engineering structures. An initialization unit is used to initialize and maintain a shared state space, wherein the characteristic parameters of the shared state space are the microenvironment activity index, the hysteresis memory state, and the virtual aggregate field. The structured editing unit is used to perform structured editing on the hysteresis memory state based on the tidal history data and the on-site monitoring data using a reinforcement learning framework, so as to update the microenvironment activity index. The first update unit is used to input the updated microenvironment activity index as a condition into the physical constraint generation model in order to update the virtual aggregate field and calculate its corresponding pumping dynamic response. The second update unit is used to update the shared state space based on the pump suction dynamics response and the updated microenvironment activity index, combined with the tidal drive boundary conditions. The output unit is used to output the predicted material properties of the island and reef engineering structure based on the updated shared state space.

Images