Green foundation material selection and structure design method based on machine learning

Abstract

The invention discloses a green foundation material selection and structure design method based on machine learning, and relates to the technical field of civil engineering and artificial intelligence, and the method comprises the steps: obtaining a resonance forbidden zone frequency and a candidate material micromorphological feature vector; determining macroscopic rigidity by utilizing a multi-scale constitutive mapping network through an iterative optimization module, calculating a coupling coefficient of an inherent frequency and a resonance forbidden zone, reconstructing microscopic stress and predicting a fatigue damage probability; if not, based on a first change rate of the damage probability with respect to the micromorphology and a second change rate of the coupling coefficient with respect to the rigidity, performing double-gradient collaborative updating on the microcosmic vector until a convergence condition is met, and finally generating a material processing parameter and a structure scheme; according to the method, the problem of macro-micro cross-scale coupling optimization is solved, and the dynamic stability of the foundation is improved.

Description

Technical Field

[0001] This invention relates to the fields of civil engineering and artificial intelligence technology, specifically to a method for selecting green foundation materials and designing structures based on machine learning. Background Technology

[0002] In the fields of civil engineering and green building, with the popularization of low-carbon and environmental protection concepts, using recycled aggregates from construction solid waste to construct foundations has become an important direction for technological development. Especially in complex environments such as near subways and high-speed railways where high-frequency vibration loads are subjected to long-term stress, the dynamic stability of the foundation structure is directly related to the safety of the superstructure.

[0003] However, according to the basic consensus of structural dynamics, the severity of the response of a foundation structure to environmental vibrations depends fundamentally on the matching relationship between the structure's own natural frequency and the frequency of the external environmental vibration source. However, when using non-homogeneous materials such as recycled aggregates, when designers attempt to increase the frictional energy dissipation (i.e., damping) of the micro-interface by increasing the sharpness or roughness of the aggregates, the overall macroscopic elastic modulus (i.e., stiffness) of the material will also undergo nonlinear dynamic drift.

[0004] If the design only pursues high damping characteristics at the microscopic level of the material and ignores the shift of the structure's natural frequency caused by stiffness changes, it is very likely that the natural frequency of the foundation will accidentally slip into the frequency band where environmental vibration energy is concentrated, thereby inducing a macroscopic resonance amplification effect. At this time, the huge dynamic stress excited by the resonance will far exceed the energy limit that the material's microscopic damping can dissipate, causing the foundation structure to suffer severe fatigue failure before reaching the expected service life. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for selecting green foundation materials and designing structures based on machine learning.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this invention discloses a method for selecting green foundation materials and designing structures based on machine learning, comprising the following steps:

[0008] Acquire the set of resonant exclusion frequencies characterizing the vibration characteristics of the target site environment, the preset structural geometric boundary parameters, and the micromorphological feature vectors of candidate foundation materials;

[0009] The micromorphological feature vector is iteratively updated until the generated fatigue damage probability value meets the preset convergence condition.

[0010] The micromorphological feature vector that satisfies the convergence condition is determined as the optimal micromorphological feature vector, and the corresponding material processing control parameters and foundation structure design scheme are generated based on it.

[0011] The iterative update process includes: inputting the micromorphological feature vector into a pre-trained multi-scale constitutive mapping network to determine the macroscopic equivalent complex stiffness matrix of the candidate foundation material;

[0012] Based on the macroscopic equivalent complex stiffness matrix and structural geometric boundary parameters, the natural frequency characteristics of the target foundation structure composed of candidate foundation materials are calculated.

[0013] The degree of overlap between the natural frequency characteristics and the resonant forbidden frequency set is calculated to generate the macroscopic resonant coupling coefficient, and the equivalent dynamic stress amplitude at the microscale is calculated based on the macroscopic resonant coupling coefficient.

[0014] The equivalent dynamic stress amplitude and the micromorphological feature vector are input into the fatigue life prediction model to generate fatigue damage probability values.

[0015] If the fatigue damage probability value does not meet the convergence condition, the first rate of change of the fatigue damage probability value with respect to the micromorphological feature vector and the second rate of change of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix are calculated, and the micromorphological feature vector is updated based on the first rate of change and the second rate of change.

[0016] Secondly, this invention discloses a green foundation material selection and structural design system based on machine learning, comprising:

[0017] The data acquisition module is used to acquire the set of resonant exclusion frequencies characterizing the vibration characteristics of the target site environment, the preset structural geometric boundary parameters, and the micromorphological feature vectors of candidate foundation materials;

[0018] The iterative optimization module is used to perform iterative update processing on the micro-morphological feature vector until the generated fatigue damage probability value meets the preset convergence condition.

[0019] The scheme generation module is used to determine the micromorphological feature vector that meets the convergence condition as the optimal micromorphological feature vector, and generate the corresponding material processing control parameters and foundation structure design scheme based on it.

[0020] The iterative optimization module includes:

[0021] Constitutive mapping unit is used to input micromorphological feature vectors into a pre-trained multi-scale constitutive mapping network to determine the macroscopic equivalent complex stiffness matrix of candidate foundation materials.

[0022] The frequency characteristic calculation unit is used to calculate the inherent frequency characteristics of the target foundation structure composed of candidate foundation materials based on the macroscopic equivalent complex stiffness matrix and structural geometric boundary parameters.

[0023] The resonant stress analysis unit is used to calculate the degree of overlap between the natural frequency characteristics and the resonant forbidden frequency set, generate the macroscopic resonant coupling coefficient, and calculate the equivalent dynamic stress amplitude at the microscale based on the macroscopic resonant coupling coefficient.

[0024] The fatigue prediction unit is used to input the equivalent dynamic stress amplitude and the micromorphological feature vector into the fatigue life prediction model to generate fatigue damage probability values.

[0025] The vector update unit is used to calculate the first rate of change of the fatigue damage probability value with respect to the micromorphological feature vector and the second rate of change of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix when the fatigue damage probability value does not meet the convergence condition, and to update the micromorphological feature vector based on the first rate of change and the second rate of change.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. By constructing a closed-loop calculation logic of micromorphology-macrostiffness-natural frequency-resonance coupling, the problem of neglecting stiffness drift when improving material damping, which leads to the structure's natural frequency accidentally slipping into the resonance forbidden zone, is overcome. By using the first rate of change of fatigue damage probability with respect to micromorphology (damping / microscopic perspective) and the second rate of change of macroscopic resonance coupling coefficient with respect to stiffness matrix (stiffness / macroscopic perspective) for dual gradient updates, the system can automatically find the global optimal solution in the multidimensional feature space that can maximize microscopic energy dissipation and accurately achieve macroscopic frequency domain avoidance, thereby significantly improving the dynamic stability and fatigue life of the foundation structure in complex vibration environment.

[0028] 2. After determining the optimal micro-morphological feature vector, the design vector is decoded into specific material processing control parameters by using a pre-set process parameter inverse regression model. This directly breaks down the barrier between theoretical calculation and on-site construction, ensuring that the calculated optimal aggregate morphology is not only in the digital model, but can be accurately manufactured by existing mechanical equipment. This greatly enhances the engineering implementation value and construction guidance significance of the green foundation material design scheme.

[0029] 3. When an iteration timeout is detected and the macroscopic resonance coupling coefficient remains high, the system automatically switches to the damping saturation-dominated mode. By forcibly eliminating the stiffness adjustment gradient and maximizing the damping weight, the system guides the material to evolve towards a high-energy-dissipation form that fully resists earthquakes. This ensures that even if perfect earthquake avoidance cannot be achieved, an effective solution can still be output to ensure structural safety by maximizing energy dissipation at the microscopic interface. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is an overall block diagram of the method in Embodiment 1 of the present invention;

[0032] Figure 2 This is a flowchart of the iterative update processing logic in the method of Embodiment 1 of the present invention;

[0033] Figure 3 This is an overall block diagram of the system in Embodiment 2 of the present invention. Detailed Implementation

[0034] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Application Overview: In the field of modern green civil engineering, especially in foundation treatment projects near strong vibration sources such as subways and high-speed railways, the use of recycled aggregates from construction solid waste to construct high-damping foundations is considered a key technical path to achieve the dual goals of low carbon emissions and vibration reduction. The construction of such high-performance foundations is essentially a synergistic process of precise spectrum control and energy dissipation at the structural dynamics level. That is, vibration energy is dissipated through frictional slippage at the micro-aggregate interface (damping effect), while the natural frequency is configured outside the resonance forbidden zone of the environmental vibration source by utilizing the stiffness distribution of the macro-structure (tuning effect), thereby ensuring the dynamic stability of the foundation system throughout its entire life cycle.

[0036] However, existing technologies lack a verification mechanism for the nonlinear coupling relationship between the evolution of microscopic material morphology and the frequency domain response of macroscopic structures. This leads to an inability to accurately identify the stiffness-damping inversion paradox and implicit resonance traps in the design process. The stiffness-damping inversion paradox manifests as designers attempting to increase microscopic damping by increasing aggregate roughness, while ignoring the inevitable nonlinear hardening drift of macroscopic equivalent stiffness caused by this physical change. The implicit resonance trap manifests as an unintentional change in macroscopic stiffness causing a shift in the structure's natural frequency, causing it to accidentally slip into the frequency band where environmental vibration energy is concentrated. As a result, a strict dynamic mapping relationship cannot be established between the selection of microscopic morphology and the setting of macroscopic frequency, leading to misjudgment of potential resonance risks and thus affecting the accuracy of foundation fatigue life assessment.

[0037] For example, in recycled concrete filling projects near subway lines, conventional design systems often only focus on the physical strength indicators of aggregates, attempting to enhance interlocking force and damping ratio by selecting angular crushed aggregates. However, they fail to anticipate that this change in microstructure will significantly improve the overall stiffness of the foundation. Furthermore, when the increase in stiffness causes the foundation's first natural frequency to shift from the safe low-frequency range to the high-frequency range, the system only records the apparent advantage of increased damping and fails to detect that the natural frequency happens to overlap with the characteristic frequency generated by subway operation. Specifically, the system misjudges high-stiffness, high-damping materials as the optimal choice, unaware that this stiffness change induces a macroscopic resonance amplification effect, resulting in the foundation bearing dynamic stress amplitudes far exceeding expectations during actual service.

[0038] If the above problems are not addressed, the foundation design system will continue to lose its ability to objectively judge the structural safety under complex dynamic environments. In particular, the failure to quantify the stiffness drift caused by micromorphology will lead to the design scheme relying excessively on empirical damping superposition, resulting in the loss of control over the structure's natural frequency and thus weakening the foundation's seismic resistance. At the same time, the failure to correct implicit resonance traps will cause macroscopic resonance energy to accumulate rapidly at the microscopic interface, making the aggregate contact surface unable to withstand high-frequency alternating stress, ultimately leading to the rapid propagation of micro-cracks and inducing premature failure of the macroscopic structure. As a result, the inaccuracy of design feedback will systematically hinder the safe application of green foundation materials in dynamic engineering and affect the achievement of the goal of long-term service of infrastructure.

[0039] Example 1:

[0040] like Figures 1-2 As shown, a machine learning-based method for selecting green foundation materials and designing structures includes the following steps:

[0041] Step S1: Obtain the set of resonant exclusion zones that characterize the vibration characteristics of the target site environment, the preset structural geometric boundary parameters, and the micromorphological feature vectors of candidate foundation materials;

[0042] First, the system utilizes an array of accelerometers deployed at the target foundation site to collect and acquire real-time time-series data of the environmental dynamic loads at the site. This data records the changes in the acceleration amplitude experienced by the foundation within a certain time window. To reveal the frequency domain characteristics hidden in the time series, the system performs a Fast Fourier Transform (FFT) on the environmental dynamic load time-series data, converting it from a time-domain signal to a frequency-domain signal, and further calculates the power spectral density (PSD) of the frequency-domain signal, thereby obtaining a spectrum of energy distribution with frequency. Based on this, the system introduces a preset safety threshold as a criterion, marking frequency intervals with energy values ​​in the power spectral density higher than the preset safety threshold as resonance-free zones. This means that environmental vibration energy is highly concentrated in these frequency bands, making it extremely easy to induce structural resonance. Finally, the system combines all frequency intervals marked as resonance-free zones to generate the final set of resonance-free zone frequencies. For example, in a typical subway line site, after the above processing, two high-energy frequency bands, [15Hz, 25Hz] and [40Hz, 50Hz], might be identified as resonance-free zones.

[0043] Secondly, obtaining the pre-defined structural geometric boundary parameters is fundamental to establishing a macroscopic mechanical model of the foundation. In practical applications, these parameters are directly derived from the architectural design drawings or BIM models, specifically including geometric dimensions such as the length, width, and depth of the foundation, as well as parameters describing the structural morphology, such as shape factors. These parameters will be used in subsequent steps to construct the mass and stiffness matrices of the foundation, thereby determining the dynamic characteristics of the structure.

[0044] Finally, for obtaining the microscopic morphological feature vectors of candidate foundation materials, this embodiment constructs a deep learning-based microscopic morphological representation mechanism. Before performing specific feature extraction, the system pre-trains a dedicated three-dimensional convolutional neural network (3D-CNN). The construction of this network takes into account the complexity of the morphology of recycled aggregates. Its input layer is designed to receive three-dimensional voxel mesh data, and the intermediate layers use multiple layers of three-dimensional convolutional kernels and pooling layers to capture the local geometric features and global topological structure of the aggregates in three-dimensional space. The output layer maps to specific physical property indicators. In actual operation, the system first obtains three-dimensional scanned voxel model data of candidate foundation materials (such as recycled concrete aggregates) through high-precision industrial CT or laser scanning equipment. These data accurately restore the real physical morphology of the aggregates. Subsequently, these voxel model data are input into the aforementioned pre-set three-dimensional convolutional neural network for feature encoding. The network output layer automatically extracts three key geometric and topological attributes: the angularity index, which characterizes the sharpness of the aggregate surface; the texture roughness index, which characterizes the surface unevenness; and the sphericity index, which characterizes the degree to which the overall shape approximates a sphere. To eliminate the influence of dimensions and facilitate unified processing in subsequent models, the system normalizes the extracted angularity index, texture roughness index, and sphericity index, mapping them to the [0,1] interval. The normalized values ​​are then combined to construct the microscopic morphology feature vector. For example, the feature vector of an aggregate with sharp edges and a rough surface might be represented as [0.85, 0.78, 0.45], where 0.85 represents a high angularity index, 0.78 represents a high texture roughness, and 0.45 represents a low sphericity. In this way, this embodiment successfully transforms the complex microscopic physical morphology of materials into a processable digital vector, laying a solid data foundation for subsequent macro-micro coupling analysis.

[0045] Step S2: Perform iterative update processing on the micro-morphological feature vector until the generated fatigue damage probability value meets the preset convergence condition;

[0046] Step S201: Input the micro-morphological feature vector into the pre-trained multi-scale constitutive mapping network to determine the macro-equivalent complex stiffness matrix of the candidate foundation material;

[0047] Given that foundation materials (especially recycled aggregate concrete) are inherently heterogeneous particle packing systems, their macroscopic mechanical response is highly dependent on the contact interactions and interfacial transfer effects between aggregate particles. Traditional fully connected neural networks struggle to capture this complex topological dependency. Therefore, this embodiment employs a Graph Neural Network (GNN) as the underlying architecture of the surrogate model. During the model's pre-training phase, the system utilizes discrete element method (DEM) or finite element method (FEM) simulations to generate a massive amount of microstructure-macroscopic response pairwise data as a training set, enabling the model to learn the ability to deduce macroscopic mechanical behavior from microscopic contact mechanisms.

[0048] In actual operation, when the system receives the current micromorphological feature vector (including the edge index, texture roughness index, and sphericity index) output in step S1, it first performs the instantiation and construction of the micromechanical graph structure. Specifically, the system generates representative volume elements (RVEs) in virtual space, maps aggregate particles to nodes in the graph structure, and assigns the micromorphological feature vector to these nodes as initial state attributes; simultaneously, it maps the interparticle transition zone (ITZ) and contact surface to edges in the graph structure. In particular, to reflect physical realism, the system sets the edge weights to be positively correlated with the texture roughness index in the micromorphological feature vector. This is because physical principles show that the roughness of the contact surface directly determines the friction coefficient and energy dissipation capacity, thus affecting the stress wave transmission efficiency.

[0049] After construction, the system utilizes the message passing mechanism of a graph neural network to perform multi-layer feature update processing on the microscopic mechanical graph structure. In each layer update iteration, the algorithm simulates the propagation process of stress waves in the particle skeleton: each node calculates a weighted sum of the feature information of its neighboring nodes based on the weights of the edges connected to it, simulating the mechanical action of surrounding particles on the central particle; subsequently, the weighted sum is transformed using a nonlinear activation function (such as ReLU or Tanh) to generate the updated node state vector. After multiple layers (e.g., 3-5 layers) of deep propagation, local microscopic features complete global diffusion and fusion within the graph structure.

[0050] Finally, the system performs a global readout operation on the graph structure after multi-layer feature update processing. This typically employs average pooling or weighted summation to aggregate the updated state vectors of all nodes into a high-dimensional vector, which is then mapped through a fully connected layer to a frequency-varying complex numerical matrix—the final macroscopic equivalent complex stiffness matrix. In this matrix, the real part physically represents the material's storage modulus, reflecting the elastic stiffness of the foundation and directly related to the structure's resistance to deformation; while the imaginary part physically represents the material's loss modulus, reflecting the foundation's damping characteristics and energy dissipation capacity. For example, if the input microscopic morphological feature vector indicates that the aggregate has high texture roughness, after the above GNN inference, the imaginary part of the output complex stiffness matrix will significantly increase, indicating that the material has a stronger macroscopic vibration energy absorption capacity. Through this process, this embodiment successfully achieves cross-scale extrapolation from microscopic geometric parameters to macroscopic complex domain mechanical parameters.

[0051] Furthermore, to ensure the robust generalization ability of the multi-scale constitutive mapping network (GNN) when faced with complex aggregate morphologies never before encountered, this embodiment employs a multi-dimensional morphological perturbation generation strategy when constructing the microstructure-macroresponse training dataset. The system uses parametric modeling techniques to randomly generate a virtual aggregate model library (containing over 10,000 morphological variants) with different aspect ratios, angles, and fractal dimensions within a preset particle size range. For each virtual sample, multi-directional uniaxial compression and shear simulations are performed to obtain its full anisotropic stiffness matrix. This training data construction method based on high-dimensional morphological space coverage ensures that the GNN model can not only memorize specific samples but also learn the deep physical mapping between microscopic geometric topology and macroscopic mechanical constitutive models.

[0052] Step S202: Based on the macroscopic equivalent complex stiffness matrix and structural geometric boundary parameters, calculate the natural frequency characteristics of the target foundation structure composed of candidate foundation materials;

[0053] The core of this step is to project abstract material mechanics parameters into a specific engineering geometric space, and through dynamic simulation calculations, reveal the inherent vibration characteristics of the foundation structure made of this specific candidate material in the physical world.

[0054] First, the system calls the preset structural geometric boundary parameters (such as the length, width, height, shape factor, and boundary constraints of the foundation) obtained in step S1, and discretizes the target foundation structure in digital space, dividing it into a finite number of grid cells. Based on the physical density distribution of the candidate materials, it assembles and generates a target foundation structure mass matrix that describes the mass distribution characteristics of the system. This matrix is ​​a large sparse symmetric matrix, and its diagonal elements accurately reflect the inertial characteristics of each node of the structure.

[0055] Simultaneously, the system needs to construct a target foundation structure stiffness matrix describing the structure's resistance to deformation. During this process, the system performs a physical separation of the macroscopic equivalent complex stiffness matrix output in step S1. Given that in structural dynamics principles, the undamped natural frequency of a system primarily depends on its elastic restoring force, i.e., the real part of the stiffness, while the imaginary part mainly affects vibration attenuation (which will be processed in subsequent steps), this embodiment explicitly extracts the real part data of the macroscopic equivalent complex stiffness matrix. This real part data characterizes the material's storage modulus (i.e., elastic modulus), which the system maps to each discrete grid element. Through the integration of element stiffness matrices, a global structural stiffness matrix is ​​constructed.

[0056] After constructing both the mass matrix and the stiffness matrix, the system transforms the physical problem into a mathematical generalized eigenvalue problem. Specifically, the system constructs its dynamic characteristic equations. (in Here is the stiffness matrix. The equation is solved using a numerical solution algorithm (such as the Lanczos algorithm or subspace iteration method), with the goal of finding a scalar that makes the equation true. (Eigenvalues) and vectors (Eigenvectors). Wherein, each eigenvalue obtained is... Each of these directly corresponds to a vibration mode of the structure, and the system is determined by the formula. The eigenvalues ​​are converted to Hertz (Hz) units, resulting in a series of frequency values ​​arranged by order. This vector of values ​​constitutes the final determined natural frequency characteristic. For example, the calculation results might show that the first-order natural frequency of the foundation structure is 18Hz and the second-order natural frequency is 26Hz. These precise values ​​provide a quantitative basis for subsequent judgments on whether it falls within the resonance exclusion zone.

[0057] Step S203: Calculate the degree of overlap between the natural frequency characteristics and the set of resonant forbidden frequencies, generate the macroscopic resonant coupling coefficient, and calculate the equivalent dynamic stress amplitude at the microscale based on the macroscopic resonant coupling coefficient;

[0058] The purpose of this step is to quantify the physical coupling relationship between the structure's natural frequency and the concentrated area of ​​environmental vibration energy, and to accurately map this macroscopic resonance risk into the actual mechanical load that the micro-aggregate interface needs to withstand.

[0059] Specifically, the system retrieves the natural frequency characteristics output in step S202 and the set of resonant forbidden frequencies determined in step S1. In physical reality, considering the existence of damping and calculation errors, the natural frequency of the foundation is not an absolute mathematical isolated point, but a frequency distribution interval with a certain bandwidth; similarly, the set of resonant forbidden frequencies consists of several high-energy-density frequency bands with lower and upper bound frequencies. In the frequency domain, the system calculates the overlapping integral area between the frequency distribution interval corresponding to the natural frequency characteristics and the set of resonant forbidden frequencies using a numerical integration algorithm. This integral area intuitively reflects the potential of the foundation structure to absorb environmental vibration energy. To standardize this physical quantity, the system maps the integral area to a normalized value, namely the macroscopic resonant coupling coefficient (…). In this mapping logic, the system incorporates two extreme boundary condition determination mechanisms: when the intersection of the frequency distribution interval corresponding to the inherent frequency characteristics and the resonant forbidden frequency set is an empty set, it indicates that the structure has successfully achieved frequency domain avoidance, and the system will... The value is assigned a preset attenuation benchmark (usually less than 1, representing the isolation effect); conversely, when the natural frequency characteristic is exactly located at the arithmetic mean of the lower and upper bound frequencies of the resonance forbidden frequency set (i.e., the energy center), it indicates that the most intense resonance has been triggered, and the system will... The value is assigned to a preset resonant amplification extreme value (this value is usually determined based on the quality factor of the structure).

[0060] After establishing the coefficient characterizing the degree of macroscopic resonance Subsequently, this embodiment further transforms it into a microscopic-scale mechanical input. This is because material failure often begins with stress concentration at the microscopic interface, and the source of this stress concentration is the environmental energy amplified by macroscopic resonance. Based on the principle of energy transfer, the system calculates the equivalent dynamic stress amplitude acting on the microscopic interface using the following preset physical conversion formula ( ):

[0061]

[0062] In this formula, This represents the total energy integral value corresponding to the set of resonant forbidden frequencies, and its square root characterizes the original amplitude level of the environmental vibration. The preset stress transfer coefficient is a dimensionless empirical parameter that depends on the geometric scale effect of the foundation structure and the degree of material homogenization. It describes the attenuation or concentration characteristics during the transfer of macroscopic average stress to microscopic local stress. Through the above calculations, this embodiment successfully transforms the abstract frequency overlap problem into a specific microscopic mechanical load, i.e. This value precisely quantifies the dynamic shear force or tensile / compressive force amplitude actually borne by the surface of aggregate particles under the chain reaction that causes changes in macroscopic stiffness due to the current micromorphology, and then triggers a certain degree of resonance, providing real physical boundary conditions for subsequent fatigue life prediction.

[0063] It is worth noting that, in order to ensure the stress transfer coefficient in the above formula ( To ensure the physical accuracy of the values, this embodiment does not simply use fixed constants, but rather employs a dynamic calibration method based on RVE boundary conditions. Specifically, during the pre-training phase, the system applies a unit macroscopic average stress to representative volumetric elements (RVEs) with different porosities and aggregate contact modes, and calculates the maximum local stress response of the aggregate interface inside the RVE through finite element micromechanical simulation; the ratio of this maximum local stress to the macroscopic average stress is defined as the reference transfer coefficient under this condition. In actual calculations, the system dynamically matches the closest value based on the current foundation structure's macroscopic porosity design parameters through table lookup interpolation or regression functions. This treatment method effectively corrects the nonlinear amplification effect of macroscopic stress in the microscopic transmission process caused by material heterogeneity.

[0064] Step S204: Input the equivalent dynamic stress amplitude and micromorphological feature vector into the fatigue life prediction model to generate fatigue damage probability values;

[0065] This step aims to combine mechanical values ​​with specific material microstructures and use artificial intelligence technology to predict the probability of fatigue failure of the foundation structure during its expected service life under the current design scheme.

[0066] Before executing the prediction step, this embodiment pre-constructs and trains a Physics-Informed Fatigue Prediction Model. This model is not simply built upon data-driven black-box logic, but rather integrates fracture mechanics theory with a deep neural network architecture. Specifically, the model's input layer is designed as a dual-channel structure, receiving scalar-type mechanical load data and vector-type morphological feature data respectively; the intermediate layer employs a multilayer fully connected network (MLP) and residual connection structure to capture high-dimensional nonlinear features; and the output layer is mapped to probability values ​​in the [0,1] interval using a Sigmoid activation function. To ensure the physical plausibility of the prediction, during the model training phase, the system introduces a loss function constraint based on Miner's linear cumulative damage theory. Iterative optimization of the model weights is performed using massive amounts of material fatigue test data (including cyclic loading failure data under different aggregate morphologies and stress levels) to accurately reflect the entire process mechanism from microcrack initiation to macroscopic fracture.

[0067] In actual operation, the system simultaneously outputs the equivalent dynamic stress amplitude at the microscale in step S203 ( The current microscopic morphological feature vector obtained in step S1 is input into the fatigue life prediction model. The model first performs a dynamic derivation of the stress concentration factor: based on the angularity index in the microscopic morphological feature vector, the model identifies regions of geometric abrupt changes on the aggregate surface and calculates the local stress concentration factor in these regions. Physics principles show that the sharper the edges (i.e., the higher the angle index), the more significant the stress concentration effect. Subsequently, the model, combined with the input equivalent dynamic stress amplitude, calculates the actual local peak stress borne by the aggregate tip or interface transition zone (ITZ). This process profoundly reveals the destructive effect of the superposition of macroscopic resonance and microscopic morphology—even if the average stress transmitted macroscopically is not large, if the aggregate shape is too sharp, it can lead to… If the stress is too high, the local peak stress may still instantly exceed the fatigue strength threshold of the material.

[0068] Finally, the model compares the calculated local peak stress with the SN curve (stress-life curve) characteristics of the material matrix, comprehensively considering the contribution of texture roughness to interfacial bond strength (higher roughness results in stronger anti-slip ability and can delay crack propagation to some extent). Ultimately, a quantified fatigue damage probability value is generated through the output layer. This value intuitively characterizes the possibility of overall structural failure due to micro-fatigue accumulation within the design service life of a foundation structure constructed using recycled aggregate in its current form under vibration excitation in the current environment. For example, an output value of 0.05 indicates an extremely low failure risk, while 0.95 indicates an extremely high risk. This value will serve as the basis for subsequent steps to determine whether to initiate morphological evolution iteration.

[0069] Step S205: If the fatigue damage probability value does not meet the convergence condition, calculate the first rate of change of the fatigue damage probability value with respect to the micromorphological feature vector and the second rate of change of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix, and update the micromorphological feature vector based on the first rate of change and the second rate of change.

[0070] The core logic of this step is that when the current microstructure of the foundation material leads to an unacceptable risk of failure in the structure, the system uses mathematical gradient guidance to automatically find an optimal evolution path in the multidimensional feature space that can both reduce micro-stress concentration and avoid macro-resonance.

[0071] Specifically, the system first executes the convergence condition judgment logic, comparing the fatigue damage probability value output in step S204 with a preset safety threshold (e.g., 5% or 0.05). If the probability value is less than or equal to the safety threshold, the system determines that the current design scheme meets the reliability requirements, the iteration terminates, and it directly jumps to the subsequent steps to generate the final scheme; otherwise, if the probability value is higher than the safety threshold, it indicates that the microstructure of the current candidate material has a high risk under the current working conditions, and the system then starts the microstructure feature vector update process based on dual gradients.

[0072] The mathematical essence of this update process is solving a non-convex optimization problem constrained by multiple physics fields. This embodiment employs a gradient descent strategy based on the chain rule. The system first calculates the first rate of change ( This gradient, or gradient of the fatigue damage probability value with respect to the current microscopic morphological feature vector, is used in the preceding fatigue life prediction model. Technically, since the fatigue life prediction model is built on a differentiable deep neural network, the system utilizes an automatic differentiation mechanism to backpropagate and calculate the partial derivative of the output layer (damage probability) with respect to the input layer (morphological vector). The physical meaning of this gradient points to how changing the aggregate morphology can most directly reduce microscopic local stress, typically guiding the aggregate to evolve towards a more rounded shape with fewer sharp edges.

[0073] Simultaneously, the system calculates the second rate of change, a composite gradient describing the sensitivity of macroscopic resonance risk to microscopic morphology. This calculation process comprises two cascaded parts: first, the system calculates the gradient of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix (…). This reflects how the stiffness needs to change to move the natural frequency out of the resonance forbidden zone; secondly, the system calculates the partial derivatives of the macroscopic equivalent complex stiffness matrix with respect to the current microscopic morphological eigenvectors ( This step, achieved through backpropagation and differentiation of the aforementioned multi-scale constitutive mapping network (GNN), reflects how the microstructure needs to change to produce the required stiffness change. The system multiplies these two parts ( The second rate of change was obtained. The physical meaning of this gradient points to how changing the aggregate morphology can achieve frequency domain vibration damping by adjusting the structural stiffness.

[0074] After acquiring the gradient information from the two dimensions mentioned above, the system corrects the microscopic morphological feature vector based on a preset update strategy. To balance the potentially conflicting objectives of reducing microscopic damage and avoiding macroscopic resonance, the system introduces a fatigue damage weighting factor (…). ), resonance risk weighting factor ( ) and learning rate ( Perform the following weighted update operation:

[0075]

[0076] Using this formula, the system generates a microscopic morphological feature vector for the next iteration. This updated vector not only contains fine-tuning instructions for the aggregate geometry (such as slightly rounding edges or increasing surface roughness), but also implicitly suggests a reshaping of the structure's macroscopic dynamic characteristics. Subsequently, the system uses this updated vector as new input and automatically returns to step S201 to re-trigger the stiffness prediction, frequency calculation, and damage assessment processes until the system finds a globally optimal morphological solution that converges the fatigue damage probability value to within a safe threshold.

[0077] In the process of performing the above-mentioned routine update of micromorphological feature vectors based on dual gradients, this embodiment also incorporates an abnormal takeover mechanism after resonance escape failure, which aims to solve the problem of oscillation or non-convergence that the algorithm may fall into when the structure’s natural frequency cannot be moved out of the environmental resonance forbidden zone by adjusting the material stiffness (i.e., frequency domain deadlock occurs).

[0078] Specifically, in each iteration loop, the system acquires and records the current iteration number in real time. ) and the calculated macroscopic resonance coupling coefficient ( A state monitoring daemon runs in the system background, continuously determining whether the current state simultaneously triggers the two critical conditions for a survival deadlock:

[0079] Does the current number of iterations exceed the preset maximum iteration threshold (e.g., set to 50 rounds)? If it does, it indicates that a solution that simultaneously satisfies fatigue and resonance constraints has not been found in the conventional non-convex optimization space search.

[0080] Is the current macroscopic resonance coupling coefficient consistently higher than the preset resonance lock-in threshold (e.g.) If it is higher, it indicates that despite multiple morphological adjustments, the natural frequency of the foundation structure is always trapped within the resonance forbidden zone where environmental energy is most concentrated, and frequency domain migration is impossible.

[0081] When both of the above conditions are met, the system determines that the conventional damping strategy has failed, and the system immediately switches the iterative update processing logic from dual-objective collaborative optimization to damping saturation-dominated mode.

[0082] After entering this damping saturation-dominated mode, the system performs a downgraded reconstruction of the update logic for the micro-morphological characteristic vectors. The core idea is to abandon seismic avoidance and focus entirely on seismic resistance. The specific execution logic is as follows:

[0083] First, the system performs a gradient decoupling operation. The system forces the resonant risk weighting factor (which is used to weight the second rate of change, i.e., the stiffness and frequency adjustment term) to be removed. The value is set to zero. Mathematically, this operation directly eliminates the gradient component of the second rate of change on the update direction of the micromorphological eigenvector. Physically, it means that the system no longer attempts to adjust the macroscopic stiffness to avoid resonance by changing the aggregate morphology, because previous iterations have proven that this approach is not feasible.

[0084] Secondly, the system performs a saturation enhancement of the damping weights. The system will use the fatigue damage weighting factor (which involves the micro-damage and damping terms) to weight the first rate of change. Set to a preset saturation threshold (this threshold is usually significantly higher than the weight in normal mode, for example, set to...). (or higher). At this point, the system calculates the product of the first rate of change and the saturation threshold to obtain the single-objective update gradient.

[0085] Finally, the system uses this single-objective update gradient to unidirectionally update the current microscopic morphological feature vector. In this mode, the optimization direction is completely dominated by maximizing microscopic damping energy dissipation, typically guiding the aggregate towards extreme roughness or specific multifractal dimensions to forcibly dissipate the enormous energy generated by resonance by maximizing interfacial friction and scattering effects. This process continues until the generated fatigue damage probability value satisfies the preset convergence condition under the new high-damping state, thus outputting a foundation design scheme that remains safe and reliable under resonance conditions.

[0086] Step S3: Determine the micromorphological feature vector that satisfies the convergence condition as the optimal micromorphological feature vector, and generate the corresponding material processing control parameters and foundation structure design scheme based on it.

[0087] Specifically, the system first locks the current micromorphological feature vector that meets the convergence condition as the optimal micromorphological feature vector. To transform this vector into executable material processing control parameters, this embodiment pre-constructs and integrates a process parameter inverse regression model. This model is built upon extensive industrial field test data, establishing a multi-dimensional mapping relationship between aggregate micromorphological changes and crushing and shaping equipment operating parameters. During actual execution, the system calls the initial micromorphological feature vector of the candidate foundation material obtained in step S1 (…). ), and calculate the morphological difference vector between the optimal vector and the initial vector ( This difference vector precisely quantifies the degree of physical modification required for raw materials to meet design requirements.

[0088] Based on the calculated morphological difference vector, the system automatically triggers different process generation logics according to its positive and negative values: if the morphological difference vector indicates that the angularity index needs to be significantly reduced (i.e., This indicates that the raw material is too sharp, posing a risk of stress concentration. The system will automatically match the shaping process and output specific control parameters for the shaping equipment (such as a centrifugal impact shaping machine). These parameters may include reducing the throwing head speed to decrease the crushing ratio, or setting specific grinding time parameters to round the edges. Conversely, if the morphology difference vector indicates that the texture roughness index needs to be significantly increased (i.e.,...), the system will automatically adjust the shaping process accordingly. This indicates that the raw material surface is too smooth and the friction damping is insufficient. The system will automatically match the crushing process and output specific control parameters for the crushing equipment (such as an impact crusher). Specifically, this includes adjusting the discharge port gap parameter to increase the probability of crushing by compression, or increasing the rotor linear speed parameter to generate more fresh fracture surfaces. In this way, the system outputs a material processing procedure sheet accurate to the equipment speed and gap in millimeters.

[0089] While generating material processing control parameters, this embodiment also executes the generation process of foundation structure design scheme in parallel.

[0090] Specifically, the system utilizes the optimal micromorphological feature vector ( The macroscopic equivalent complex stiffness matrix (including storage modulus and loss modulus) determined by [the algorithm], combined with the preset structural geometric boundary parameters obtained in step S1, drives a pre-set foundation dynamics structure generation engine. This engine integrates a multiphysics-coupled finite element solver and a genetic optimization algorithm, and its operating logic is as follows:

[0091] First, the engine performs optimal gradation inversion, even though the microstructure (aggregate shape) has changed from... This is certain, but the ultimate realization of macroscopic stiffness still depends on the way the particles are packed. The engine is based on... The sphericity and angularity indices are used to calculate the theoretical packing density and number of contact points under different particle size distributions using the compressible packing model (CPM). The system uses the macroscopic stiffness determined in step S201 as the objective function and searches inversely for the optimal Taylor grading curve parameters (such as the maximum particle size). With gradation index For example, if the target stiffness is high, the engine might output a gradation index. A dense gradation scheme is used to maximize the interlocking effect between aggregates.

[0092] Secondly, the engine performs a dynamic layered design. Considering that the propagation characteristics of seismic waves in the foundation are closely related to the soil layer thickness (i.e., standing wave effect), the system calculates the critical layer thickness that can avoid vertical resonance based on the material's shear wave velocity (derived from the real part of stiffness and density) and the center frequency of the environmental resonance exclusion zone. If the calculation finds that a single homogeneous fill may induce standing wave resonance, the system will automatically generate a stiffness gradient composite design scheme, that is, output a set of compaction control indicators that vary with depth (e.g., 98% compaction of the bottom layer and 93% compaction of the surface layer), dissipating vibration energy by artificially creating a difference in wave impedance.

[0093] Finally, the system integrates the above calculation results and outputs a digital drawing that can guide construction, specifically including:

[0094] Aggregate gradation batching sheet: clearly specify the mass percentage of each particle size range (e.g., 5-10mm, 10-20mm);

[0095] Structural construction parameter table: precisely defines the layered filling thickness of the foundation (e.g., 30cm per layer), the optimum moisture content, and the target dry density;

[0096] Quality control thresholds: Based on fatigue life prediction results, set the lower limit for compaction testing and the upper limit for porosity at the construction site.

[0097] This in-depth implementation process not only solved the problem of what materials to use (aggregate form), but also the problem of how to stack the materials (gradation and structure), thereby ensuring that the excellent performance at the micro level can be transferred to the macro structural level without damage.

[0098] In summary, the embodiments of this invention not only mathematically deconstruct the nonlinear entanglement between the microscopic morphology and macroscopic dynamic properties of heterogeneous materials through a high-precision multi-scale mapping network, but also logically overcome the engineering paradox of high damping accompanied by stiffness hardening leading to frequency domain runaway through a physics-driven dual-gradient iteration mechanism. This closed-loop feedback, which deduces from microscopic constitutive model to macroscopic resonance and then maps environmental energy back to microscopic damage, essentially endows the foundation design system with a cross-scale adaptive evolutionary capability. This allows it to actively seek the optimal balance point between material morphology and structural parameters under the constraints of geological environment vibration, thereby reversing the passive situation in traditional design that only focuses on static strength and ignores dynamic matching for recycled materials. This establishes a new technical benchmark for the transformation of green geotechnical engineering from experience-based trial and error to precise and intelligent design.

[0099] Example 2:

[0100] like Figure 3 As shown, a machine learning-based green foundation material selection and structural design system includes:

[0101] The data acquisition module is used to acquire the set of resonant exclusion frequencies characterizing the vibration characteristics of the target site environment, the preset structural geometric boundary parameters, and the micromorphological feature vectors of candidate foundation materials;

[0102] The iterative optimization module is used to perform iterative update processing on the micro-morphological feature vector until the generated fatigue damage probability value meets the preset convergence condition.

[0103] The scheme generation module is used to determine the micromorphological feature vector that meets the convergence condition as the optimal micromorphological feature vector, and generate the corresponding material processing control parameters and foundation structure design scheme based on it.

[0104] The iterative optimization module includes:

[0105] Constitutive mapping unit is used to input micromorphological feature vectors into a pre-trained multi-scale constitutive mapping network to determine the macroscopic equivalent complex stiffness matrix of candidate foundation materials.

[0106] The frequency characteristic calculation unit is used to calculate the inherent frequency characteristics of the target foundation structure composed of candidate foundation materials based on the macroscopic equivalent complex stiffness matrix and structural geometric boundary parameters.

[0107] The resonant stress analysis unit is used to calculate the degree of overlap between the natural frequency characteristics and the resonant forbidden frequency set, generate the macroscopic resonant coupling coefficient, and calculate the equivalent dynamic stress amplitude at the microscale based on the macroscopic resonant coupling coefficient.

[0108] The fatigue prediction unit is used to input the equivalent dynamic stress amplitude and the micromorphological feature vector into the fatigue life prediction model to generate fatigue damage probability values.

[0109] The vector update unit is used to calculate the first rate of change of the fatigue damage probability value with respect to the micromorphological feature vector and the second rate of change of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix when the fatigue damage probability value does not meet the convergence condition, and to update the micromorphological feature vector based on the first rate of change and the second rate of change.

[0110] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0111] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0112] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for selecting green foundation materials and designing structures based on machine learning, characterized in that, Includes the following steps: Acquire the set of resonant exclusion frequencies characterizing the vibration characteristics of the target site environment, the preset structural geometric boundary parameters, and the micromorphological feature vectors of candidate foundation materials; The micromorphological feature vector is iteratively updated until the generated fatigue damage probability value meets the preset convergence condition. The micromorphological feature vector that satisfies the convergence condition is determined as the optimal micromorphological feature vector, and the corresponding material processing control parameters and foundation structure design scheme are generated based on it. The iterative update process includes: inputting the micromorphological feature vector into a pre-trained multi-scale constitutive mapping network to determine the macroscopic equivalent complex stiffness matrix of the candidate foundation material; Based on the macroscopic equivalent complex stiffness matrix and the structural geometric boundary parameters, the natural frequency characteristics of the target foundation structure composed of the candidate foundation materials are calculated. The degree of overlap between the inherent frequency characteristics and the set of resonant forbidden frequencies is calculated to generate a macroscopic resonant coupling coefficient, and the equivalent dynamic stress amplitude at the microscale is calculated based on the macroscopic resonant coupling coefficient. The equivalent dynamic stress amplitude and the micromorphological feature vector are input into the fatigue life prediction model to generate fatigue damage probability values. If the fatigue damage probability value does not meet the convergence condition, then calculate the first rate of change of the fatigue damage probability value with respect to the micromorphological feature vector and the second rate of change of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix, and update the micromorphological feature vector based on the first rate of change and the second rate of change.

2. The method for selecting and designing green foundation materials based on machine learning according to claim 1, characterized in that: The process of obtaining the set of resonant forbidden frequencies includes: Acquire time-series data of dynamic environmental loads at the target site; Perform a Fast Fourier Transform on the environmental dynamic load time series data to obtain a frequency domain signal, and calculate the power spectral density of the frequency domain signal; The frequency ranges in the power spectral density with energy values ​​higher than a preset safety threshold are marked as resonance forbidden zones. All frequency ranges marked as resonance forbidden zones are combined to generate the resonance forbidden zone frequency set.

3. The method for selecting and designing green foundation materials based on machine learning according to claim 2, characterized in that: The process of obtaining the microscopic morphological feature vector includes: Obtain three-dimensional scanning voxel model data of candidate foundation materials; The 3D scanned voxel model data is feature-encoded using a pre-set 3D convolutional neural network to extract the angularity index, texture roughness index, and sphericity index, which characterize the geometric topological properties of the aggregate. The edge index, texture roughness index, and sphericity index are normalized, and the normalized values ​​are combined to construct the micromorphological feature vector.

4. The method for selecting and designing green foundation materials based on machine learning according to claim 3, characterized in that: The process of determining the macroscopic equivalent complex stiffness matrix includes: The multi-scale constitutive mapping network is a proxy model based on graph neural networks; A micromechanical graph structure is constructed based on the micromorphological feature vector, wherein the node attributes of the graph structure correspond to the micromorphological feature vector, and the edge weights of the graph structure are positively correlated with the texture roughness index in the micromorphological feature vector. The micromechanical graph structure is updated in multiple layers using the message passing mechanism of a graph neural network: in each layer, the features of adjacent nodes are weighted and aggregated and nonlinearly activated based on the edge weights to generate updated node state vectors; a global readout operation is performed on the graph structure after the multi-layer update, and the node state vectors of the whole graph are fused and mapped into a complex numerical matrix that varies with frequency. The complex numerical matrix is ​​determined as the macroscopic equivalent complex stiffness matrix, wherein the real part of the macroscopic equivalent complex stiffness matrix represents the energy storage modulus of the material, and the imaginary part represents the loss modulus of the material.

5. The method for selecting and designing green foundation materials based on machine learning according to claim 4, characterized in that: The process of determining the inherent frequency characteristics includes: Construct the mass matrix of the target foundation structure based on the aforementioned structural geometric boundary parameters; The stiffness matrix of the target foundation structure is constructed based on the real part of the macroscopic equivalent complex stiffness matrix. The eigenvalue vector is obtained by solving the generalized eigenvalue problem of the mass matrix and the stiffness matrix; The frequency corresponding to the feature value vector is determined as the inherent frequency feature.

6. The method for selecting and designing green foundation materials based on machine learning according to claim 5, characterized in that: The generation process of the macroscopic resonance coupling coefficient includes: Calculate the integral area of ​​the inherent frequency feature in the frequency domain that falls within the set of resonant forbidden frequencies; The normalized value of the integral area is determined as the macroscopic resonance coupling coefficient; The resonant forbidden frequency set has a lower bound frequency and an upper bound frequency; When the intersection of the frequency distribution interval corresponding to the inherent frequency characteristic and the frequency set of the resonance forbidden zone is an empty set, the macroscopic resonance coupling coefficient is taken as a preset attenuation reference value. When the inherent frequency characteristic is located at the arithmetic mean of the lower bound frequency and the upper bound frequency, the macroscopic resonance coupling coefficient is taken as a preset resonance amplification extreme value.

7. The method for selecting and designing green foundation materials based on machine learning according to claim 6, characterized in that: The microscopic morphological feature vector is updated based on the first rate of change and the second rate of change. The specific calculation formula is as follows: in, This represents the micromorphological feature vector used in the next iteration. This represents the current micromorphic feature vector. This represents the preset learning rate. This represents the fatigue damage weighting factor. Indicates the resonance risk weighting factor; This represents the first rate of change, which is the gradient of the fatigue damage probability value with respect to the current micromorphological feature vector; This represents the gradient of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix; This represents the partial derivative of the macroscopic equivalent complex stiffness matrix with respect to the current microscopic morphological eigenvector.

8. The method for selecting and designing green foundation materials based on machine learning according to claim 7, characterized in that: The iterative update process for micromorphic feature vectors also includes: The current iteration number and the macroscopic resonance coupling coefficient are obtained in real time, and it is determined whether the following conditions are met simultaneously: the iteration number exceeds the preset maximum iteration threshold, and the macroscopic resonance coupling coefficient is higher than the preset resonance locking threshold. If so, the iterative update process will be switched to the damping saturation-dominated mode.

9. The method for selecting and designing green foundation materials based on machine learning according to claim 8, characterized in that: In the damping saturation-dominated mode, the micromorphological feature vector is updated, including: The resonance risk weight factor is set to zero, and the gradient component of the second rate of change on the micromorphological feature vector update direction is eliminated. The fatigue damage weighting factor is set to a preset saturation threshold. Calculate the product of the first rate of change and the saturation threshold to obtain the single-objective update gradient; The current micromorphological feature vector is updated according to the single-objective update gradient until the preset convergence condition is met.

10. A machine learning-based system for selecting green foundation materials and designing structures, characterized in that, include: The data acquisition module is used to acquire the set of resonant exclusion frequencies characterizing the vibration characteristics of the target site environment, the preset structural geometric boundary parameters, and the micromorphological feature vectors of candidate foundation materials; The iterative optimization module is used to perform iterative update processing on the micromorphological feature vector until the generated fatigue damage probability value meets the preset convergence condition. The scheme generation module is used to determine the micromorphological feature vector that satisfies the convergence condition as the optimal micromorphological feature vector, and generate the corresponding material processing control parameters and foundation structure design scheme based on it. The iterative optimization module includes: Constitutive mapping unit is used to input the micro-morphological feature vector into a pre-trained multi-scale constitutive mapping network to determine the macro-equivalent complex stiffness matrix of the candidate foundation material. The frequency characteristic calculation unit is used to calculate the inherent frequency characteristics of the target foundation structure composed of the candidate foundation material based on the macroscopic equivalent complex stiffness matrix and the structural geometric boundary parameters. The resonant stress analysis unit is used to calculate the degree of overlap between the natural frequency characteristics and the resonant forbidden frequency set, generate the macroscopic resonant coupling coefficient, and calculate the equivalent dynamic stress amplitude at the microscale based on the macroscopic resonant coupling coefficient. The fatigue prediction unit is used to input the equivalent dynamic stress amplitude and the micromorphological feature vector into the fatigue life prediction model to generate fatigue damage probability values. The vector update unit is used to calculate the first rate of change of the fatigue damage probability value with respect to the micromorphological feature vector and the second rate of change of the macroscopic resonance coupling coefficient with respect to the macroscopic equivalent complex stiffness matrix when the fatigue damage probability value does not meet the convergence condition, and update the micromorphological feature vector based on the first rate of change and the second rate of change.

Images