An unordered metamaterial and an unordered metamaterial customization method based on deep learning

Abstract

This invention discloses a method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning. The method includes the following steps: S1. A dataset generation module performs batch finite element numerical simulations to obtain the stress-strain curves of the disordered metamaterials; S2. A forward prediction model is trained, using a multilayer perceptron to forward predict the stress-strain curves of the disordered metamaterials, outputting stress-strain curve feature points uniformly sampled by displacement; S3. A disordered metamaterial inverse generation network model is constructed based on the collaborative optimization of the generator and predictor; the target curve is input into the generator, which outputs the corresponding disordered metamaterial microstructure distribution, and the predictor predicts and outputs the predicted curve based on the generated structure; by minimizing the difference loss between the target curve and the predicted curve, performance-driven structure optimization and parameter updating are achieved, generating a metamaterial structure that best matches the target performance. This method can achieve efficient and interpretable inverse design under limited sample conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent design of metamaterials, and in particular to a method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning. Background Technology

[0002] Disordered metamaterials are a class of artificial materials that achieve specific macroscopic mechanical properties through microstructure design, and are widely used in energy absorption, deformation control, and impact protection. However, their complex structures, high degree of design freedom, and strong nonlinear mechanical responses mean that their macroscopic properties are influenced by multiple factors such as microstructure topology, geometry, and spatial distribution, resulting in highly complex structure-performance relationships. Different structures may produce similar responses, increasing design uncertainty and optimization difficulty.

[0003] Currently, design methods for disordered metamaterials in engineering design largely rely on finite element forward simulation combined with optimization algorithms (such as genetic algorithms, particle swarm optimization, and Bayesian optimization) for structure search and performance matching. While these methods can improve design accuracy to some extent, they require extensive iterative calculations and simulation evaluations, resulting in high computational complexity, slow convergence speed, and a tendency to get trapped in local optima. They also struggle to achieve efficient exploration in high-dimensional design spaces, have limited model generalization capabilities, and are unsuitable for large-scale or real-time design tasks. Furthermore, traditional design methods based on geometric optimization or heuristic search rely solely on structural morphological differences for optimization, failing to establish a differentiable mapping relationship from performance to structure, leading to deviations between the generated structure and the target stress-strain response.

[0004] Existing deep learning reverse engineering methods often focus solely on minimizing geometric differences, neglecting the physical continuity and manufacturability constraints of metamaterial structures. This leads to the presence of non-physical suspended or isolated units in the generated results, compromising the overall mechanical integrity of the material. Furthermore, while optimization algorithms (such as genetic algorithms and particle swarm optimization) can introduce some physical inspiration, they struggle to guarantee gradient differentiability, limiting the collaborative optimization capabilities of the model and the deep network. Simultaneously, training data often fails to cover all target performance spaces, resulting in poor generalization ability of the model under unknown performance objectives. Summary of the Invention

[0005] To address the problems of existing technologies, this invention aims to provide a deep learning-based method for customizing the nonlinear mechanical properties of disordered metamaterials. This method combines deep learning with intelligent optimization strategies, constructing a dual-network collaborative framework of a performance evaluation proxy model (f-NN) and an inverse generation model (i-NN). It introduces continuously differentiable structural encoding and a multi-objective physical constraint loss function to achieve high-precision mapping from the target stress-strain curve to a suitable microstructure. Furthermore, the generated results are posteriorly screened and fine-tuned using genetic algorithms or Bayesian optimization to further enhance design diversity and global optimality, enabling efficient and customized design of disordered metamaterials under complex nonlinear mechanical properties. This method can achieve efficient and interpretable inverse design under limited sample conditions and generates diverse structural solutions through temperature sampling, making it applicable to fields such as energy absorption, protective structures, and smart materials.

[0006] To achieve the above objectives, this invention provides a method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning, the method comprising the following steps: S1. The dataset generation module performs batch finite element numerical simulations to obtain stress-strain curves of disordered metamaterials; S2. Train a positive prediction model and use a multilayer perceptron to make positive predictions on the stress-strain curve of the disordered metamaterial. The output is the stress-strain curve feature points sampled uniformly according to displacement. S3. Construct a disordered metamaterial inverse generation network model based on the collaborative optimization of the generator and predictor; the generator adopts an Encoder-Decoder architecture, inputting the target curve into the generator, which transforms the target curve into a disordered metamaterial microstructure distribution; the predictor is the forward prediction model trained in step S2, which predicts and outputs stress-strain curve feature points uniformly sampled by displacement based on the generated structure, generating stress-strain prediction curves; by minimizing the difference loss between the target curve and the prediction curve, performance-driven structural optimization and parameter update are achieved, generating a metamaterial structure that best matches the target performance in terms of force-displacement response.

[0007] Furthermore, the dataset generation module uses Abaqus in conjunction with Python scripts to perform finite element numerical simulations in batches, obtaining stress-strain curves of irregular metamaterials.

[0008] Furthermore, the finite element numerical simulation is as follows: Two rigid plates are applied at the top and bottom to compress the disordered metamaterial. The top plate is used as a loading plate and moves downward at a constant speed of 15 m / s. All degrees of freedom except for the compression direction are constrained. The bottom plate is a support plate with all displacements restricted. Finite element calculations are performed using plane strain elements. The nodes can only move freely in the plane, while the degrees of freedom outside the plane are restricted. Finally, the nonlinear response curve of the disordered metamaterial under dynamic impact is obtained.

[0009] Furthermore, the nonlinear response curve of the disordered metamaterial under dynamic impact was discretized into 61 points; the stress-strain curve was reconstructed based on the 61 points using spline interpolation.

[0010] Furthermore, delete the first point out of the aforementioned 61 points. Then, normalize the remaining 60 discrete points after deleting the first point.

[0011] Furthermore, all labels are individually normalized to the range [0, 1].

[0012] Furthermore, the multilayer perceptron makes positive predictions of the stress-strain curves of disordered metamaterials as follows: First, the microstructure of the disordered metamaterial is discretized into a 20×20 grid, and the type and orientation information of each element is encoded into a multi-channel vector, which is then normalized and input into the MLP model.

[0013] Furthermore, the MLP model includes an input layer, several hidden layers, and an output layer. ReLU activation and BatchNorm normalization are used to enhance training stability. The output is a curve feature point uniformly sampled according to displacement, which fully represents the force-displacement response of the structure.

[0014] Furthermore, the MLP model employs the SmoothL1 loss function and combines the AdamW optimizer with adaptive learning rate scheduling to achieve fitting and convergence, thereby completing the prediction of the mechanical properties of disordered metamaterials.

[0015] The beneficial effects of this invention are as follows: Addressing the problems of high ambiguity, low performance matching accuracy, and insufficient structural physical rationality in traditional reverse design, this invention constructs a dual-network collaborative framework of a performance evaluation proxy model and a reverse generation model. Through continuously differentiable structural encoding and physical constraint loss design, it achieves an accurate mapping from the target stress-strain curve to the structural distribution. This method significantly improves the model's interpretability, generalization ability, and design efficiency, and can efficiently generate diverse disordered structures that meet the target performance.

[0016] (1) This invention uses deep learning technology to construct a performance-driven disordered metamaterial reverse design model, thereby improving design accuracy and computational efficiency. By introducing a deep learning-based generative network, taking the target stress-strain curve as input, and utilizing the powerful nonlinear fitting capability and parallel computing characteristics of neural networks, an end-to-end mapping between performance and structure is achieved, significantly improving design accuracy and computational efficiency, and reducing dependence on high-cost simulation samples and optimization iterations.

[0017] (2) This invention constructs a multi-objective learning mechanism that integrates physical constraints and continuous encoding to improve the physical rationality and generalization ability of the model. By introducing a continuously differentiable structural encoding method and physical adjacency rule constraints in the model training, the class probability of building blocks is combined with spatial adjacency relationships to construct a composite loss function of performance matching, class frequency and domain consistency, thereby achieving joint optimization of performance-driven and physical constraints. Through this learning mechanism that decouples physical constraints, the model can not only accurately fit the target stress-strain response, but also generate diverse design results with reasonable structure and consistent performance under unknown target conditions, thereby significantly improving the stability, physical rationality and generalization ability of disordered metamaterial reverse design. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the dataset generation module performing finite element numerical simulation on disordered metamaterials in the deep learning-based method for customizing the nonlinear mechanical properties of disordered metamaterials according to the present invention. Figure 2 This is a schematic diagram of the stress-strain curve for predicting disordered metamaterials; Figure 3 This is a schematic diagram of the reverse generative network model architecture; Figure 4 This is a schematic diagram of the microstructure of disordered metamaterials; Figure 5 It is a specific discretization process and coding diagram of the microstructure of disordered metamaterials; Figure 6 This is a schematic diagram of characteristic points in the stress-strain curve. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0020] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0021] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0022] The following combination Figures 1-6 Specific embodiments of the present invention will be described in detail below. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the present invention.

[0023] This invention discloses a deep learning-based method for customizing the nonlinear mechanical properties of disordered metamaterials. This method represents a novel inverse design approach for disordered metamaterials based on performance-driven and physical constraints. By constructing a dual-network collaborative system of a performance evaluation proxy model and an inverse generation model, it achieves high-precision mapping from target stress-strain properties to microstructure space, thereby improving the stability, interpretability, and generalization ability of the inverse design model. This method can achieve efficient and interpretable inverse design under finite sample conditions and generates diverse structural solutions through temperature sampling, making it applicable to fields such as energy absorption, protective structures, and smart materials.

[0024] To explore the diversity of solutions in the high-dimensional latent space, this method employs a temperature-based sampling strategy instead of the deterministic argmax operation. A temperature parameter T is introduced into the Softmax function to adjust the smoothness of the probability distribution: a lower T value makes the distribution sharper, tending towards deterministic selection; while a higher T value makes the distribution flatter, thus increasing randomness through more uniform sampling. This allows the model to generate multiple different but all-satisfying structural solutions for the same mechanical objective, thereby broadening the exploration range of the design space and enhancing generalization ability.

[0025] The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to the present invention includes three core steps: data preparation, forward surrogate model training, and reverse generative network design.

[0026] For information on the generation methods of disordered metamaterials, see [link to documentation]. Figure 4 The entire metamaterial structure is divided into ten sub-regions (p1-p10 below) along the vertical direction. Then, a random probability distribution for the generation of building block types is assigned to each sub-region, and the structures are assembled into the entire disordered metamaterial. The specific method is as follows: S1. Dataset generation module, used to perform finite element calculations on various different disordered metamaterials to obtain multiple corresponding stress-strain curves and form a dataset.

[0027] This invention utilizes the commercial software Abaqus in conjunction with Python scripts to perform batch finite element numerical simulations, obtaining stress-strain curves for disordered metamaterials. A large dataset was obtained from finite element simulations of numerous possible structures. For example... Figure 1 As shown, two rigid plates are applied at the top and bottom to compress the disordered structure. The top plate acts as a loading plate, moving downwards at a constant speed of 15 m / s, with all degrees of freedom constrained except in the compression direction. The bottom plate is a support plate with all displacements restricted. Finite element analysis is performed using plane strain elements, where nodes can only move freely within the plane, while out-of-plane degrees of freedom are restricted. The nonlinear response curve of the irregular metamaterial under dynamic impact is obtained, and the curve is discretized into 61 points. Using spline interpolation, these 61 points are sufficient to reconstruct the stress-strain curve, as shown in the figure. Figure 2 As shown. To better train the neural network, the first point is manually removed, as it equals 0 for any curve. All labels are then individually normalized to the range [0, 1] to ensure the efficiency and accuracy of deep learning model training.

[0028] The finite element method involves discretizing the structure into plane strain elements and nodes, and then performing calculations.

[0029] In this embodiment, 61 discrete points are preferred because this number is sufficient to reconstruct the stress-strain curve. Too many points would lead to a large dataset and increased computational difficulty, while too few points would make it difficult to characterize the stress-strain curve. Since the first point in the finite element method is 0, this point is deleted. The label refers to the remaining 60 discrete points after deletion, and these discrete points are normalized.

[0030] S2. Training a positive prediction model A multilayer perceptron (MLP) was used to positively predict the stress-strain curves of disordered metamaterials: first, the microstructure of the disordered metamaterial was discretized into a 20×20 grid. The microstructure consists of three building blocks that make up the irregular metamaterial. All subsequent units are considered building blocks. The three types of building blocks include Cross-shaped blocks, Diamond-shaped blocks, and T-shaped blocks, with T-shaped blocks exhibiting four deformation directions. The specific discretization process of the disordered metamaterial microstructure is as follows: Figure 5 As shown.

[0031] The type and orientation information of each unit are encoded into a multi-channel vector, normalized, and then input into the MLP model. The MLP model consists of an input layer, several hidden layers, and an output layer, employing ReLU activation and BatchNorm normalization to enhance training stability. The output is a curve feature point uniformly sampled according to displacement, fully representing the force-displacement response of the structure. The feature points are as follows: Figure 6 Orange dots are selected uniformly in the middle. The Smooth L1 loss function is used, combined with the AdamW optimizer and adaptive learning rate scheduling to achieve high-precision fitting and stable convergence, thereby quickly predicting the mechanical properties of disordered metamaterials.

[0032] Figure 4 On the far left are three building blocks, and on the right are their rotations. Only the arrow-shaped (T) building block has rotation; the other two do not have direction. Therefore, as shown... Figure 5 As shown, the T-type block is represented by four numbers: 2, 1, -2, and 1, while the other two types of blocks are represented by 0 and 1.

[0033] In this invention, the MLP model constructs an end-to-end mapping from microscopic topology to macroscopic mechanical properties: First, the initial 20×20 irregular disordered metamaterial microstructure is discretized and numerically encoded, and then normalized into a single-channel feature tensor (1×20×20) as input; subsequently, this tensor is flattened into a 400-dimensional feature vector within the model, and through linear weighting, ReLU nonlinear activation, and Dropout regularization of multiple fully connected layers, the implicit structure-performance correlation features are extracted; finally, the output layer projects it into a one-dimensional vector containing 60 discrete points, which can be visualized and connected to reconstruct a complete stress-strain prediction curve; during the training phase, the network parameters are updated using gradient descent by calculating the Smooth L1 Loss between the prediction vector and the finite element simulation data (i.e., the dataset data), thereby achieving millisecond-level accurate prediction of the mechanical properties of any input structure.

[0034] Smooth L1 loss function Smooth L1 Loss: Represents the smoothing L1 loss function. It is a metric used to measure prediction accuracy; the smaller the value, the more accurate the prediction. x represents the prediction error, which is equal to the difference between the predicted value and the true value. When the prediction error is less than a set threshold, the Smooth L1 loss function is 0.5x. 2 The Smooth L1 loss function, where x is the absolute value of x minus 0.5, represents a quadratic relationship when the prediction error is greater than or equal to a set threshold, and represents a linear relationship. Therefore, gradient descent is used to update the network parameters, enabling millisecond-level accurate prediction of the mechanical properties of any input structure.

[0035] Disordered metamaterials are a class of artificial materials formed by combining different types of building blocks (also known as microstructural units) on periodic or aperiodic substrates. Their macroscopic mechanical properties are determined by their microscopic topological morphology and spatial distribution patterns. For example... Figure 4 As shown, in one specific embodiment, the basic building blocks selected in this invention include six types of units: Cross, Diamond, and four types of T-shaped blocks with different deformation directions. The arrangement and combination of different building blocks in a local area can significantly affect the overall stress-strain response characteristics.

[0036] S3. Constructing a reverse generative network model like Figure 3 As shown, a reverse design framework for disordered metamaterials based on co-optimization of the generator and predictor is presented. The generator adopts an Encoder-Decoder architecture, taking the target stress-strain eigenvector as input and outputting the corresponding disordered metamaterial microstructure distribution; the predictor (i.e., the forward prediction model) predicts the mechanical response curve of the generated structure. By minimizing the difference loss between the target curve and the predicted curve, performance-driven structure optimization and parameter updates are achieved, making the generated metamaterial highly matched with the target performance in terms of force-displacement response.

[0037] The feature vector obtained from the curve is derived from 60 discrete points obtained through uniform sampling. The output corresponds to the disordered metamaterial microstructure distribution; in a specific embodiment, the output is a 20x20 encoded map, which is... Figure 5 The three numbered coded diagrams are shown. Based on the spatial location of the coded diagrams, the corresponding types of building blocks and their directions are filled in to obtain the overall disordered metamaterial.

[0038] To achieve accurate inverse design of disordered metamaterials, this paper constructs a generative network based on an Encoder-Decoder architecture and utilizes a pre-trained Multilayer Perceptron (MLP) as a performance evaluator to build closed-loop supervised training. The specific process is as follows: First, in the generation stage, dimensionality expansion and reconstruction occur. The input 60-dimensional target stress feature vector is mapped and nonlinearly transformed by two fully connected layers of the encoder, and reshaped into a high-dimensional feature tensor (128x8x8) with 128 channels and a spatial size of 8x8. Subsequently, this tensor enters the decoder, and through layer-by-layer transposed convolution or bilinear interpolation upsampling operations, the number of channels is gradually compressed while the spatial resolution is restored. Finally, the feature tensor is reconstructed into a single-channel distribution map of the disordered metamaterial microstructure (20x20). Second, in the MLP-based differential evaluation stage, the generated disordered metamaterial microstructure is fed into the parameter-frozen MLP for forward prediction, and the output is a predicted stress-strain curve with the same dimension as the input. Finally, the Smooth L1 Loss between the predicted curve and the target curve is calculated, and the AdamW optimizer is used to back-update the weight parameters of the encoder and decoder in the generator network based on this difference, thereby achieving an accurate inverse mapping from the low-dimensional performance curve to the high-dimensional topology.

[0039] This method circumvents the ill-conditioned nature of the original problem, namely, the existence of multiple structures exhibiting similar mechanical behaviors in macroscopic responses. It achieves a precise mapping from target performance to manufacturable structures, providing an efficient solution for the intelligent customization of complex mechanical properties.

[0040] This invention proposes a reverse design method for disordered metamaterials based on performance-driven and physical constraints. By constructing a dual-network collaborative mechanism of a performance evaluation proxy model and a reverse generation network, an end-to-end mapping from target stress-strain properties to the microstructure distribution of disordered metamaterials is achieved. This method introduces a continuously differentiable structure encoding mechanism and a physical constraint loss function.

[0041] While maintaining gradient transferability, the model ensures the physical rationality and fabrication feasibility of the generated structure, thereby significantly improving the model's performance matching accuracy, physical interpretability, and generalization ability to unknown targets. The unknown target refers to a randomly defined stress-strain curve. Here, generalization ability is interpreted as the ability to generate corresponding disordered metamaterials for the randomly defined target's stress-strain curve.

[0042] This invention proposes a deep learning-based inverse design method based on performance-driven and physical constraints. Addressing the problems of high ambiguity, low performance matching accuracy, and poor physical rationality in existing disordered metamaterial design, a dual-network collaborative framework of a performance evaluation proxy model and an inverse generation network is constructed. Continuously differentiable structure encoding and a multi-objective physical constraint loss function are introduced to achieve high-precision mapping from the target stress-strain curve to the disordered metamaterial.

[0043] Any process or method described in the flowcharts of this invention or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, which can be implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device. The computer-readable medium can be any medium containing a program for storage, communication, propagation, or transmission for use by the execution system, apparatus, or device, including read-only memory, magnetic disks, or optical disks.

[0044] In the description of this specification, references to terms such as "embodiment," "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 present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, those skilled in the art can combine or combine the different embodiments or examples described in this specification and the features therein without causing contradiction.

[0045] While embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and alterations to the above embodiments within the scope of the present invention.

Claims

1. A method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning, characterized in that, The method includes the following steps: S1. The dataset generation module performs batch finite element numerical simulations to obtain stress-strain curves of disordered metamaterials; S2. Train a positive prediction model and use a multilayer perceptron to make positive predictions on the stress-strain curve of the disordered metamaterial. The output is the stress-strain curve feature points sampled uniformly according to displacement. S3. Construct a disordered metamaterial inverse generation network model based on the collaborative optimization of the generator and predictor; the generator adopts an Encoder-Decoder architecture, inputting the target curve into the generator, which transforms the target curve into a disordered metamaterial microstructure distribution; the predictor is the forward prediction model trained in step S2, which predicts and outputs stress-strain curve feature points uniformly sampled by displacement based on the generated structure, generating stress-strain prediction curves; by minimizing the difference loss between the target curve and the prediction curve, performance-driven structural optimization and parameter update are achieved, generating a metamaterial structure that best matches the target performance in terms of force-displacement response.

2. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 1, characterized in that, The dataset generation module uses Abaqus in conjunction with Python scripts to perform finite element numerical simulations in batches, obtaining stress-strain curves of irregular metamaterials.

3. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 2, characterized in that, The finite element numerical simulation is as follows: Two rigid plates are applied at the top and bottom to compress the disordered metamaterial. The top plate is used as a loading plate and moves downward at a constant speed of 15 m / s. All degrees of freedom except for the compression direction are constrained. The bottom plate is a support plate with all displacements restricted. Finite element calculations are performed using plane strain elements. The nodes can only move freely in the plane, while the degrees of freedom outside the plane are restricted. Finally, the nonlinear response curve of the disordered metamaterial under dynamic impact is obtained.

4. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 3, characterized in that, The nonlinear response curve of the disordered metamaterial under dynamic impact was discretized into 61 points; the stress-strain curve was reconstructed based on the 61 points using spline interpolation.

5. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 4, characterized in that, Delete the first point out of the 61 points mentioned above.

6. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 5, characterized in that, Normalize the remaining 60 discrete points after deleting the first point.

7. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 6, characterized in that, Normalize all labels individually to the range [0, 1].

8. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 1, characterized in that, The multilayer perceptron (MLP) performs positive prediction of the stress-strain curve of disordered metamaterials as follows: First, the microstructure of the disordered metamaterial is discretized into a 20×20 grid, and the type and orientation information of each element is encoded into a multi-channel vector, which is then normalized and input into the MLP model.

9. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 8, characterized in that, The MLP model includes an input layer, several hidden layers, and an output layer. It employs ReLU activation and BatchNorm normalization to enhance training stability. The output consists of curve feature points uniformly sampled according to displacement, which fully characterize the force-displacement response of the structure.

10. The method for customizing the nonlinear mechanical properties of disordered metamaterials based on deep learning according to claim 9, characterized in that, The MLP model employs the Smooth L1 loss function and combines the AdamW optimizer with adaptive learning rate scheduling to achieve fitting and convergence, thereby completing the prediction of the mechanical properties of disordered metamaterials.

Images