Methods, devices, electronic equipment and storage media for predicting the mechanical properties of materials

Abstract

This invention relates to the field of materials mechanics technology, providing a method, device, electronic device, and storage medium for predicting the mechanical properties of materials. The method includes: acquiring the actual physical condition parameters of the structure under test and inputting them into a pre-constructed target constitutive equation to obtain a mechanical property prediction result; the step of determining the target constitutive equation includes: generating multiple candidate constitutive equations using a large language model; performing multi-dimensional evaluation on each candidate constitutive equation and determining a scalar reward value corresponding to each candidate constitutive equation based on the evaluation results; and iteratively updating the parameters of the large language model based on the scalar reward value to determine the target constitutive equation for the target material. This invention drives iterative optimization of the large language model through multi-dimensional evaluation, achieving a deep integration of physical constraints and data-driven discovery, resulting in a final equation with strong physical rationality and high reliability of the mechanical property prediction results.

Description

Technical Field

[0001] This invention relates to the field of materials mechanics, and in particular to a method, apparatus, electronic device, and storage medium for predicting the mechanical properties of materials. Background Technology

[0002] In the fields of materials science and mechanical engineering, establishing constitutive equations describing the stress-strain relationship of materials is a core prerequisite for predicting the mechanical properties of materials and performing finite element simulations of structures. Currently, the establishment of constitutive equations has gradually shifted from traditional manual derivation by domain experts to automated search and discovery using algorithms such as genetic programming or large language models. Existing automated discovery techniques primarily use mathematical data fit as the sole optimization objective, generating candidate equations by fitting experimental data.

[0003] However, because such technologies lack the necessary constraints on physical laws, they often result in insufficient physical rationality in practical applications. For example, the generated equations may have defects such as inconsistent dimensions, mathematical properties that do not match mechanical performance, or excessively complex expressions, making it difficult for the final equations to truly reflect the physical properties of materials and to be effectively applied to complex engineering simulation scenarios. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for predicting the mechanical properties of materials, in order to address the shortcomings of existing automated equation determination techniques, which rely excessively on data fitting, resulting in constitutive equations that lack physical rationality and are difficult to apply to engineering simulations.

[0005] This invention provides a method for predicting the mechanical properties of materials, comprising: Obtain the actual physical working parameters of the structural component under test, which is made of the target material; The actual physical working condition parameters are input into the pre-constructed target constitutive equation to obtain the predicted mechanical properties of the structure under test. The steps for determining the target constitutive equation include: Multiple candidate constitutive equations for characterizing the mechanical properties of the target material are generated using a pre-trained large language model; Based on the stress-strain experimental dataset and variable dimensional information of the target material, a multi-dimensional evaluation is performed on each candidate constitutive equation, and a scalar reward value corresponding to each candidate constitutive equation is determined according to the evaluation results. The multi-dimensional evaluation includes at least physical constraint evaluation and data fitting evaluation. Based on the scalar reward value, the parameters of the large language model are iteratively updated, and the target constitutive equation of the target material is determined according to the large language model after the parameters are iteratively updated.

[0006] According to the material mechanical property prediction method provided by the present invention, the evaluation results include at least two of the following: dimensional consistency verification results, differentiability verification results, goodness of fit, and simplicity. The method involves a multi-dimensional evaluation of each candidate constitutive equation based on the stress-strain experimental dataset and variable dimensional information of the target material, including: Physical property analysis is performed on the candidate constitutive equations to determine the physical constraint verification results of the candidate constitutive equations. The physical constraint verification results include the dimensional consistency verification results and the differentiability verification results. Based on the stress-strain experimental dataset, the undetermined parameters in the candidate constitutive equation are optimized, and the goodness of fit, which reflects the degree of data fit, is calculated. The complexity parameter of the candidate constitutive equation is statistically analyzed to obtain the simplicity that reflects the complexity of the expression.

[0007] According to a method for predicting the mechanical properties of materials provided by the present invention, the step of performing physical property analysis on the candidate constitutive equations and determining the physical constraint verification results of the candidate constitutive equations includes: The candidate constitutive equations are converted into equation strings, and the equation strings are input into a preset symbol analysis module; Based on the symbol analysis module and the variable dimension information, check whether all terms and operations in the equation string conform to the preset dimension rules to obtain the dimension consistency verification result. Based on the symbol analysis module, the continuity of the equation string within the domain is checked to obtain the differentiability verification result.

[0008] According to a method for predicting the mechanical properties of materials provided by the present invention, the step of determining the scalar reward value corresponding to each candidate constitutive equation based on the evaluation results includes: If the dimensional consistency verification result is not consistent, or the differentiability verification result is not differentiable, the preset penalty value is used as the scalar reward value corresponding to the candidate constitutive equation. If the dimensional consistency verification result is consistent and the differentiability verification result is differentiable, the scalar reward value corresponding to the candidate constitutive equation is calculated based on the goodness of fit and the simplicity.

[0009] According to a method for predicting the mechanical properties of materials provided by the present invention, the step of determining the scalar reward value corresponding to each candidate constitutive equation based on the evaluation results includes: The scalar reward value corresponding to the candidate constitutive equation is calculated by weighted summation based on the dimensional consistency verification result, the differentiability verification result, the goodness of fit, and the simplicity.

[0010] According to a method for predicting the mechanical properties of materials provided by the present invention, determining the target constitutive equation of the target material based on the large language model after iterative parameter updates includes: After each update of the parameters of the large language model, candidate constitutive equations that pass the physical constraint evaluation and whose scalar reward value is greater than a preset threshold are selected from the multiple candidate constitutive equations, and the selected candidate constitutive equations are stored in the empirical group storage space. When generating candidate constitutive equations in the next iteration, historical constitutive equations are extracted from the experience group storage space and input as prompts into the large language model to guide the large language model in generating equations. If the number of iterations reaches the preset termination condition, the iteration stops, and the candidate constitutive equation with the highest scalar reward value is extracted from the experience group storage space as the target constitutive equation.

[0011] According to a method for predicting the mechanical properties of materials provided by the present invention, the actual physical condition parameters include load data, temperature data, and boundary condition parameters corresponding to the structure under test. The step of inputting the actual physical condition parameters into a pre-constructed target constitutive equation to obtain the predicted mechanical properties of the structure under test includes: The target constitutive equation is used as the material control parameter of the finite element simulation model, and the actual physical working condition parameters are input into the finite element simulation model so that the finite element simulation model can perform simulation calculations on the structure under test and obtain mechanical simulation calculation results. Based on the mechanical simulation calculation results, the stress distribution data, deformation data, and failure risk analysis data of the structure under test are determined under the actual physical working conditions, and the stress distribution data, deformation data, and failure risk analysis data are used as the mechanical performance prediction results.

[0012] The present invention also provides a material mechanical property prediction device, comprising: The model building module is used to pre-construct the target constitutive equations for the target material. The parameter acquisition module is used to acquire the actual physical working parameters of the structural component under test, which is made of the target material. The performance prediction module is used to input the actual physical condition parameters into the target constitutive equation to obtain the mechanical performance prediction results of the structure under test. Specifically, the model building module is used for: Multiple candidate constitutive equations for characterizing the mechanical properties of the target material are generated using a pre-trained large language model; Based on the stress-strain experimental dataset and variable dimensional information of the target material, a multi-dimensional evaluation is performed on each candidate constitutive equation, and a scalar reward value corresponding to each candidate constitutive equation is determined according to the evaluation results. The multi-dimensional evaluation includes at least physical constraint evaluation and data fitting evaluation. Based on the scalar reward value, the parameters of the large language model are iteratively updated, and the target constitutive equation of the target material is determined according to the large language model after the parameters are iteratively updated.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the material mechanical property prediction method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the material mechanical property prediction method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the material mechanical property prediction method as described above.

[0016] The material mechanical property prediction method, device, electronic device, and storage medium provided by this invention combine the powerful generation capabilities of a pre-trained large language model with a multi-dimensional evaluation system that includes physical constraints and data fitting. This not only enables the system to accurately select constitutive equations that conform to objective physical constraints and take into account the accuracy of experimental data fitting from a massive number of candidate equations, but also drives the large language model to actively learn and internalize physical common sense such as dimensional laws and mathematical smoothness during the generation process through parameter iterative updates based on scalar reward values. This solves the physical absurdity problem that is prone to occur in traditional automated modeling methods, improves the generalization ability and prediction accuracy of the constructed target constitutive equation under complex actual working conditions, and provides a scientific and reliable model support for the mechanical property analysis of the structural components under test. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a schematic diagram of the material mechanical property prediction method provided by the present invention; Figure 2 This is a schematic diagram of the process for determining the target constitutive equation provided by the present invention; Figure 3 This is a flowchart illustrating the method for determining the material constitutive equation based on multi-objective reinforcement learning provided by the present invention. Figure 4 This is a schematic diagram of the material mechanical property prediction device provided by the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] In the fields of materials science and mechanical engineering, establishing constitutive equations that accurately describe the stress-strain relationship of materials is one of the core tasks. Constitutive equations are not only the foundation for understanding the macroscopic mechanical behavior of materials, but also key input models for structural strength assessment, life prediction, and finite element simulation analysis in fields such as aerospace, automotive lightweighting, equipment manufacturing, and civil engineering.

[0021] Traditional constitutive equations are primarily derived through empirical derivation by domain experts. These experts conduct numerous mechanical experiments involving tension, compression, shear, or cyclic loading to obtain raw experimental data, and then manually derive or fit empirical formulas using physical intuition. However, this process is not only time-consuming and labor-intensive, but also highly dependent on the expert's knowledge base, making it difficult to address the nonlinear mechanical characteristics emerging in novel composite materials or complex working conditions.

[0022] In recent years, with the development of artificial intelligence technology, some automated discovery methods have begun to be introduced into this field. For example, genetic programming (GP) algorithms are used to perform heuristic searches within the mathematical expression space, or the powerful knowledge and reasoning capabilities of large language models (LLMs) are used to generate candidate mathematical equations. These techniques have improved the efficiency of constitutive equation discovery to some extent, but they still have the following significant drawbacks in practical engineering applications.

[0023] First, existing technologies typically use data fit (such as mean squared error, MSE) as the sole optimization objective. This makes the system prone to discovering equations that, while mathematically well-fitting experimental curves, are physically absurd. For example, equations often contain errors such as directly adding or subtracting physical quantities with different dimensions (inconsistent dimensions), violating fundamental physical laws.

[0024] Secondly, equations generated purely through data-driven processes often lack constraints on continuity and differentiability. In stress-strain modeling, if the generated equations are discontinuous or non-differentiable within the domain, they cannot be embedded in finite element simulation software for numerical calculations, making it difficult to reflect the true physical response of solid mechanics.

[0025] Furthermore, in pursuit of higher fitting accuracy, systems often generate extremely complex expressions. This overfitting phenomenon not only reduces the model's generalization ability but also causes the equations to lose their simplicity and interpretability, making it difficult for domain experts to understand the underlying mechanical mechanisms.

[0026] To address these shortcomings, this invention provides a method for predicting the mechanical properties of materials. The method can be executed by an electronic device, such as a server, a distributed computing cluster, a personal computer, or a computing module integrated into a materials design and simulation system. In practical applications, the executing entity can flexibly select computing resources based on the type of material to be tested and the scale of the simulation.

[0027] Figure 1 This is a schematic diagram of the material mechanical property prediction method provided by the present invention, as shown below. Figure 1 As shown, the method includes: Step 110: Obtain the actual physical working parameters of the structural component under test, which is made of the target material.

[0028] Specifically, in the field of materials science and mechanical engineering, this invention aims to solve the problem of predicting the mechanical properties of engineering materials such as metallic materials, composite materials, and polymer materials under actual service environments. In this step, the system first needs to define the specific boundary conditions and external force characteristics of the structural component under test (such as lightweight aluminum alloy sheets for automotive bodies). Actual physical condition parameters may include, but are not limited to, the load data (such as pressure, tension, and shear force) experienced by the structural component, ambient temperature data, displacement constraints, strain rate requirements, and specific geometric parameters. These parameters constitute the prerequisites for the material to generate a mechanical response.

[0029] Step 120: Input the actual physical condition parameters into the pre-constructed target constitutive equation to obtain the predicted mechanical properties of the structure under test.

[0030] Specifically, the target constitutive equation refers to a mathematical expression that can accurately describe the stress-strain relationship of the target material. In this embodiment of the invention, the equation is no longer an empirical formula that relies on manual fitting based on expert experience, but rather an optimal solution automatically discovered through an artificial intelligence iterative process described later. By substituting the working condition parameters into the equation, the predicted mechanical properties of the structure under test under a specific stress state can be calculated. These results can be expressed as stress distribution, strain distribution, deformation, yield behavior, hardening characteristics, and failure risk assessment, etc.

[0031] To ensure the reliability of the above prediction results, the focus of this invention is on the determination process of the target constitutive equation. Traditional automatic discovery techniques (such as genetic programming) or purely data-driven LLM generation schemes often only focus on data fit, which can easily lead to equations with inconsistent dimensions or absurd physical meanings. Figure 2 This is a schematic diagram of the process for determining the target constitutive equation provided by the present invention, as shown below. Figure 2 As shown, the steps for determining the target constitutive equation include: Step 210: Generate multiple candidate constitutive equations to characterize the mechanical properties of the target material using a pre-trained large language model.

[0032] Specifically, in this step, LLM leverages its strong knowledge base and logical reasoning capabilities to generate a set of symbolic mathematical equation strings. For example, for the hardening curve of metallic materials, LLM might generate a string of equations in the form of... Candidate constitutive equations ,in and Represents physical variables. These are parameters to be determined.

[0033] Step 220: Based on the stress-strain experimental dataset and variable dimensional information of the target material, perform a multi-dimensional evaluation on each candidate constitutive equation, and determine the scalar reward value corresponding to each candidate constitutive equation according to the evaluation results. The multi-dimensional evaluation includes at least physical constraint evaluation and data fitting evaluation.

[0034] Specifically, the system uses real data obtained from experimental equipment (i.e., stress-strain experimental datasets). Based on the dimensional information of the variables, a multi-dimensional evaluation is performed on each generated candidate constitutive equation. Here, multi-dimensional evaluation refers to scoring the quality of the equation from different perspectives, including at least data fit evaluation and physical constraint evaluation.

[0035] Data fit evaluation refers to the process of evaluating data fit through... Undetermined parameters in Optimization is performed by calculating the error (such as mean square error) between the equation's predicted values ​​and experimental observations, evaluating its ability to reproduce known data, and improving its performance on the dataset. The optimal fit is found above. The result obtained from the data fitting evaluation is denoted as... .

[0036] Physical constraint evaluation is crucial to ensuring the feasibility of a scheme. It involves checking and verifying dimensional consistency and mathematical-physical properties (such as continuity and differentiability). The result of dimensional consistency verification is denoted as... It can be either pass or fail; the result of the differentiability verification is denoted as It can also be either pass or fail.

[0037] Subsequently, a gating reward mechanism can be applied to determine the scalar reward value corresponding to each candidate constitutive equation. The gating reward mechanism introduced in this embodiment of the invention has a veto power attribute, meaning that physical constraints are set as hard indicators. If an equation is determined to be inconsistent in the dimensional check (i.e., the dimensional consistency verification fails), or if it contains physical absurdities such as non-differentiability / discontinuity within its domain, the system will directly assign it a fixed penalty value. (e.g., -10); only when the equation is physically reasonable will its final scalar reward value be calculated based on weighted indices such as goodness of fit and simplicity. .

[0038] In addition to using a gating reward mechanism (hard constraint) to determine the scalar reward value corresponding to each candidate constitutive equation, a weighted summation (soft constraint) method can also be used. That is, a weighted summation is performed based on the dimensional consistency verification results, differentiability verification results, goodness of fit, simplicity, etc., to calculate the scalar reward value of the candidate constitutive equation.

[0039] Step 230: Based on the scalar reward value, iteratively update the parameters of the large language model, and determine the target constitutive equation of the target material according to the large language model after iteratively updating the parameters.

[0040] Specifically, after calculating the scalar reward value, the system uses a policy optimization algorithm (such as GRPO or PPO) to guide the LLM using the scalar reward value as a feedback signal. During this reinforcement learning adaptation process, the LLM gradually learns to avoid expressions that violate physical common sense and actively generates equations that can accurately fit the data and are reasonable in terms of dimensions and physical properties. Upon reaching the termination condition, the best-performing equation is selected from the generated sequence as the target constitutive equation.

[0041] The method provided in this invention combines the powerful generation capabilities of a pre-trained large language model with a multi-dimensional evaluation system that includes physical constraints and data fitting. This not only enables the system to accurately select constitutive equations from a massive pool of candidate equations that conform to objective physical constraints and take into account the accuracy of experimental data fitting, but also drives the large language model to actively learn and internalize physical common sense such as dimensional laws and mathematical smoothness during the generation process through parameter iterative updates based on scalar reward values. This solves the physical absurdity problem that is easily generated by traditional automated modeling methods, improves the generalization ability and prediction accuracy of the constructed target constitutive equation under complex actual working conditions, and provides a scientific and reliable model support for the mechanical performance analysis of the structural components under test.

[0042] Based on any of the above embodiments, the evaluation results include at least two of the following: dimensional consistency verification results, differentiability verification results, goodness of fit, and simplicity; correspondingly, in step 220, the multi-dimensional evaluation of each candidate constitutive equation based on the stress-strain experimental dataset and variable dimensional information of the target material includes: Step 221: Perform physical property analysis on the candidate constitutive equations to determine the physical constraint verification results of the candidate constitutive equations. The physical constraint verification results include the dimensional consistency verification results and the differentiability verification results.

[0043] Specifically, this process aims to determine whether candidate constitutive equations violate fundamental physical laws. The results of physical constraint verification are examined primarily from two dimensions: dimensional consistency (…). This involves checking whether all terms and operations in the equation (such as addition, subtraction, multiplication, division, exponentiation, logarithm, etc.) fully conform to pre-defined rules of physical quantity dimensions, for example, stress units cannot be directly added to strain units; secondly, it involves verifying differentiability. This involves checking whether the equations are continuous and differentiable everywhere within the domain of material mechanics (such as the normal strain interval), because most solid mechanical behaviors are macroscopically continuous and smooth. The verification results for these two dimensions are usually expressed as Boolean values ​​of pass / fail or satisfy / dissatisfy.

[0044] Step 222: Optimize the undetermined parameters in the candidate constitutive equation based on the stress-strain experimental dataset, and calculate the goodness of fit, which reflects the degree of data fit.

[0045] Specifically, since LLM often generates expression structures containing symbolic parameters (such as...) This step will utilize numerical optimization algorithms (such as least squares or gradient descent) on the experimental dataset. The optimal parameter set is found. Based on this, the goodness of fit, which reflects the degree of data fit, is calculated. This indicator can be characterized by the reciprocal of the mean squared error (MSE) or other statistical correlation coefficients, and is used to measure the accuracy of the equation in numerical prediction.

[0046] Step 223: Calculate the complexity parameters of the candidate constitutive equations to obtain the simplicity that reflects the complexity of the expression.

[0047] Specifically, in materials science, overly complex equations (such as expressions with dozens of operators) often imply overfitting and a lack of interpretability. Therefore, this invention evaluates whether an equation conforms to the principle of simplicity by statistically analyzing complexity parameters such as the number of operators, the number of variables, and the nesting level within the equation string. The simpler the equation, the higher its simplicity. The higher the score, the better.

[0048] Based on any of the above embodiments, step 221 specifically includes: Step 2211: Convert the candidate constitutive equation into an equation string and input the equation string into a preset symbol analysis module; Step 2212: Based on the symbol analysis module and combined with the variable dimension information, check whether all terms and operations in the equation string conform to the preset dimension rules to obtain the dimension consistency verification result. Step 2213: Based on the symbol analysis module, check the continuity of the equation string within the domain to obtain the differentiability verification result.

[0049] Specifically, since the raw output generated by the large language model is usually text conforming to mathematical writing conventions, the system first extracts and normalizes it into an equation string for subsequent rigorous logical verification. This string is then sent to the symbolic analysis module. Here, the symbolic analysis module is the core component in this embodiment of the invention for processing symbolic mathematical operations. It can be built based on a symbolic computation library and has the ability to perform syntax parsing, term separation, differentiation, and simplification of mathematical expressions. Through this module, complex equations are parsed into an operational syntax tree, allowing for independent checking of each operator and operand.

[0050] Next, the system uses a symbolic analysis module, combined with variable dimensional information, to check whether all terms and operations in the equation string conform to preset dimensional rules, thus obtaining a dimensional consistency verification result. In practice, the system pre-stores a variable dimensional information table, recording the dimensional attributes of all physical quantities involved in the equation and constant terms. The symbolic analysis module traverses the syntax tree: checking addition and subtraction operations, requiring that the physical quantities on both sides of the operator have exactly the same dimension; checking transcendental functions (such as exp, ln, sin, etc.), requiring that their internal parameters be dimensionless after elimination; checking both sides of the equal sign, requiring that the combined dimension of the output on the left (such as stress) and the calculated result on the right be consistent. If any node in the syntax tree violates the above preset dimensional rules, the dimensional consistency verification is deemed unsatisfactory, i.e., the dimensional consistency verification result is a failure.

[0051] Subsequently, the system uses the symbolic analysis module to check the continuity of the equation string within its domain to obtain the differentiability verification result. In materials mechanics modeling, constitutive equations must be able to reflect the smooth response of the material under load. The symbolic analysis module will analyze candidate constitutive equations... For the independent variable in the equation (such as strain) Perform sign differentiation to check within the actual service domain of the material (e.g., strain range). The LLM check determines whether the equation has mathematical singularities, discontinuities, or non-differentiable points. For example, if the equation generated by LLM contains terms with denominators that may be zero or operations that take even roots of negative numbers, and this occurs within the domain, the equation will be considered non-differentiable. A successful check is only output if the equation is continuous throughout the entire domain and has a first-order (or higher) derivative.

[0052] This invention, through the introduction of a symbolic analysis module, transforms abstract physical laws into computer-executable syntax tree checking rules, thereby automating the checks on dimensional consistency and mathematical smoothness. This mechanism ensures that every candidate constitutive equation entering the subsequent fitting stage is logically sound, thus reducing resource waste caused by invalid computations and guaranteeing that the final determined constitutive equation can be directly applied to engineering simulation scenarios.

[0053] Based on any of the above embodiments, step 220, determining the scalar reward value corresponding to each candidate constitutive equation based on the evaluation results, includes: If the dimensional consistency verification result is not consistent, or the differentiability verification result is not differentiable, the preset penalty value is used as the scalar reward value corresponding to the candidate constitutive equation. If the dimensional consistency verification result is consistent and the differentiability verification result is differentiable, the scalar reward value corresponding to the candidate constitutive equation is calculated based on the goodness of fit and the simplicity.

[0054] Specifically, based on the aforementioned multi-dimensional evaluation indicators, this embodiment of the invention employs a core gating reward mechanism for scalarization. If the dimensional consistency verification result is non-consistent (i.e., dimensional consistency check fails), or the differentiability verification result is non-differentiable (i.e., differentiability check fails), a preset penalty value will be directly applied. ) as scalar reward value This means that physical constraints have veto power. For example, [the following text is incomplete and likely refers to a separate topic:] Set it to a large negative value (e.g., -10). Even if an equation has an extremely high fit to the data points, the system will severely punish it if its dimensions are incorrect or if it contains mathematical singularities, thereby shielding it from interference from physically absurd solutions.

[0055] If the dimensional consistency verification result is consistent (i.e., passes the dimensional consistency check) and the differentiability verification result is differentiable (i.e., passes the differentiability check), then the normal evaluation logic is entered, which calculates the scalar reward value based on the goodness of fit and simplicity. .

[0056] At this point, the scalar reward value is usually calculated using a weighted summation method, i.e. By adjusting the weighting coefficients and This can guide the model to find the optimal constitutive equation that can both accurately predict mechanical behavior and is simple and intuitive enough, while ensuring physical correctness.

[0057] The method provided in this invention introduces a gating reward mechanism with veto power, which forcibly embeds physical common sense into the feedback loop of reinforcement learning. This not only ensures the physical logic rigor of the discovered equations, but also improves the engineering practical value and interpretability of the equations by checking their simplicity. This effectively solves the problem that traditional algorithms generate complex and absurd equations.

[0058] Based on any of the above embodiments, step 220, determining the scalar reward value corresponding to each candidate constitutive equation based on the evaluation results, includes: The scalar reward value corresponding to the candidate constitutive equation is calculated by weighted summation based on the dimensional consistency verification result, the differentiability verification result, the goodness of fit, and the simplicity.

[0059] Specifically, in some optional embodiments, a soft constraint mechanism (weighted summation strategy) can be used to replace the aforementioned gated hard constraint mechanism for evaluating and calculating the reward of candidate constitutive equations. Specifically, the system no longer assigns veto power to physical constraints, but instead maps multi-dimensional evaluation results numerically. For example, for dimensional consistency verification results... and differentiability verification results If the candidate constitutive equation is fully satisfied, it is mapped to 100 points (or 1.0); if it is partially satisfied or not satisfied at all, it is assigned a continuous score between [0, 100] or 0 points based on the degree to which it violates physical common sense. For goodness of fit... and simplicity Similarly, normalization is performed to obtain the corresponding score.

[0060] Subsequently, the system obtains the pre-configured weight parameters for the above four evaluation dimensions, denoted as follows: , , as well as The weighted summation formula is used to calculate the first... The final scalar reward value of each candidate constitutive equation : Understandably, in actual engineering deployments, the aforementioned weight parameters are not static and can be carefully tuned to balance exploration and exploitation. For example, in the early exploration phase of reinforcement learning training, the weight of physical constraints can be appropriately reduced (e.g., by lowering the weight of the physical constraints). and This approach increases the weights of goodness of fit and simplicity. This setting allows large language models to make mistakes in early explorations. Because some equations with minor physical flaws may have highly promising mathematical skeletons (e.g., containing excellent nonlinear fitting terms), directly rejecting them with hard constraints could lead the system into local optima. Allowing these potentially flawed equations to persist and participate in the next iteration increases the diversity of the model's solution space exploration.

[0061] As the number of iterations increases, the system can dynamically increase the weight parameters of physical constraints, forcing the large language model to gradually converge to regions that conform to physical common sense in later generation.

[0062] Based on any of the above embodiments, step 230, determining the target constitutive equation of the target material according to the large language model after iterative parameter updates, includes: After each update of the parameters of the large language model, candidate constitutive equations that pass the physical constraint evaluation and whose scalar reward value is greater than a preset threshold are selected from the multiple candidate constitutive equations, and the selected candidate constitutive equations are stored in the empirical group storage space. When generating candidate constitutive equations in the next iteration, historical constitutive equations are extracted from the experience group storage space and input as prompts into the large language model to guide the large language model in generating equations. If the number of iterations reaches the preset termination condition, the iteration stops, and the candidate constitutive equation with the highest scalar reward value is extracted from the experience group storage space as the target constitutive equation.

[0063] Specifically, after updating the parameters in each round of reinforcement learning, the system selects candidate constitutive equations from multiple candidates that pass the physical constraint evaluation and have a scalar reward value greater than a preset threshold. This involves a dual selection logic: the first selection requires the candidate constitutive equation to completely pass the aforementioned dimensional consistency and differentiability verifications, proving its physical existence; the second selection is based on its scalar reward value. The result must be higher than a preset threshold to indicate that the equation has demonstrated high quality in terms of fitting accuracy and simplicity.

[0064] Subsequently, the system stores the selected candidate constitutive equations in the empirical population storage space. This storage space is responsible for storing all physically plausible and high-performing operator combinations and equation structures discovered during the historical search process. This space is dynamically updated with each iteration, replacing older, lower-scoring equations with newer, higher-scoring ones, thus maintaining the advanced nature of the empirical population.

[0065] Next, when generating candidate constitutive equations for the next time, the system extracts historical constitutive equations from the experience group storage space and inputs them as prompts into the large language model. This step fully utilizes the contextual learning capabilities of the large language model. By transforming these excellent historical constitutive equations into prompts, the LLM, when generating the next batch of candidate equations, no longer blindly generates them from scratch, but can perform variations, combinations, or logical transfers based on these successful examples, thereby improving the directionality and efficiency of the search.

[0066] This invention constructs an experience group storage space and transforms it into LLM prompts, forming a closed-loop experience accumulation and knowledge guidance mechanism. This enables the model to learn from beneficial experiences in historical searches, thereby accelerating the exploration speed of high-potential constitutive equations and ensuring the stability and convergence of the search process.

[0067] Furthermore, the system monitors the current iteration status in real time. It stops iterating when the preset termination condition is reached. This preset termination condition can be set in several ways: one is the maximum number of iterations (e.g., 50 or 100 loops) to ensure reasonable allocation of computational resources; another is a reward convergence condition, where the model is considered to have discovered the optimal equation for the dataset when the growth rate of the highest reward value in the experience population falls below a certain minimum value for several consecutive generations, thus terminating the iteration prematurely.

[0068] Once the iteration stops, the system will extract the candidate constitutive equation with the highest scalar reward value from the experience population storage space. This serves as the final objective constitutive equation. Since the empirical population storage space always holds candidate solutions that are optimal throughout the generations, the final extracted equation... It must be the mathematical description that best performs in terms of physical constraints, fitting accuracy, and expression simplicity. This equation will be output and embedded in subsequent material mechanical property prediction modules or finite element simulation modules, serving as the core algorithm for characterizing the mechanical properties of materials.

[0069] Based on any of the above embodiments, the actual physical condition parameters include the load data, temperature data, and boundary condition parameters corresponding to the structure under test; correspondingly, step 120 specifically includes: Step 121: Use the target constitutive equation as the material control parameter of the finite element simulation model, and input the actual physical working condition parameters into the finite element simulation model so that the finite element simulation model can perform simulation calculations on the structure under test and obtain mechanical simulation calculation results. Step 122: Based on the mechanical simulation calculation results, determine the stress distribution data, deformation data, and failure risk analysis data of the structural component under test under the actual physical working conditions, and use the stress distribution data, deformation data, and failure risk analysis data as the mechanical performance prediction results.

[0070] Specifically, after determining the target constitutive equation, the system will complete the closed loop from mathematical formula to engineering prediction through the following steps: First, clarify the actual physical operating condition parameters. In specific application fields such as aerospace, automotive lightweighting, equipment manufacturing, and civil engineering, the service environment of the structural components under test (such as automotive body panels, aero-engine blades, etc.) is extremely complex. The operating condition parameters obtained in this step are multi-dimensional, including load data (such as specific values ​​or force curves for tension, compression, shear, and cyclic loading), temperature data (reflecting the thermodynamic response of materials under high-temperature or variable-temperature environments), and boundary condition parameters (such as displacement constraints of the structural component, fixed support positions, contact friction coefficients, etc.). These parameters together constitute a stress environment model of the structural component in the real physical world.

[0071] Secondly, the optimal constitutive equation discovered in the aforementioned steps A material subroutine or material property definition module recognizable by the simulation software is written. This equation, serving as a material control parameter, specifies the stress-strain evolution logic that each mesh element in the simulation model should follow under stress. Subsequently, the collected loads, temperatures, and boundary conditions are applied to the simulation model as external inputs, and the simulation calculation is initiated. Because this constitutive equation has passed rigorous dimensional verification and differentiability checks, the simulation model exhibits extremely high numerical stability and convergence during complex nonlinear iterative solutions, and can realistically simulate the material response under different loads, temperatures, or strain rates.

[0072] Finally, the system extracts key data from the output file of the finite element solver. The mechanical property prediction results are represented in the spatial dimension as stress distribution data (e.g., stress concentration areas in the stamping process of automotive sheet metal) and deformation data (e.g., springback, degree of local thinning, etc.); in the failure determination dimension, they are represented as failure risk analysis data (e.g., cracking risk, yield behavior, or degree of hardening determined based on stress-strain state).

[0073] Taking the stamping analysis of lightweight aluminum alloy sheets for automotive bodies as an example, this invention utilizes constitutive equations that are automatically discovered and meet the requirements of dimensional consistency and continuous differentiability. This allows for highly accurate prediction of stress concentration points, springback, local thinning, and cracking risks in body panels during the stamping process. Compared to empirical equations obtained solely through data fitting, which may contain physical flaws, the prediction results obtained by this invention significantly improve the reliability and realism of the simulation results due to the high degree of consistency between its mathematical logic and physical common sense.

[0074] This invention, through deep coupling of the target constitutive equation with mature finite element simulation technology, achieves an automated transition from experimental data to engineering decisions, ensuring the accuracy of material mechanical response prediction under complex working conditions.

[0075] Based on any of the above embodiments Figure 3 This is a flowchart illustrating the method for determining the material constitutive equation based on multi-objective reinforcement learning provided by the present invention, as shown below. Figure 3 As shown, the process for determining and applying the material constitutive equation provided in this embodiment of the invention is as follows: S1: Input and initialization.

[0076] The system first acquires basic input data, including: (1) stress-strain experimental dataset. (1) It records the actual mechanical response of the target material under different experimental conditions; (2) The dimensional information of the variables is used for subsequent physical validity verification; (3) The pre-trained Large Language Model (LLM) serves as the policy network for generating equations. In the initialization phase, the system will also establish an experience group storage space to dynamically record excellent individual equations that emerge during the iteration process.

[0077] S2: Generate candidate constitutive equations.

[0078] Viewing LLM as a policy network, it generates a set of policies based on prompts (such as...). (Number) candidate symbolic mathematical equations For example, the generated candidate equations are in the form of ,in Represents response, Represents stress, These are parameters to be determined. This step leverages the powerful logical combination capabilities of LLM, which can automatically explore an extremely wide space of mathematical operators.

[0079] S3: Multi-dimensional assessment.

[0080] For each generated candidate equation For multi-dimensional evaluation and verification: (a) Parameter fitting: using numerical optimization algorithms to fit the parameters. Undetermined parameters in Optimize it to fit the dataset The optimal fit is achieved, and the goodness of fit is calculated. .

[0081] (b) Physical analysis: The equation string is sent to the symbol analysis module, and the dimensional rules are used to check whether the terms conform to dimensional consistency. (Pass / Fail).

[0082] (c) Mathematical property check: Analyzing the continuity and differentiability of the equations within the service domain of the material. (pass / fail), and calculate the conciseness of the expression ( (a numerical value).

[0083] In addition, the multi-dimensional assessment here and It can be replaced or supplemented according to the actual application scenario. For example, when discovering an energy equation, an energy conservation check can be added; when discovering a dynamic equation, a symmetry or boundary condition check can be added.

[0084] S4: Gated reward calculation.

[0085] This step combines multi-dimensional evaluation metrics into a single scalar reward signal. This invention employs hard-constraint gating logic in its embodiments: if the equation For failure or If it fails, it is directly judged as physically absurd and a fixed penalty value is imposed. Only equations that pass physical verification are used to calculate positive reward values ​​based on their fitting accuracy and simplicity.

[0086] S5: Strategy Adaptation.

[0087] Using the above scalar rewards As a feedback signal for reinforcement learning, gradients are calculated and the LLM model parameters are updated using policy optimization algorithms (such as GRPO or PPO).

[0088] S6: Group Update and Output.

[0089] The system selects equations that pass the physics check and have high reward values ​​in the current iteration and stores them in the experience population. In subsequent iterations, these excellent equations will serve as prompts to guide the LLM to perform deeper derivations. When the preset number of iterations or convergence conditions are reached, the system extracts the target constitutive equation with the highest composite reward from the experience population. As a final finding.

[0090] After obtaining the target constitutive equation, this embodiment of the invention applies it to actual production. Specifically, the actual physical operating parameters of the structure under test (such as stress load, ambient temperature, and boundary constraints) can be obtained, and the target constitutive equation is embedded into the finite element simulation software as the core material model. Through simulation calculation, the stress distribution, deformation, and failure risk prediction results of the structure under specific operating conditions are finally output.

[0091] Furthermore, the method provided in this embodiment of the invention is not limited to stress-strain modeling, but can also be widely applied to any field that requires discovering mathematical equations that meet specific constraints from data, such as fluid mechanics, heat transfer, biochemistry, and financial modeling.

[0092] The material mechanical property prediction device provided by the present invention is described below. The material mechanical property prediction device described below can be referred to in correspondence with the material mechanical property prediction method described above.

[0093] Based on any of the above embodiments Figure 4 This is a schematic diagram of the material mechanical property prediction device provided by the present invention, as shown below. Figure 4 As shown, the device includes: Model building module 410 is used to pre-build the target constitutive equation of the target material; The parameter acquisition module 420 is used to acquire the actual physical working condition parameters of the structural component under test, which is made of the target material. The performance prediction module 430 is used to input the actual physical condition parameters into the target constitutive equation to obtain the mechanical performance prediction results of the structure under test. Specifically, the model building module 410 is used for: Multiple candidate constitutive equations for characterizing the mechanical properties of the target material are generated using a pre-trained large language model; Based on the stress-strain experimental dataset and variable dimensional information of the target material, a multi-dimensional evaluation is performed on each candidate constitutive equation, and a scalar reward value corresponding to each candidate constitutive equation is determined according to the evaluation results. The multi-dimensional evaluation includes at least physical constraint evaluation and data fitting evaluation. Based on the scalar reward value, the parameters of the large language model are iteratively updated, and the target constitutive equation of the target material is determined according to the large language model after the parameters are iteratively updated.

[0094] The apparatus provided in this invention combines the powerful generation capabilities of a pre-trained large language model with a multi-dimensional evaluation system that includes physical constraints and data fitting. This not only enables the system to accurately select constitutive equations from a massive pool of candidate equations that conform to objective physical constraints and take into account the accuracy of experimental data fitting, but also drives the large language model to actively learn and internalize physical common sense such as dimensional laws and mathematical smoothness during the generation process through parameter iterative updates based on scalar reward values. This solves the physical absurdity problem that is easily generated by traditional automated modeling methods, improves the generalization ability and prediction accuracy of the constructed target constitutive equation under complex actual working conditions, and provides a scientific and reliable model support for the mechanical performance analysis of the structural components under test.

[0095] Based on any of the above embodiments, the evaluation results include at least two of the following: dimensional consistency verification results, differentiability verification results, goodness of fit, and simplicity; the model building module includes a multi-dimensional evaluation unit, which is used for: Physical property analysis is performed on the candidate constitutive equations to determine the physical constraint verification results of the candidate constitutive equations. The physical constraint verification results include the dimensional consistency verification results and the differentiability verification results. Based on the stress-strain experimental dataset, the undetermined parameters in the candidate constitutive equation are optimized, and the goodness of fit, which reflects the degree of data fit, is calculated. The complexity parameter of the candidate constitutive equation is statistically analyzed to obtain the simplicity that reflects the complexity of the expression.

[0096] Based on any of the above embodiments, the multidimensional evaluation unit is further used for: The candidate constitutive equations are converted into equation strings, and the equation strings are input into a preset symbol analysis module; Based on the symbol analysis module and the variable dimension information, check whether all terms and operations in the equation string conform to the preset dimension rules to obtain the dimension consistency verification result. Based on the symbol analysis module, the continuity of the equation string within the domain is checked to obtain the differentiability verification result.

[0097] Based on any of the above embodiments, the multidimensional evaluation unit is further used for: If the dimensional consistency verification result is not consistent, or the differentiability verification result is not differentiable, the preset penalty value is used as the scalar reward value corresponding to the candidate constitutive equation. If the dimensional consistency verification result is consistent and the differentiability verification result is differentiable, the scalar reward value corresponding to the candidate constitutive equation is calculated based on the goodness of fit and the simplicity.

[0098] Based on any of the above embodiments, the multidimensional evaluation unit is further used for: The scalar reward value corresponding to the candidate constitutive equation is calculated by weighted summation based on the dimensional consistency verification result, the differentiability verification result, the goodness of fit, and the simplicity.

[0099] Based on any of the above embodiments, the model building module is specifically used for: After each update of the parameters of the large language model, candidate constitutive equations that pass the physical constraint evaluation and whose scalar reward value is greater than a preset threshold are selected from the multiple candidate constitutive equations, and the selected candidate constitutive equations are stored in the empirical group storage space. When generating candidate constitutive equations in the next iteration, historical constitutive equations are extracted from the experience group storage space and input as prompts into the large language model to guide the large language model in generating equations. If the number of iterations reaches the preset termination condition, the iteration stops, and the candidate constitutive equation with the highest scalar reward value is extracted from the experience group storage space as the target constitutive equation.

[0100] Based on any of the above embodiments, the actual physical operating condition parameters include the load data, temperature data, and boundary condition parameters corresponding to the structure under test; correspondingly, the performance prediction module is specifically used for: The target constitutive equation is used as the material control parameter of the finite element simulation model, and the actual physical working condition parameters are input into the finite element simulation model so that the finite element simulation model can perform simulation calculations on the structure under test and obtain mechanical simulation calculation results. Based on the mechanical simulation calculation results, the stress distribution data, deformation data, and failure risk analysis data of the structure under test are determined under the actual physical working conditions, and the stress distribution data, deformation data, and failure risk analysis data are used as the mechanical performance prediction results.

[0101] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a material mechanical property prediction method. This method includes: acquiring actual physical condition parameters of a structural component under test, the structural component being composed of a target material; inputting the actual physical condition parameters into a pre-constructed target constitutive equation to obtain a predicted mechanical property result for the structural component under test; wherein the step of determining the target constitutive equation includes: generating multiple candidate constitutive equations characterizing the mechanical properties of the target material using a pre-trained large language model; performing multi-dimensional evaluation on each candidate constitutive equation based on the stress-strain experimental dataset and variable dimensional information of the target material, and determining a scalar reward value corresponding to each candidate constitutive equation based on the evaluation results, the multi-dimensional evaluation including at least physical constraint evaluation and data fitting evaluation; iteratively updating the parameters of the large language model based on the scalar reward value, and determining the target constitutive equation of the target material based on the iteratively updated parameters of the large language model.

[0102] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the material mechanical property prediction method provided by the above methods. The method includes: acquiring actual physical condition parameters of a structural component under test, the structural component under test being composed of a target material; inputting the actual physical condition parameters into a pre-constructed target constitutive equation to obtain a mechanical property prediction result of the structural component under test; wherein, the step of determining the target constitutive equation includes: generating multiple candidate constitutive equations for characterizing the mechanical properties of the target material using a pre-trained large language model; performing multi-dimensional evaluation on each candidate constitutive equation based on the stress-strain experimental dataset and variable dimensional information of the target material, and determining a scalar reward value corresponding to each candidate constitutive equation according to the evaluation results, the multi-dimensional evaluation including at least physical constraint evaluation and data fitting evaluation; iteratively updating the parameters of the large language model based on the scalar reward value, and determining the target constitutive equation of the target material based on the large language model after iterative parameter updates.

[0104] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for predicting the mechanical properties of materials provided by the methods described above. This method includes: acquiring actual physical condition parameters of a structural component under test, the structural component being composed of a target material; inputting the actual physical condition parameters into a pre-constructed target constitutive equation to obtain a predicted mechanical property result for the structural component under test; wherein the step of determining the target constitutive equation includes: generating multiple candidate constitutive equations characterizing the mechanical properties of the target material using a pre-trained large language model; performing a multi-dimensional evaluation on each candidate constitutive equation based on the stress-strain experimental dataset and variable dimensional information of the target material, and determining a scalar reward value corresponding to each candidate constitutive equation based on the evaluation results, the multi-dimensional evaluation including at least physical constraint evaluation and data fitting evaluation; iteratively updating the parameters of the large language model based on the scalar reward value, and determining the target constitutive equation of the target material based on the large language model after iterative parameter updates.

[0105] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0106] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of predicting mechanical properties of a material, characterized by, include: Obtain the actual physical working parameters of the structural component under test, which is made of the target material; The actual physical working condition parameters are input into the pre-constructed target constitutive equation to obtain the predicted mechanical properties of the structure under test. The steps for determining the target constitutive equation include: Multiple candidate constitutive equations for characterizing the mechanical properties of the target material are generated using a pre-trained large language model; Based on the stress-strain experimental dataset and variable dimensional information of the target material, a multi-dimensional evaluation is performed on each candidate constitutive equation, and a scalar reward value corresponding to each candidate constitutive equation is determined according to the evaluation results. The multi-dimensional evaluation includes at least physical constraint evaluation and data fitting evaluation. Based on the scalar reward value, the parameters of the large language model are iteratively updated, and the target constitutive equation of the target material is determined according to the large language model after the parameters are iteratively updated.

2. The method of predicting mechanical properties of a material according to claim 1, wherein The evaluation results include at least two of the following: dimensional consistency verification results, differentiability verification results, goodness of fit, and simplicity. The evaluation of each candidate constitutive equation based on the stress-strain experimental dataset and variable dimensional information of the target material involves multiple dimensions, including: Physical property analysis is performed on the candidate constitutive equations to determine the physical constraint verification results of the candidate constitutive equations. The physical constraint verification results include the dimensional consistency verification results and the differentiability verification results. Based on the stress-strain experimental dataset, the undetermined parameters in the candidate constitutive equation are optimized, and the goodness of fit, which reflects the degree of data fit, is calculated. The complexity parameter of the candidate constitutive equation is statistically analyzed to obtain the simplicity that reflects the complexity of the expression.

3. The method of predicting mechanical properties of a material according to claim 2, wherein The step of performing physical property analysis on the candidate constitutive equations and determining the physical constraint verification results of the candidate constitutive equations includes: The candidate constitutive equations are converted into equation strings, and the equation strings are input into a preset symbol analysis module; Based on the symbol analysis module and the variable dimension information, check whether all terms and operations in the equation string conform to the preset dimension rules to obtain the dimension consistency verification result. Based on the symbol analysis module, the continuity of the equation string within the domain is checked to obtain the differentiability verification result.

4. The method for predicting the mechanical properties of materials according to claim 2, characterized in that, The step of determining the scalar reward value corresponding to each candidate constitutive equation based on the evaluation results includes: If the dimensional consistency verification result is not consistent, or the differentiability verification result is not differentiable, the preset penalty value is used as the scalar reward value corresponding to the candidate constitutive equation. If the dimensional consistency verification result is consistent and the differentiability verification result is differentiable, the scalar reward value corresponding to the candidate constitutive equation is calculated based on the goodness of fit and the simplicity.

5. The method for predicting the mechanical properties of materials according to claim 2, characterized in that, The step of determining the scalar reward value corresponding to each candidate constitutive equation based on the evaluation results includes: The scalar reward value corresponding to the candidate constitutive equation is calculated by weighted summation based on the dimensional consistency verification result, the differentiability verification result, the goodness of fit, and the simplicity.

6. The method for predicting the mechanical properties of materials according to claim 1, characterized in that, Determining the target constitutive equation of the target material based on the large language model after iterative parameter updates includes: After each update of the parameters of the large language model, candidate constitutive equations that pass the physical constraint evaluation and whose scalar reward value is greater than a preset threshold are selected from the multiple candidate constitutive equations, and the selected candidate constitutive equations are stored in the empirical group storage space. When generating candidate constitutive equations in the next iteration, historical constitutive equations are extracted from the experience group storage space and input as prompts into the large language model to guide the large language model in generating equations. If the number of iterations reaches the preset termination condition, the iteration stops, and the candidate constitutive equation with the highest scalar reward value is extracted from the experience group storage space as the target constitutive equation.

7. The method for predicting the mechanical properties of materials according to any one of claims 1 to 6, characterized in that, The actual physical condition parameters include the load data, temperature data, and boundary condition parameters corresponding to the structure under test. The step of inputting the actual physical condition parameters into the pre-constructed target constitutive equation to obtain the predicted mechanical properties of the structure under test includes: The target constitutive equation is used as the material control parameter of the finite element simulation model, and the actual physical working condition parameters are input into the finite element simulation model so that the finite element simulation model can perform simulation calculations on the structure under test and obtain mechanical simulation calculation results. Based on the mechanical simulation calculation results, the stress distribution data, deformation data, and failure risk analysis data of the structure under test are determined under the actual physical working conditions, and the stress distribution data, deformation data, and failure risk analysis data are used as the mechanical performance prediction results.

8. A device for predicting the mechanical properties of materials, characterized in that, include: The model building module is used to pre-construct the target constitutive equations for the target material. The parameter acquisition module is used to acquire the actual physical working parameters of the structural component under test, which is made of the target material. The performance prediction module is used to input the actual physical condition parameters into the target constitutive equation to obtain the mechanical performance prediction results of the structure under test. Specifically, the model building module is used for: Multiple candidate constitutive equations for characterizing the mechanical properties of the target material are generated using a pre-trained large language model; Based on the stress-strain experimental dataset and variable dimensional information of the target material, a multi-dimensional evaluation is performed on each candidate constitutive equation, and a scalar reward value corresponding to each candidate constitutive equation is determined according to the evaluation results. The multi-dimensional evaluation includes at least physical constraint evaluation and data fitting evaluation. Based on the scalar reward value, the parameters of the large language model are iteratively updated, and the target constitutive equation of the target material is determined according to the large language model after the parameters are iteratively updated.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the material mechanical property prediction method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the material mechanical property prediction method as described in any one of claims 1 to 7.

Images