A tire rubber material modulus optimization method and system based on a large language model (LLM)

Abstract

The present application relates to the technical field of intelligent tire design, and particularly relates to a tire pattern automatic generation method and system based on an LLM model. In view of the problems of strong coupling of rubber modulus of a plurality of components such as a tread, a base rubber, and a sidewall, and low efficiency of traditional finite element trial and error, the present application first parameterizes and defines the modulus of each component within an engineering feasible interval, obtains rolling resistance samples and constructs a data set based on a two-dimensional / three-dimensional finite element rolling model of a tire under a given speed and load; performs structured numerical fitting and supervised training on the LLM, establishes reverse mapping from the modulus vector, inputs a target output candidate in an inference stage, obtains the target output candidate through interval constraint and finite element recalculation verification, and forms a closed loop iteration through back-feeding fine tuning. The scheme can quickly reverse the modulus combination and ensure physical consistency, improve the efficiency and accuracy of achieving the rolling resistance target, and can be extended to multi-performance index optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent tire design technology, and in particular to a method and system for optimizing the modulus of tire rubber materials based on Large Language Model (LLM). Background Technology

[0002] Rolling resistance of radial tires is closely related to vehicle energy consumption, range, and carbon emissions. Rolling resistance primarily originates from the periodic deformation of the tire during rolling contact with the ground and the viscoelastic energy dissipation of the rubber material. It is influenced by tire structural design, rubber material formulation, and the material parameters of various components. In engineering practice, to reduce rolling resistance, researchers typically work on multiple component material systems, including the tread compound, base compound, sidewall compound, inner liner, and tread rubber. By adjusting the material modulus and energy dissipation characteristics, they alter the strain energy distribution and heat generation level, thereby reducing rolling resistance while maintaining durability and handling. However, tires are multi-material, multi-layered composite structures, and the moduli of the materials in each component are strongly coupled: a change in the modulus of a single component can cause a chain reaction of changes in the contact patch, stress-strain field, belt end strain energy, and thermo-mechanical coupling state, leading to a complex trade-off between rolling resistance and performance indicators such as durability, wear, and noise. Therefore, how to quickly determine the effective mapping between the multi-component modulus combination and the target rolling resistance within the feasible parameter range of engineering, and to realize reverse design oriented towards the target index, is a key technical problem in the synergistic optimization of tire materials and structures.

[0003] In existing technologies, a common approach is forward optimization based on finite element simulation: a three-dimensional finite element model of the tire structure is established with boundary conditions such as road surface / drum test conditions. Batch simulation calculations are performed by changing material or structural parameters, and then the optimal solution is selected using evaluation indicators such as rolling resistance, strain energy density, and stress-strain at key points. This approach can reflect the impact of material coupling of multiple tire components on performance, but it often relies on a large number of trial calculations. Especially when multiple components are simultaneously adjusted, the parameter combination space grows exponentially, leading to a significant increase in simulation computation and engineering time costs.

[0004] Chinese patent (CN120297079A) discloses a simulation optimization method, device, and system for matching the modulus of multiple rubber compounds in tires. This scheme constructs a finite element model of the tire and road surface, and uses orthogonal experimental design to perform finite element calculations on the modulus properties of different rubber compounds / mesh matching in the tire simulation model, extracting data such as stress, strain, and strain energy. Further, it calculates modulus characteristic values ​​to obtain the optimal modulus ratio of the multiple rubber compounds. This technology demonstrates the application of orthogonal experimental design and finite element simulation in matching the modulus of multiple rubber compounds, which can improve the efficiency and usability of simulation optimization to a certain extent. However, such methods are still generally forward selection: they require pre-enumerating or designing a large number of modulus combinations and simulating them one by one, ultimately selecting a better solution from the sample set. When the target expands from optimal stress / strain energy to a specified rolling resistance target value, there is still a lack of a reverse inference capability that can directly output the modulus combination that satisfies the target. At the same time, orthogonal experimental methods have limited coverage of high-dimensional continuous parameter spaces, and the sample size will still expand rapidly when facing more complex working conditions (speed, load, temperature, etc.).

[0005] In addition, there are also intelligent methods for material parameter inversion. Chinese patent (CN116522510B) discloses a method and system for inverting tire material parameters for wheel performance simulation. This scheme establishes a finite element simulation model based on tire radial and lateral stiffness test data. Through error calculation and optimization processes, the simulation curves are made to closely match the test curves. A neural network model is introduced to achieve rapid material parameter inversion using non-destructive stiffness test data. This type of technology illustrates that in the tire field, a parameter inversion approach combining experimental / simulation data with machine learning / neural networks has emerged. However, in terms of applicable objects, their inversion targets are mostly concentrated on stiffness test curves and related material parameter calibration, emphasizing the improvement of simulation accuracy and test curve fitting, rather than directly generating material modulus combinations for rolling resistance target values ​​in reverse. In terms of model expression, traditional neural networks usually use fixed-dimensional numerical inputs and outputs. For scenarios with multiple components, multiple working conditions, and complex sample structures, the model's expansion and generalization capabilities are still greatly affected by the scale of training data, feature engineering, and network structure selection. In terms of engineering closed loop, although it includes iterative optimization, it mostly uses the minimum error as the convergence target, lacking an efficient inverse design closed loop mechanism that is oriented towards a given target index → ​​multiple solution candidates → physical verification → self-learning backfeed.

[0006] On the other hand, Chinese patent (CN117715771A) discloses a system and method for real-time estimation of tire rolling resistance. This solution utilizes sensors such as tire liner or valve stem to obtain real-time signals, and combines this with information such as tire steady-state values ​​(obtained through drum testing or finite element analysis) and wear conditions to estimate or correct rolling resistance in real time. This type of technology focuses on rolling resistance estimation and condition correction during vehicle use, which is helpful for online monitoring and model compensation of rolling resistance. However, its core is not the reverse engineering of material modulus during the R&D stage. Even with the introduction of finite element analysis, it mainly serves steady-state benchmarks or calibration, and it is difficult to directly solve the problem of how to quickly deduce the combination of multi-component moduli from the target rolling resistance during the R&D stage.

[0007] However, the above technologies generally have the following shortcomings: (1) Lack of reverse generation capability for target indicators: Most schemes still mainly rely on enumeration / design parameters → simulation → screening, making it difficult to directly output multi-component modulus combinations that meet the target under given target rolling resistance conditions; (2) Insufficient efficiency of high-dimensional continuous parameter space: When optimizing multi-component modulus at the same time, the parameter dimension is high, and discrete designs such as orthogonal experiments have limited coverage of continuous space, and the simulation cost increases significantly with the increase of dimension; (3) Insufficient model generalization and closed-loop self-learning: Traditional neural networks mostly rely on fixed features and data scale, and when facing multiple working conditions, multiple constraints, and multiple solution output requirements, it is difficult to simultaneously take into account feasibility, diversity, and rapid iteration; (4) Insufficient unity of physical verification and model update chain: Existing schemes often separate the simulation calculation model training / calibration scheme output into multiple links, lacking a unified closed-loop process of data representation, reverse reasoning, finite element recalculation verification, and incremental update.

[0008] Therefore, there is an urgent need for a tire rubber material modulus optimization technology that can utilize the authenticity of finite element simulation data and combine it with stronger generation and mapping expression capabilities (such as LLM adaptation capability for structured numerical data) to achieve rapid reverse design of target rolling resistance → modulus combination under engineering feasibility constraints, and achieve continuous iterative optimization through finite element recalculation and data backfeedback, thereby reducing R&D trial and error costs, shortening development cycle and improving the stability of target achievement. Summary of the Invention

[0009] The technical objective of this invention is to address the problems of strong coupling of modulus parameters of rubber materials in multiple tire components, low optimization efficiency due to reliance on numerous finite element calculations, and difficulty in achieving reverse design of target indicators. This invention provides a material modulus optimization method and system based on a closed-loop fusion of Large Language Model (LLM) and finite element rolling simulation data. Under engineering feasibility constraints, this method can quickly deduce and output the optimal or alternative modulus combinations of rubber materials for each component from a given rolling speed, load condition, and target rolling resistance value. Furthermore, simulation verification and data feedback iterative updates improve prediction accuracy and optimization stability.

[0010] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution: Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution: A method for optimizing the modulus of tire rubber materials based on Large Language Modeling (LLM) includes the following steps: S1, Obtain the set of tire rubber components, which includes multiple components to be optimized, each component to be optimized having a corresponding component number; Set the allowable range of elastic modulus for each component to be optimized; S2, under rolling speed and load conditions, establish a tire finite element rolling model, generate N sets of modulus vectors in batches, each set of modulus vectors includes the elastic modulus of each component to be optimized, and obtain the rolling resistance corresponding to each set of modulus vectors through finite element simulation; S3, construct the dataset, in which each sample includes rolling speed, load, rolling resistance and the corresponding modulus vector; S4. The LLM is adapted to structured numerical input and output, and an inverse mapping model is trained based on the dataset. This enables the inverse mapping model to output a predictive modulus vector based on the input rolling speed, load, and rolling resistance. S5. During the inference phase, the target rolling resistance and target operating conditions are input. The target operating conditions include the target speed and target load. The inverse mapping model outputs candidate modulus vectors and constrains the candidate modulus vectors to the allowable range of the elastic modulus of each component to be optimized, thus obtaining the constrained modulus vectors. The constrained modulus vectors are input into the tire finite element rolling model for finite element recalculation to obtain the verification rolling resistance. When the absolute value of the difference between the verification rolling resistance and the target rolling resistance is not greater than the allowable error threshold, the constrained modulus vector is output as the optimization result. Otherwise, the target speed, target load, verification rolling resistance, and constrained modulus vector are added to the dataset as refeed samples, and the inverse mapping model is updated.

[0011] Preferably, in step S1, the component to be optimized includes at least two of the following: tread, base rubber, sidewall, carcass, inner liner, and triangular rubber. And / or, in step S2, the finite element rolling model includes a two-dimensional finite element mesh or a three-dimensional finite element mesh, and ground boundary conditions and rolling boundary conditions are set; the finite element simulation is one of static rolling simulation or dynamic rolling simulation; And / or, in step S2, N sets of modulus vectors are generated using one of orthogonal experimental design, Latin hypercube sampling, or Sobol sequence sampling, and a convergence flag is recorded for each set of simulation results. The convergence flag is used to indicate the convergence status of the corresponding set of simulations.

[0012] Preferably, in step S3, the dataset is divided into a training set, a validation set, and a test set, with a ratio of 90%:5%:5%. The training set is used to train the inverse mapping model, the validation set is used for model selection, and the test set is used for generalization evaluation.

[0013] Preferably, in step S4, the structured numerical input-output adapter uses a numerical quantization step size to discretize the rolling speed, load, rolling resistance and elastic modulus of each component to be optimized into a labeled sequence input LLM, wherein the numerical quantization step size is the quantization resolution. Alternatively, in step S4, the structured numerical input-output adapter uses a numerical embedding function to map the rolling speed, load, rolling resistance and elastic modulus of each component to be optimized into vectors and then inputs them into the LLM. The numerical embedding function is a mapping function from continuous numerical values ​​to the embedding space. And / or, in step S4, the loss function for training the inverse mapping model is obtained by weighted summation of the modulus prediction error term and the constraint violation penalty term; the modulus prediction error term is the average of the squared differences between the predicted modulus and the true modulus of each component to be optimized; the constraint violation penalty term is used to penalize model outputs that do not meet the allowable range of elastic modulus or the set of engineering constraints; the set of engineering constraints is a set of constraints used to limit the candidate modulus vectors to meet the requirements of tire structural safety, deformation range and engineering manufacturability.

[0014] Preferably, the set of engineering constraints includes at least a structural stress upper limit constraint and / or a deformation upper limit constraint; the structural stress upper limit constraint is used to limit the equivalent stress obtained from finite element simulation to not exceed the maximum allowable equivalent stress; the deformation upper limit constraint is used to limit the displacement or deformation obtained from finite element simulation to not exceed the maximum allowable displacement or maximum allowable deformation.

[0015] Preferably, in step S5, the inverse mapping model outputs M groups of candidate modulus vectors, where M is the number of candidates. Each group of candidate modulus vectors has a corresponding candidate index, which is numbered sequentially from 1 to M. The candidate modulus vectors are generated using one of beam search, Top-k sampling, or temperature sampling. The retention count in Top-k sampling is used to limit the number of candidates retained during sampling, and the temperature parameter in temperature sampling is used to control the randomness of the generation of candidate modulus vectors. The retention count and the temperature parameter are used together to control the diversity of candidate solutions. And / or, in step S5, the constraint projection operator is used to constrain the candidate modulus vector, which is used to map the candidate modulus vector into a feasible modulus vector that simultaneously satisfies the allowable range of elastic modulus and the set of engineering constraints. And / or, in step S5, the conditions for triggering the refeed update include at least the number of new samples reaching the threshold for the number of new samples that triggers the update, or the verification error exceeding the error threshold that triggers the update.

[0016] Preferably, the LLM is a numerical generation LLM for inverse generation of tire multi-component modulus, comprising: (1) Numerical embedding module, used to map input working conditions and component indices into embedding vectors, wherein the input working conditions include rolling speed, load and rolling resistance; (2) Component alignment encoding module, used to establish a one-to-one correspondence between each component to be optimized and its corresponding output modulus. The component alignment encoding module adopts position encoding or alignment encoding. (3) A continuous regression decoder head is used to directly output a continuous modulus vector, the continuous modulus vector including the prediction modulus of each component to be optimized, and the continuous regression decoder head is used to map the hidden state of LLM into a continuous value. Furthermore, the LLM undergoes supervised learning during the training phase using samples containing rolling speed, load, rolling resistance, and corresponding modulus vectors, thereby obtaining the ability to inversely generate modulus combinations from the target resistance.

[0017] Secondly, a tire rubber material modulus optimization system based on large language model LLM includes: a parameterization and constraint module, a finite element simulation module, a dataset management module, an LLM training and inference module, and a physical verification and backfeeding iteration module. The parameterization and constraint module is used to execute step S1; The finite element simulation module is used to execute step S2; The dataset management module is used to execute step S3; The LLM training and inference module is used for the output of candidate modulus vectors in steps S4 and S5. The physical verification and backfeed iteration module is used to perform constraint processing, finite element recalculation, optimization result output and data backfeed update in step S5.

[0018] Thirdly, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0019] Fourthly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the method.

[0020] This invention combines tire finite element rolling simulation data with the inverse generation capability of Large Language Model (LLM), changing the traditional forward trial-and-error mode of tire rubber material modulus optimization, which involves manually setting modulus combinations, performing finite element simulations group by group, and selecting the best results. It can directly deduce candidate modulus combinations for multiple rubber components, such as the tread, base rubber, sidewall, carcass, inner liner, and triangular rubber, after inputting target rolling resistance, target speed, and target load. This significantly reduces the number of blind simulations and improves the efficiency of material parameter matching. Because this invention establishes an inverse mapping relationship between working conditions, rolling resistance, and modulus vectors using rolling resistance samples generated by the finite element rolling model during the training phase, LLM can not only process textual information but also predict continuous modulus parameters for multiple components through structured numerical adaptation, numerical embedding, and continuous regression output. This solves the problem of insufficient generalization ability of traditional neural networks or experimental design methods in high-dimensional, multi-component, and strongly coupled parameter spaces.

[0021] Furthermore, this invention does not simply rely on the model output results. Instead, after the LLM outputs the candidate modulus vectors, it constrains them to the allowable range of elastic modulus of each component and the engineering constraints, and then feeds them back into the tire finite element rolling model for recalculation and verification. The usability of the candidate modulus combination is determined by verifying the error between the rolling resistance and the target rolling resistance. This LLM reverse reasoning—constraint processing—finite element recalculation—error judgment process ensures that the model output results have physical consistency and engineering feasibility, avoiding invalid solutions that, although the model predictions are reasonable, do not meet the material manufacturing window, structural safety constraints, or rolling simulation verification requirements.

[0022] Meanwhile, this invention employs a backfeeding iteration mechanism. When the verified rolling resistance fails to reach the allowable error threshold, the target working condition, the recalculated verified rolling resistance, and the constrained modulus vector are added as new samples to the dataset and used to update the inverse mapping model. This allows the model's local mapping relationship near the target rolling resistance to be continuously corrected and converged. Therefore, this invention can continuously accumulate effective samples under different tire structures, different speed load conditions, and different material windows, improving the model's prediction accuracy and stability for low rolling resistance target regions, and forming a continuously optimized closed-loop design system. Attached Figure Description

[0023] Figure 1 Tire material distribution diagram.

[0024] Figure 2 Component name and two-dimensional finite element mesh diagram.

[0025] Figure 3 Three-dimensional finite element model diagram.

[0026] Figure 4This is a flowchart of an LLM-based method for optimizing the modulus of tire rubber materials.

[0027] Figure 5 A schematic diagram of the model structure for LLM structured numerical adaptation and reverse generation.

[0028] Figure 6 This is a schematic diagram of closed-loop verification and reinjection fine-tuning. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0030] I. Terminology Explanation

[0031] 1. Tire rubber component assembly This refers to the collection of rubber components in a tire that participate in optimizing the material modulus. Common components include the tread, base rubber, sidewall, carcass rubber, inner liner, protective rubber, and tread rubber. For component serial number, For the number of parts, .

[0032] 2. Elastic modulus Modulus range : Indicates components The elastic modulus used in finite element simulation (preferably in MPa). , These represent the lower and upper limits of the engineering feasibility for the modulus of the component, respectively.

[0033] 3. Modulus vector : An ordered vector representing the modulus of each component: ;in, It is a modulus vector; For the first Components The modulus; This refers to the number of components.

[0034] 4. Parameters for rolling operation : The rolling speed (preferably in km / h) is the rolling speed. This is the normal load (preferably in N). In implementation, the operating condition dimension can be further expanded, such as tire pressure. ,temperature Road surface roughness wait.

[0035] 5. Rolling resistance Rolling resistance (preferably in N) refers to the rolling resistance of a tire under given operating conditions, which can be obtained through finite element simulation or testing. This invention primarily uses finite element simulation output.

[0036] 6. Sample number With sample size : Indicates the first Index of group modulus combination samples ; This indicates the number of data sample groups used for training / validation / testing.

[0037] 7. Dataset Sample Structure: The dataset consists of multiple sets of samples, each preferably represented as an input-output pair:

[0038] ;

[0039] in, For the first Sample speed; For the first Sample payload; For the first Sample rolling resistance; For the first Sample modulus vector; angle brackets This represents the input / output pair of a sample.

[0040] 8. Target rolling resistance With the allowable error threshold : The desired rolling resistance; This is the allowable error threshold used to determine whether the target is met.

[0041] 9. LLM and Inverse Mapping Model LLM stands for Large Language Model or its structured numerical adaptation version. This represents the model mapping function used for inverse reasoning after training, enabling the inverse generation of the modulus vector from the target resistance and operating conditions:

[0042] ;

[0043] in, The modulus vector predicted by the model; This is the inverse mapping model after training; Input operating conditions and rolling resistance.

[0044] 10. Number of candidate solutions With candidate number During the inference phase, the model can output multiple sets of candidate modulus vectors. For the number of candidates, Candidate number, .

[0045] 11. Constrained modulus vector With verification of rolling resistance : This represents the result after constraining the prediction modulus vector to the feasible interval and engineering constraints. Indicates to Verification of rolling resistance obtained through finite element recalculation.

[0046] II. Overview of the System Structure of the Invention

[0047] At the system implementation level, this invention is preferably deployed as a modular optimization system, the system structure of which can be abstracted into the following modules (which can be software modules in the same computing device, or multiple nodes deployed in a distributed manner):

[0048] Parameterization and Constraints Module: Responsible for component assembly Definition, Modulus Range Set and maintain engineering constraint rules.

[0049] Finite element modeling and simulation module: responsible for establishing tire finite element models, performing batch simulation calculations, and outputting rolling resistance. And necessary intermediate fields (grounding mode, stress and strain, convergence information, etc.).

[0050] Dataset Management Module: Responsible for structured storage of samples, training / validation / test partitioning, version management, sample merging and deduplication during backfeeding.

[0051] LLM numerical adaptation and training module: responsible for converting structured numerical values ​​into an LLM-processable representation and training the inverse mapping model. 1. Perform model evaluation and export.

[0052] Reasoning and Loop Validation Module: Responsible for inputting the target Operating conditions Generate candidate modulus vectors, perform constraint processing and finite element recalculation verification, output the final modulus combination, and trigger refeed fine-tuning when the error threshold is not met.

[0053] III. Specific Technical Route for Implementing the Method of the Invention

[0054] like Figure 4 As shown, to facilitate engineering implementation, this embodiment breaks down the method flow into five steps, S1–S5, and provides reproducible implementation details and parameter selection suggestions for each step. The overall approach is as follows:

[0055] Step S1: Parametric Definition and Constraint Modeling (Engineering Window Confirmation)

[0056] Step S2: Finite element batch simulation sampling (sample generation)

[0057] Step S3: Dataset preparation and quality control (formation of trainable data)

[0058] Step S4: LLM Structured Numerical Adaptation and Supervised Training (Inverse Mapping Learning)

[0059] Step S5: Target rolling resistance reverse reasoning, finite element recalculation verification, and backfeed iteration (closed-loop optimization).

[0060] IV. Specific implementation methods of steps S1–S5

[0061] (I) Step S1, Parametric Definition and Constraint Modeling

[0062] Step S1.1, Component set definition and number mapping

[0063] Combination Figure 1 The tire material distribution shown is preferably such that the tire structure is divided into several components according to material regions. In one specific embodiment, a total of rubber-related components are taken. Examples of components include: carcass rubber, protective rubber, inner liner, base rubber, crown belt rubber, second belt layer rubber, first belt layer rubber, tread rubber, sidewall rubber, and triangular rubber. Reinforcing skeleton layers (such as bead wire, belt layer skeleton, carcass skeleton, etc.) can be used as fixed material parameters and are not considered optimization variables in this invention; however, optimization variables (such as optimizing the equivalent modulus or volume fraction of the cord layer) can be included in extended embodiments. In this embodiment, it is preferred to keep them fixed to reduce dimensionality.

[0064] To ensure reproducibility, it is preferable to establish a component name-component serial number system. —A mapping table of finite element partition IDs and material card IDs, for example:

[0065] : Carcass rubber, corresponding to FE zone ID=1, material card ID=Rubber Carcass;

[0066] : Protective coating, corresponding to FE partition ID=2, material card ID=Rubber Chafer.

[0067] This mapping table is used for batch assignment of values ​​by automated scripts and for structured recording of datasets to avoid human error.

[0068] Step S1.2, setting the modulus range and the set of engineering feasibility constraints

[0069] For each component Set the adjustable range of the modulus The range can be determined from the following sources: historical design databases, material testing (DMA / tensile small strain range), capability boundaries of formulation and vulcanization systems, or enterprise standard windows.

[0070] In one embodiment, initial reference values ​​(e.g., 4 MPa for the tread and 3.5 MPa for the base rubber) can be given based on the rubber material modulus table for finite element analysis, and the relative change ratio can be taken. Formation interval:

[0071] ;

[0072] in, For components Lower limit of modulus; The upper limit; For reference modulus; To allow for relative changes (e.g., 0.2 represents ±20%).

[0073] In addition to interval constraints, it is preferable to define a set of engineering constraints. It is used to eliminate unusable solutions during the reasoning phase. It may include, but is not limited to:

[0074] 1) Structural safety constraints: equivalent stress in critical areas Not exceeding the allowed limit Strain in key components No more than ;

[0075] 2) Deformation constraint: Maximum radial deformation under rolling conditions No more than ;

[0076] 3) Simulation solvability constraints: Solver convergence flags ;

[0077] 4) Manufacturing feasibility constraints: The formulation / vulcanization window corresponding to the modulus combination can be achieved (which can be determined by empirical rules or regression models).

[0078] Engineering constraints can be implemented using threshold rules or scoring functions. For ease of subsequent loop closure, threshold rules are preferred.

[0079] (ii) Step S2, finite element batch simulation sampling

[0080] Step S2.1, Finite Element Model Establishment

[0081] Combination Figure 2 and Figure 3This step involves establishing a two-dimensional / three-dimensional finite element model of the tire. The two-dimensional model can be used for preliminary mesh and partitioning verification, while the three-dimensional model is used for rolling simulation to calculate rolling resistance. Key aspects of the model include:

[0082] Geometry and Partitioning: Partition by component to ensure each Corresponding to a unique material domain;

[0083] Material Card: Rubber Area Assignment; the reinforcing layer is assigned a value based on a fixed equivalent modulus (e.g., 21000MPa for steel wire coils);

[0084] Boundary conditions: Apply tire pressure (if necessary), apply normal load. Apply rolling speed Define the grounding contact and friction model;

[0085] Solution type: Steady-state rolling, quasi-static rolling, and explicit dynamic rolling can be used, depending on the software platform (ABAQUS, ANSYS, MSC.Marc, etc.).

[0086] Step S2.2, Sampling Strategy and Sample Generation

[0087] To generate training data, batch generation is required. Group Modulus Vector And obtain through simulation The following sampling methods are preferred:

[0088] Orthogonal experimental design: suitable for rapid coverage of low / medium dimensions;

[0089] Latin hypercube sampling: suitable for high-dimensional continuous spatial coverage;

[0090] Sobol sequences: suitable for low-difference uniform coverage.

[0091] Each set of sample outputs must contain at least: , , , Convergence indicators Simulation version number .in Whether convergence has occurred can be represented by 0 or 1; This is for consistency during subsequent power-on.

[0092] (III) Step S3, Dataset Preparation and Quality Control

[0093] Step S3.1, Structured Data Format

[0094] To facilitate training and data tracing, it is preferable to use a uniform structured format to store data, such as JSONL or tables. Recommended fields for each sample:

[0095] part_names: List of component names (and...) (One-to-one correspondence)

[0096] E: Modulus list (corresponding) )

[0097] V: Speed

[0098] F: Load

[0099] R: Rolling resistance

[0100] conv: convergence

[0101] solver_ver: Simulation version

[0102] mesh_tag: Mesh tag (optional).

[0103] Step S3.2, Data Cleaning and Partitioning

[0104] The following quality control measures are preferred:

[0105] 1) Delete non-convergent samples ( ) or marked as low weight;

[0106] 2) To Outliers should be statistically tested (e.g., exceeding the mean ± 3 times the standard deviation) and manually verified.

[0107] 3) If duplicates exist Alternatively, repeat the operating conditions, retain the latest version, or perform averaging (depending on the simulation noise level).

[0108] Then the dataset Divide into training sets proportionally. Validation set With test set The preferred ratio is 90%:5%:5%. A layered strategy can be adopted when dividing the range to ensure that different rolling resistance ranges or different operating condition ranges are covered.

[0109] (iv) Step S4, LLM structured numerical adaptation and supervised training

[0110] like Figure 5As shown, step S4 transforms the traditional LLM's sequence modeling capability for natural language into inverse mapping learning capable of handling structured continuous numerical values, and outputs modulus vectors that correspond one-to-one with tire components. This step not only needs to be trainable, but also needs to be stable, have controllable output, and be reproducible in engineering. Therefore, this implementation provides a structured disclosure from six aspects: input representation, model structure, training objective, training process, hyperparameter selection, evaluation, and derivation.

[0111] Step S4.1, Input-output representation (numerical serialization scheme A: quantization marker sequence)

[0112] In one implementation, continuous numerical values and Mapping to a discrete labeled sequence allows LLM to generate symbolic modulus buckets without an additional numerical regression head, which are then decoded into continuous values. Its advantages are simplicity and stability; its disadvantage is that the resolution is limited by the quantization step size.

[0113] Define quantization step size These are used for quantizing speed, load, rolling resistance, and modulus, respectively. The quantization function can be:

[0114] ;

[0115] in, For quantization operators; The value to be quantified; This is the quantization step size; This indicates rounding to the nearest integer.

[0116] Define a tag template (example only):

[0117] Input sequence: V_ <qv> F_ <qf> R_ <qr>

[0118] Output sequence: E1_ <qe1> E2_ <qe2>...In_ <qen>;

[0119] in , , , .

[0120] Decoding continuous values: in, These are continuous values ​​obtained through decoding; Number the buckets; This corresponds to the quantization step size.

[0121] To ensure that the decoded value still falls within the range A clipping operation can be performed after decoding:

[0122] ;

[0123] in, The modulus after trimming; These are the upper and lower limits of the interval; This is the decoding modulus.

[0124] Step S4.2, Input-output representation (numerical embedding scheme B: continuous numerical embedding + regression decoding head)

[0125] In another preferred embodiment, the LLM is numerically embedded and structurally adapted to continuous regression output, enabling the model to directly output a continuous modulus vector. This avoids quantization errors and is suitable for high-precision engineering applications.

[0126] 4.2.1 Numerical Embedding

[0127] Each consecutive numerical value is mapped to an embedding vector. A multilayer perceptron (MLP) numerical embedding function can be used. :

[0128] ;

[0129] in, For numerical values The embedding vector; For numerical embedding functions; For non-linear activation functions (such as GELU / ReLU); This is the weight matrix; This is the bias vector.

[0130] To maintain scale consistency, input Prioritize normalization (standardization):

[0131] ;

[0132] in, Normalized value; The mean of this variable in the training set; The standard deviation is denoted as .

[0133] For input Embedded respectively This information, along with the component location code, is then sent into the LLM backbone.

[0134] 4.2.2 Component Alignment and Position Coding

[0135] To ensure a one-to-one correspondence between the output modulus and the components, a component index is introduced. Alignment encoding (This can be a learnable embedding). For example:

[0136] ;

[0137] in, For the first Component position / alignment encoding; This is the encoding function.

[0138] Conceptually, the input embeddings are combined into a sequence of tokens:

[0139] token0: Work status token, content is... fused vector

[0140] token : No. Component token, content is .

[0141] The fusion method can be linear projection after stitching:

[0142] ;

[0143] in, This is the working condition fusion vector; For fusion parameters; This indicates vector concatenation.

[0144] 4.2.3 LLM Backbone and Regression Decoder

[0145] The LLM backbone can be a Transformer decoder structure. Inputting the above token sequence into the backbone yields the hidden state of each component token. Then use the regression decoder head. Output continuous modulus:

[0146] ;

[0147] in, To predict the modulus; For the first Component hidden state; This is the regression weight vector; For bias.

[0148] To ensure the output falls within the range, a bounded mapping (preferably a sigmoid mapping to the range) can be used after the decoder:

[0149] ;

[0150] in, This is the output of the unconstrained regression. It is the sigmoid function; These are the upper and lower limits of the interval.

[0151] This interval mapping method is smoother than simple clipping, which helps stabilize training and naturally satisfies interval constraints.

[0152] Step S4.3, Training Objective and Loss Function (Supervised Learning)

[0153] The training objective of this invention is to learn the inverse mapping: input Output Supervised learning is employed, with the basic loss being the mean squared error (MSE). Under continuous regression scheme B, the optimal loss function is:

[0154] ;

[0155] in, This is the mean square error loss; Number of components; To predict the modulus; This is the true modulus.

[0156] To further constrain project feasibility, a penalty term can be added to form the total loss:

[0157] ;

[0158] in, Total loss; For penalty weighting; This is the penalty function for engineering constraints.

[0159] penalty function Can be based on constraint sets Definition. For example, regarding stress and deformation constraints:

[0160] ;

[0161] in, The equivalent stress index is obtained from the fast proxy model or the previous round of recalculation; The upper limit; Displacement / deformation index; This is the upper limit. If it's inconvenient to obtain data in real-time during the training phase... This option can be temporarily disabled or a simple, calculable proxy constraint can be used (e.g., relative order of modulus, stricter upper and lower limits for key components, etc.); then, strict constraints can be imposed by finite element recalculation in the closed loop of step S5.

[0162] Step S4.4, Training Process (Data Organization, Batch Processing, Optimizer, Stopping Criteria)

[0163] 4.4.1 Training Data Organization

[0164] Will Each sample organization in the input With tags If serialization scheme A is used, then and All are token sequences; if embedding scheme B is used, then It is a continuous numerical vector.

[0165] 4.4.2 Batch Processing and Mini-Batch

[0166] Mini-batch training is used, with each batch size being [size missing]. For numerical regression tasks, Options include 32, 64, or 128. For stable training, it is recommended to start with a smaller amount. Start it up, then gradually increase the size.

[0167] 4.4.3 Optimizer and Learning Rate Strategy

[0168] The AdamW optimizer is preferred, and a learning rate is used. With weight decay Cosine annealing or piecewise decay can be used. The learning rate can be updated in the following form (illustrated):

[0169] ;

[0170] in, For the first Step learning rate; The initial maximum learning rate; Minimum learning rate; This is the current training step; This represents the total number of steps.

[0171] 4.4.4 Early Termination and Model Selection

[0172] In the validation set Upload monitoring to verify errors or When continuous No improvement was observed in any assessment cycle (decline was less than the threshold). Stop early to save the optimal model parameters. Among them, To reduce the patience value for early cessation; The minimum improvement threshold.

[0173] Step S4.5, Hyperparameter Selection Recommendations

[0174] To facilitate implementation, this invention provides a feasible hyperparameter suggestion window (not constituting a limitation):

[0175] LLM backbone layers : 6–24 (6–12 can be selected for smaller data volumes, and 12–24 can be selected for larger data volumes)

[0176] Hidden Dimensions :256–1024

[0177] Number of attention heads :4–16

[0178] Learning rate : arrive

[0179] Weight decay : 0.01–0.1

[0180] Batch size : 32–128

[0181] Number of training rounds 10–200 (combined with early stop)

[0182] Interval mapping sigmoid temperature (optional): Used to control the smoothness of the output distribution.

[0183] When the dataset size When the number of data points is small (e.g., a few thousand), it is recommended to use a smaller model, strong regularization, and early stopping to avoid overfitting; when... For larger datasets (e.g., tens of thousands of data points), a larger model can be used to improve the fitting ability.

[0184] Step S4.6, Model Evaluation Metrics

[0185] In addition to MSE, it is recommended to also assess:

[0186] Mean Absolute Error (MAE):

[0187] ;

[0188] in, Mean absolute error; Number of components; The meaning is the same as before.

[0189] Interval compliance rate: Predictive modulus falls within The proportion within (the interval mapping scheme can reach 100%).

[0190] Recalculation Consistency Index: The model output modulus is used for finite element recalculation, and the results are compared. With the goal The error distribution is the key to closing the loop in step S5.

[0191] The above evaluation ensures that the model not only makes accurate numerical predictions, but also meets the standards when used for finite element recalculation.

[0192] (v) Step S5: Target rolling resistance reverse reasoning / finite element recalculation verification / recharge iteration

[0193] like Figure 6 As shown, step S5 embeds the reverse generation capability of LLM into a closed loop of physical simulation verification—data backfeeding—incremental fine-tuning, thereby ensuring that the output not only meets the statistical accuracy of prediction but also satisfies physical consistency and engineering usability. This step determines that the present invention is more feasible to implement than simple machine learning fitting. The following disclosure covers eight aspects: inference input, candidate generation, multi-solution strategy, constraint projection, finite element recalculation, acceptance criteria, backfeeding mechanism, and incremental training.

[0194] Step S5.1, Inference Input and Target Definition

[0195] Input the target operating condition and target rolling resistance during inference: .in:

[0196] Target rolling speed (e.g., 80 km / h);

[0197] Target load (e.g., 2966N);

[0198] Target rolling resistance (e.g., 20N).

[0199] Model outputs candidate modulus vectors or If a multi-candidate strategy is adopted, the output will be... Group of candidates, among which The number of candidates (e.g., 5 or 10).

[0200] Step S5.2, Candidate Generation Strategy (Single Solution vs. Multiple Solutions)

[0201] 5.2.1 Single Solution Output

[0202] In continuous regression schemes, the model can directly output a single modulus vector. This allows for a quick solution, suitable for real-time or rapid iteration scenarios.

[0203] 5.2.2 Multiple Candidate Outputs

[0204] To improve the probability of compliance and support engineering trade-offs, this invention preferably outputs multiple candidates. Multiple candidates can be implemented in the following ways:

[0205] Random perturbation sampling: Adding noise to the mean solution of the model output before decoding:

[0206] ;

[0207] in, For the first Component output is not constrained; For the first Candidate perturbation terms (e.g., Gaussian noise); Candidate index.

[0208] Temperature sampling / Top-k: Temperature can be used in discrete labeled sequence schemes. Top-k retention numbers Control diversity. The larger the value, the higher the diversity, but the error may also increase. The larger the value, the wider the candidate space.

[0209] Beam search: Using beam width in sequence output scenarios. Before generation The sequence with the highest score. Among them, For the width of the bundle; To retain the number of candidates.

[0210] The significance of multiple candidates lies in the fact that the same target rolling resistance often corresponds to multiple modulus combinations (multiple solutions). Multiple candidates can increase the probability of finding a solution that meets the error threshold, while providing engineers with more options (such as taking into account durability / stress margin).

[0211] Step S5.3, Constraint Projection and Engineering Feasibility Screening

[0212] For each candidate First, apply the interval constraints, then apply the engineering constraint filtering.

[0213] 5.3.1 Interval Constraints (Hard Constraints)

[0214] If the model has already used interval mapping, then the interval is naturally satisfied; otherwise, pruning is performed:

[0215] ;

[0216] in, For the first Candidates in Cutting modulus of the component; To predict the modulus; For the interval.

[0217] get .

[0218] 5.3.2 Set of Engineering Constraints (Combination of hardware and software)

[0219] Engineering constraints can be divided into quick-filter constraints and strict recalculation constraints:

[0220] Rapid screening constraints (pre-reasoning): For example, limiting the relative modulus relationship of key components (the difference between the tread and base rubber modulus should not exceed a threshold), and limiting the triangular rubber modulus to not be lower than a certain value. These constraints can be determined without finite element analysis and are used to reduce the proportion of invalid candidates entering recalculation.

[0221] Strict recalculation constraints (judgment after recalculation): such as stress not exceeding limits, deformation not exceeding limits, local energy density not exceeding limits, etc., require finite element output.

[0222] To achieve reproducibility, it is recommended to... Write it into a rule table. Each rule includes: rule name, applicable component, threshold, and violation handling method (removal / demotion / recording).

[0223] Step S5.4, Finite Element Recalculation Verification

[0224] For each one that passed the pre-screening The rolling resistance was verified by writing it into the finite element material card and performing rolling simulation. Simultaneously, stress, deformation, and other indicators are extracted for constraint determination. The difference between recalculation and step S2 is that step S2 is used for sampling and library construction, while step S5 is used to verify whether the candidates meet the objectives and constraints. Usually, the number of recalculations is much smaller than the number of samples in step S2.

[0225] To improve efficiency, the following engineering strategies can be adopted:

[0226] 1) Simulation cache: If If the results are consistent with or similar to historical simulation samples, existing results can be read directly (the approximate matching threshold can be defined by Euclidean distance).

[0227] 2) Multi-level simulation: First, use a simplified model for rapid estimation. Then, the optimal candidate is verified using a high-precision model.

[0228] 3) Parallel computing: for Several candidate solutions were submitted in parallel.

[0229] Step S5.5, Acceptance Criteria and Output Sorting

[0230] 5.5.1 Target Error Determination

[0231] When the following conditions are met: ;in, For the first Candidate recalculated rolling resistance; Target rolling resistance; If the allowable error threshold (e.g., 0.5N or 2% of the target) is set, the candidate is considered to have met the target.

[0232] 5.5.2 Multi-indicator sorting (optional)

[0233] When multiple candidates meet the criteria, the results can be sorted and output according to their overall scores. The overall score can be defined as:

[0234] ;

[0235] in, For the first Candidate overall score (lower is better); As weight; These are the stress and deformation parameters obtained through recalculation; This is the upper limit threshold.

[0236] If the project only focuses on rolling resistance, then it can be taken as follows: The rest are 0.

[0237] The final output can be the optimal single solution. It can also output a Top-K solution set for engineers to choose from.

[0238] Step S5.6, Recharge Mechanism

[0239] Refeedback is triggered when all candidates fail to meet the error threshold, or when they do meet it but the error is too large or the constraint margin is insufficient. The standard format for refeedback samples is:

[0240] ;

[0241] in, This is a recalculation of the actual rolling resistance. These are candidate moduli. The significance of refeeding is to add points that the model considers to be potentially compliant but are actually not to the training set, so that the model can learn a more realistic mapping in that region and gradually eliminate systematic biases.

[0242] To ensure data quality, the following trigger conditions can be set for data reflow:

[0243] The number of new samples has reached the threshold Or, the verification error exceeds the threshold. Or, the sample density in the target region (a certain rolling resistance zone) is insufficient. Among these, The threshold for the number of new samples to trigger fine-tuning; This is the error threshold.

[0244] Step S5.7, Incremental Fine-tuning and Version Management

[0245] After refilling, the model When making incremental adjustments, it is recommended to follow these engineering rules:

[0246] 1) Freeze strategy: Can freeze the lower-level layers of the LLM (as before) (Layer) Only fine-tunes the upper layers and regression head to reduce overfitting and computational costs.

[0247] 2) Learning rate strategy: The fine-tuned learning rate is usually smaller than the initial training learning rate, for example, the initial training... Fine-tuning arrive .

[0248] 3) Mixed training: During fine-tuning, the re-feedback data is mixed with the original training set in a proportional manner (e.g., re-feedback: original = 1:4) to prevent catastrophic forgetting.

[0249] 4) Version number: The model version generated with each fine-tuning. Each time the simulation model is updated, a simulation version is generated. And record it in the sample to ensure traceability consistency.

[0250] V. Specific Application Examples:

[0251] 1.1 Definition of Test Objects and Components

[0252] like Figure 1 As shown in the attached figure, taking a sample passenger radial tire as the object, the tire rubber components are divided into sections as shown in the attached figure, and a component set is defined. The fixed components are arranged in the following order:

[0253] : fetus; Protective coating; Inner lining layer; Base adhesive; Crown layer; Second belt layer; First belt layer; : Tire tread; : Side of the tire; : Triangle rubber. Among them, For component serial number, , Meanwhile, the reinforcing material layers are input as non-optimization variables into the finite element model, including: wire coils, crown belt skeleton, carcass skeleton, and belt layer skeleton.

[0254] 1.2 Laboratory material parameter measurement (modulus, tan...) (density, Poisson's ratio)

[0255] To ensure consistency between the prototype, simulation, and dataset, the laboratory conducted the following tests on the rubber materials of each component and generated material card parameters:

[0256] elastic modulus The equivalent elastic modulus (unit MPa) of the small strain linear segment is used for linearized material input in finite element rolling simulation;

[0257] Poisson's ratio The rubber parts are approximately incompressible. (Enhancement layer) );

[0258] density Unit is ;

[0259] Loss factor tan : Measured by DMA at a preset temperature and frequency, used to characterize the viscoelastic energy dissipation level (for calibrating the material energy dissipation contribution to rolling resistance).

[0260] In this embodiment, the material card uses a rubber material modulus table for finite element analysis as the calibrated input (example values ​​are shown below, which are directly used for finite element modeling and sample generation):

[0261] fetus MPa , ,tan ;

[0262] Protective Gel MPa , ,tan ;

[0263] Triangle rubber MPa , ,tan ;

[0264] wire ring MPa , ,tan The remaining skeleton layers are the same as in the table.

[0265] Summary of parameter definitions: For the first Component elastic modulus; Poisson's ratio; Density; tan This is the loss factor.

[0266] 1.3 Tire prototype preparation and rolling resistance bench test (to form a real label)

[0267] The rubber compounds corresponding to the above materials are mixed according to the component formulation, calendered / extruded, and molded and vulcanized according to the component assembly process to obtain the sample tire. Steady-state rolling resistance tests are conducted using a rolling resistance test bench (drum tester / rolling resistance tester), with the following control conditions:

[0268] Scrolling speed: km / h;

[0269] Normal load: N.

[0270] Parameter definition: The scrolling speed; Normal load; rolling resistance Output for benchtop (unit: N).

[0271] In this example, to construct the initial training data, four sets of component modulus combinations were prepared (Data 1–Data 4, corresponding to four sets of rubber compound schemes or four sets of equivalent modulus calibration schemes), and the corresponding measured rolling resistance values ​​were obtained. The following real samples were formed (in component order) Record):

[0272] Sample 1 (Data 1): , , , ;

[0273] Sample 2 (Data 2): , , , ;

[0274] Sample 3 (Data 3): , , , ;

[0275] Sample 4 (Data 4): , , , .

[0276] in, For the first Sample modulus vector In this example, the initial .

[0277] 2.1 Establishment of 2D Mesh and 3D Model

[0278] like Figure 2 As shown, the tire profile is divided into two-dimensional meshes, ensuring that each component area is aligned with... Corresponding; then according to the appendix Figure 3 The finite element model is rotated / spread to generate a 3D solid and mesh, and the contact relationship between the ground surface and the tire, rolling boundary conditions and load conditions are set.

[0279] Key points:

[0280] 1) Consistency of material assignment: The two-dimensional partition ID is bound to the material card ID to ensure replacement. Batch assignment can be scripted at any time;

[0281] 2) Consistency of operating conditions: The finite element simulation uses the same operating conditions as the test bench. km / h N (add tire pressure and other boundary measures if necessary);

[0282] 3) Output consistency: Simulated output rolling resistance As a label, with the stand Align with the same units of measurement (for subsequent learning and verification).

[0283] 2.2 Automated scripts to augment the dataset (to a trainable size)

[0284] Based on the actual test samples (4 items), in order to cover the continuous parameter space, automated scripts were used in each component interval. The system generates more modulus combinations and performs finite element rolling simulation to obtain an expanded sample:

[0285] ;

[0286] in, The rolling resistance is output by finite element method.

[0287] Parameter definition: For the first Adjustable range of component modulus; For the first Rolling resistance obtained from group simulations / experiments; total number of samples: .

[0288] To ensure that the data is based on real experiments, the amplification data follows these principles:

[0289] Using four real samples from the test bench as anchor points, the sample is amplified and sampled more densely in its neighborhood (to increase the learning density of the model for engineering feasible regions).

[0290] Each simulation records the version number, mesh label, and convergence flag, and discards non-convergent samples.

[0291] A finer sampling resolution is used for key components (tread, sidewall, base rubber, and wing rubber) to improve the fitting ability of the rolling resistance sensitive area.

[0292] 3.1 Training Data Organization and Input / Output Definition

[0293] Organize the dataset according to the sample structure as follows:

[0294] enter: ;

[0295] Output: .

[0296] The training objective is to learn the inverse mapping. :

[0297] ;

[0298] in, The predicted modulus vector output by the model; This is the trained LLM inverse mapping model.

[0299] 3.2 LLM Structured Adaptation

[0300] To enable LLM to stably process continuous numerical values, this embodiment adopts a structure of numerical embedding + component alignment encoding + continuous regression head:

[0301] Numerical embedding functions Will Mapped to a vector;

[0302] Component alignment coding Guarantee output With components One-to-one correspondence;

[0303] Return to head Direct output of continuous modulus .

[0304] The following formula can be used: ;in, For LLM to the first The hidden state of the component token; For regression decoding function; To output the modulus.

[0305] 3.3 Reverse Reasoning Input and Output

[0306] Set target operating conditions and target rolling resistance:

[0307] load N;

[0308] speed km / h;

[0309] Target rolling resistance N.

[0310] Model output (by) order):

[0311] Predicted modulus vector: (Unit: MPa)

[0312] Simultaneously output a summary of operating conditions (for consistency verification): , .

[0313] Parameter definition: Target rolling resistance; For the target operating condition; Output modulus for LLM; For the first Component modulus.

[0314] 3.4 Finite element recalculation verification and error calculation

[0315] Will Write into the finite element material card and replace accordingly. The modulus parameter, the rest Poisson's ratio ,density tan The parameters of the skeleton layer remain fixed as described in the aforementioned table, under the same working conditions. km / h Perform rolling simulation at N to obtain the recalculated rolling resistance: N.

[0316] Define relative error for:

[0317] ;

[0318] in, The percentage is the relative error. Finite element method for recalculating rolling resistance; The target rolling resistance.

[0319] Substitute the values: N, N, obtained .

[0320] Conclusion: Under the reverse design task with a target rolling resistance of 20N, the multi-component modulus combination output by the LLM of this invention has an error of only 1.5% with the target after finite element recalculation, proving that this invention can achieve high-precision reverse design and physical consistency verification.

[0321] 3.5 Comparison with traditional forward trial-and-error screening

[0322] Under the same objective, the traditional process is usually: manually setting multiple modulus schemes → performing finite element analysis on each scheme → selecting the scheme closest to 20N. Since this example involves the modulus of 10 components, the combination space is extremely large. If each component only takes 5 discrete levels, the number of combinations is... Even with orthogonal / sampling methods, a large number of simulations are required to approximate the target. This invention, however, forms a mapping through a single training iteration. Then, candidates can be directly output during the inference phase. It only requires a small number of recalculations (e.g., 1–10 times) to meet the target or enter closed-loop fine-tuning, significantly reducing the number of simulation trial and error attempts and R&D time costs; at the same time, simulation recalculation-reflow ensures the usability and interpretable physical consistency of the output solution.

[0323] Furthermore, when the initial inference output is Threshold not met Time (e.g., setting) (N or 1% target value), the recalculated real sample will be reinjected: ;in, To recalculate the actual rolling resistance, This corresponds to the modulus vector. Subsequently, the model parameters are incrementally fine-tuned with a small learning rate to make the local mapping of the model more accurate near the target rolling resistance, thereby improving its performance on subsequent similar targets (e.g., ...). , N) Further reduce errors and improve the first-pass yield during inference. Parameter definition: The allowable error threshold; refluxing samples are used to update model parameters, so that... It is more stable within the target range.

[0324] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0325] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0326] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0327] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0328] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0329] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0330] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0331] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.< / qen> < / qe1> < / qr> < / qf> < / qv>

Claims

1. A method for optimizing the modulus of tire rubber materials based on Large Language Modeling (LLM), characterized in that, The method includes the following steps: S1, Obtain the set of tire rubber components, which includes multiple components to be optimized, each component to be optimized having a corresponding component number; Set the allowable range of elastic modulus for each component to be optimized; S2, under rolling speed and load conditions, establish a tire finite element rolling model, generate N sets of modulus vectors in batches, each set of modulus vectors includes the elastic modulus of each component to be optimized, and obtain the rolling resistance corresponding to each set of modulus vectors through finite element simulation; S3, construct the dataset, in which each sample includes rolling speed, load, rolling resistance and the corresponding modulus vector; S4. The LLM is adapted to structured numerical input and output, and an inverse mapping model is trained based on the dataset. This enables the inverse mapping model to output a predictive modulus vector based on the input rolling speed, load, and rolling resistance. S5. During the inference phase, the target rolling resistance and target operating conditions are input. The target operating conditions include the target speed and target load. The inverse mapping model outputs candidate modulus vectors and constrains the candidate modulus vectors to the allowable range of the elastic modulus of each component to be optimized, thus obtaining the constrained modulus vectors. The constrained modulus vectors are input into the tire finite element rolling model for finite element recalculation to obtain the verification rolling resistance. When the absolute value of the difference between the verification rolling resistance and the target rolling resistance is not greater than the allowable error threshold, the constrained modulus vector is output as the optimization result. Otherwise, the target speed, target load, verification rolling resistance, and constrained modulus vector are added to the dataset as refeed samples, and the inverse mapping model is updated.

2. The method according to claim 1, characterized in that: In step S1, the component to be optimized includes at least two of the following: tread, base rubber, sidewall, carcass, inner liner, and triangular rubber. And / or, in step S2, the finite element rolling model includes a two-dimensional finite element mesh or a three-dimensional finite element mesh, and ground boundary conditions and rolling boundary conditions are set; The finite element simulation is either a static rolling simulation or a dynamic rolling simulation. And / or, in step S2, N sets of modulus vectors are generated using one of orthogonal experimental design, Latin hypercube sampling, or Sobol sequence sampling, and a convergence flag is recorded for each set of simulation results. The convergence flag is used to indicate the convergence status of the corresponding set of simulations.

3. The method according to claim 1, characterized in that: In step S3, the dataset is divided into a training set, a validation set, and a test set, with a ratio of 90%:5%:5%. The training set is used to train the inverse mapping model, the validation set is used for model selection, and the test set is used for generalization evaluation.

4. The method according to claim 1, characterized in that: In step S4, the structured numerical input-output adapter uses a numerical quantization step size to discretize the rolling speed, load, rolling resistance and elastic modulus of each component to be optimized into a labeled sequence input LLM, wherein the numerical quantization step size is the quantization resolution; Alternatively, in step S4, the structured numerical input-output adapter uses a numerical embedding function to map the rolling speed, load, rolling resistance and elastic modulus of each component to be optimized into vectors and then inputs them into the LLM. The numerical embedding function is a mapping function from continuous numerical values ​​to the embedding space. And / or, in step S4, the loss function for training the inverse mapping model is obtained by weighted summation of the modulus prediction error term and the constraint violation penalty term; the modulus prediction error term is the average of the squared differences between the predicted modulus and the true modulus of each component to be optimized; the constraint violation penalty term is used to penalize model outputs that do not meet the allowable range of elastic modulus or the set of engineering constraints; the set of engineering constraints is a set of constraints used to limit the candidate modulus vectors to meet the requirements of tire structural safety, deformation range and engineering manufacturability.

5. The method according to claim 4, characterized in that: The set of engineering constraints includes at least upper limit constraints on structural stress and / or upper limit constraints on deformation; the upper limit constraints on structural stress are used to limit the equivalent stress obtained from finite element simulation to not exceed the maximum allowable equivalent stress; The upper limit constraint on deformation is used to limit the displacement or deformation obtained from finite element simulation to not exceed the maximum allowable displacement or maximum allowable deformation.

6. The method according to claim 1, characterized in that: In step S5, the inverse mapping model outputs M groups of candidate modulus vectors, where M is the number of candidates. Each group of candidate modulus vectors has a corresponding candidate number, which is numbered sequentially from 1 to M. The candidate modulus vector is generated using one of the following methods: beam search, Top-k sampling, or temperature sampling. In Top-k sampling, the retention number is used to limit the number of candidates retained during sampling. In temperature sampling, the temperature parameter is used to control the randomness of the candidate modulus vector generation. The retention number and the temperature parameter are used together to control the diversity of candidate solutions. And / or, in step S5, the constraint projection operator is used to constrain the candidate modulus vector, which is used to map the candidate modulus vector into a feasible modulus vector that simultaneously satisfies the allowable range of elastic modulus and the set of engineering constraints. And / or, in step S5, the conditions for triggering the refeed update include at least the number of new samples reaching the threshold for the number of new samples that triggers the update, or the verification error exceeding the error threshold that triggers the update.

7. The method according to claim 1, characterized in that: The LLM is a numerical generation LLM for inverse generation of tire multi-component modulus, including: (1) Numerical embedding module, used to map input working conditions and component indices into embedding vectors, wherein the input working conditions include rolling speed, load and rolling resistance; (2) Component alignment encoding module, used to establish a one-to-one correspondence between each component to be optimized and its corresponding output modulus. The component alignment encoding module adopts position encoding or alignment encoding. (3) A continuous regression decoder head is used to directly output a continuous modulus vector, the continuous modulus vector including the prediction modulus of each component to be optimized, and the continuous regression decoder head is used to map the hidden state of LLM into a continuous value. Furthermore, the LLM undergoes supervised learning during the training phase using samples containing rolling speed, load, rolling resistance, and corresponding modulus vectors, thereby obtaining the ability to inversely generate modulus combinations from the target resistance.

8. A tire rubber material modulus optimization system based on Large Language Model (LLM), characterized in that, include: The module includes parameterization and constraint, finite element simulation, dataset management, LLM training and inference, and physical verification and backfeeding iteration. The parameterization and constraint module is used to perform step S1 in claim 1; The finite element simulation module is used to perform step S2 in claim 1; The dataset management module is used to perform step S3 in claim 1; The LLM training and inference module is used to execute the candidate modulus vector output in steps S4 and S5 of claim 1. The physical verification and backfeed iteration module is used to perform constraint processing, finite element recalculation, optimization result output and data backfeed update in step S5 of claim 1.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images