Intelligent design method for mechanical parts based on large model

Abstract

This invention discloses an intelligent design method for mechanical parts based on a large model. The method first extracts the dimensional parameters of the parts and sets effective value boundaries, constructing a parametric universal template without geometric interference. Then, it discretizes sampling within the boundaries and constructs a high-fidelity simulation dataset through automated parameter-driven modeling and batch finite element solution. Next, based on this simulation dataset, it utilizes efficient parameter fine-tuning technology to enable the large language model to learn physical mapping laws, achieving inverse parameter optimization. Finally, through automated cross-software physical verification, it transforms real physical errors into error feedback signals and uses reinforcement learning to continuously correct the large model in a closed-loop manner. This invention solves the problems of low efficiency and difficulty in inverse optimization in traditional forward simulation, realizing intelligent design of mechanical parts with operational reliability and long lifespan.

Description

Technical Field

[0001] This invention relates to the field of collaborative technology between large-scale artificial intelligence models and intelligent assisted design in mechanical engineering. Specifically, it relates to an intelligent design method for mechanical parts driven by a large language model, possessing a closed loop of real physical verification and self-evolving anti-illusion capabilities. Background Technology

[0002] As high-end equipment manufacturing (such as heavy-duty powertrains, aero engines, and wind power equipment) continues to develop towards lightweighting, high power density, and extremely high reliability, the design and development of key components in mechanical equipment (such as various complex gears, bearings, drive shafts, casings, and rotating bodies) are facing unprecedented challenges from the cross-coupling of multidisciplinary physical fields. Existing engineering research and development methods have the following core pain points in their application in the field of mechanical aided design:

[0003] (1) Historical data in vertical fields are severely isolated, making it difficult to reuse implicit patterns in multidimensional design parameters:

[0004] The design of complex mechanical components involves not only intricate three-dimensional geometry and topology but also constraints imposed by materials mechanics and machining processes. Over the past few decades, manufacturing companies have accumulated a vast amount of drawings and expert parameter tuning experience. However, this extremely valuable heterogeneous data is often scattered in unstructured form across different industrial software systems, lacking a semantic-level retrieval and structured alignment pipeline between input features (geometric dimensions) and output responses (physical performance). When developing new components, designers struggle to reuse underlying engineering mechanics principles across platforms, leading to the repeated rework of a large amount of low-level interference-avoidance trial-and-error work.

[0005] (2) Traditional forward simulation has extremely low computational efficiency, and multi-objective inverse optimization faces the "curse of dimensionality":

[0006] Traditional forward simulation calculations are inefficient and involve long iterations: the development of traditional mechanical parts generally adopts a forward trial-and-error process of "initial dimensions given by engineers' experience → manual 3D modeling in CAD → importing into CAE software for mesh generation and mechanical simulation → readjusting parameters after discovering that stress or rotational inertia does not meet the standards." Modern parts typically require extreme lightweighting while ensuring long fatigue life and meeting the ultimate stress under various working conditions. A complete verification cycle is extremely long, and often only a locally suboptimal solution is obtained rather than a globally optimal solution. Summary of the Invention

[0007] Technical problems to be solved

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a large model-driven intelligent design method to solve the technical problems of slow forward calculation, difficulty in reverse parameter optimization, and lack of fine-tuning data in vertical fields in traditional mechanical design.

[0009] Technical solution

[0010] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows:

[0011] A mechanical intelligent design method based on high-fidelity simulation data includes the following steps:

[0012] S1: Constructing a parametric universal template: Extract the high-dimensional dimensional parameter set and associated constraint relationships of the target mechanical parts, and set effective boundary conditions for each dimensional parameter to avoid geometric interference and reconstruction failure. This is then solidified into the corresponding parametric universal template in the 3D design software, ensuring that the generated 3D model has a valid topology and is free from geometric interference.

[0013] S2: Automated construction of high-fidelity simulation dataset: High-dimensional spatial discrete sampling is performed under strict valid value boundary conditions to generate multiple sets of discrete and valid dimension parameter matrices; the 3D design software is automatically called to batch convert the dimension parameter matrices into 3D model files, which are then imported into finite element simulation software. Physical simulation is performed under set operating parameters to accurately obtain multiple target performance indicators corresponding to the model; the dimension parameter matrices, operating parameters, and target performance indicators are feature-aligned to construct a high-fidelity simulation dataset.

[0014] S3: Domain-wide large model internalization fine-tuning: The network parameters of the base large language model are updated using efficient parameter fine-tuning technology, so that it implicitly internalizes the complex nonlinear physical simulation mapping law and obtains a cognitive-level intelligent design model.

[0015] S4: Inverse parameter optimization: Receive the expected target performance indicators and working status instructions input by the user, call the cognitive-level intelligent design model to perform inverse parameter deduction, and output the preliminary predicted combination of design size parameters;

[0016] S5: Cross-software physical closed-loop verification: intercept the preliminary predicted combination of design dimension parameters, automatically drive the 3D design software to reconstruct the physical verification model through the underlying interface, and drive the finite element simulation software to perform physical solution to obtain the real physical verification performance index under the dimension combination.

[0017] S6: Physical Feedback Reinforcement Learning and Continuous Self-Evolution Correction: Compare the actual physical verification performance index with the expected target performance index, calculate the physical error value. If the physical error value exceeds the preset engineering tolerance range, convert the physical error value into an error feedback signal, use the reinforcement learning algorithm to reverse correct the weight parameters of the model, and trigger the re-execution of steps S4 to S6 until the error value is within the preset engineering tolerance range, and output the final component design parameters.

[0018] A method for intelligent design of mechanical parts based on a large model, comprising:

[0019] Offline data generation module: used to perform steps S1 and S2;

[0020] Model training and knowledge internalization module: used to perform step S3;

[0021] Knowledge retrieval and reverse reasoning module: used to execute step S4;

[0022] Physical closed-loop and anti-hallucination error correction module: used to execute steps S5 and S6.

[0023] Beneficial effects

[0024] Compared with the prior art, the present invention has the following significant advantages:

[0025] (1) Innovate the traditional forward trial-and-error design paradigm to achieve efficient reverse derivation of multi-dimensional parameters of mechanical parts.

[0026] This invention breaks through the limitations of traditional physical solvers that rely on a one-way trial-and-error approach based on given parameters. Through efficient parameter fine-tuning technology, high-fidelity simulation data spanning multi-dimensional feature parameters and multiple physical responses is implicitly internalized into the neural network weights of a large language model. This endows the large language model with the ability to deduce implicit physical laws even in the absence of explicit mechanical equations. Whether it's gears, bearings, or complex mechanical equipment, by inputting the expected target performance and operating condition commands, the model can deduce the structural dimension scheme. Transforming time-consuming simulation iterations into second-level responses through natural language interaction greatly accelerates the development speed of high-end equipment products.

[0027] (2) Solving the vertical data shortage in industrial AI and building a high-fidelity data pipeline:

[0028] To address the severe lack of high-quality training data for the practical application of AI in machinery, this invention proposes a multi-dimensional constrained spatial discrete point-scattering mechanism with "effective value boundary conditions." By strictly limiting the engineering boundary domains of various dimensional parameters, it minimizes the risk of 3D model reconstruction failures and interference errors caused by distorted parameters. This method completely integrates the underlying interfaces of CAD and CAE, realizing a pathway for constructing industrial structured datasets: "constrained point-scattering → fully automatic topology modeling → automatic mesh generation → batch load solving."

[0029] (3) The physical anti-illusion mechanism of "CAE in the loop" ensures the engineering rigor of the drawings:

[0030] Overcoming the subjective limitations of general large models that rely excessively on human feedback (RLHF), this invention uses "objective physical error" calculated by an external industrial solver as a hard error feedback signal for reinforcement learning, effectively avoiding the risk of the model generating non-physical or meaningless parameters. The system has the ability to continuously self-evolve and become more accurate with use. Attached Figure Description

[0031] Figure 1 The main flowchart of an intelligent design method for mechanical parts based on a large model provided in an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the architecture and operational logic of a mechanical intelligent design system based on high-fidelity data internalization and physical feedback closed loop, as provided in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, multi-dimensional parameter collaborative optimization logic, and physical feedback closed-loop mechanism of this invention clearer, the data flow and underlying algorithm control logic of the system architecture of this invention will be further explained in detail below with reference to the accompanying drawings and specific embodiments.

[0034] This invention essentially provides a novel intelligent design paradigm that transcends traditional forward black-box trial and error in mechanical design. To facilitate a detailed explanation of its implementation, this embodiment selects the intelligent reverse design of a typical complex rotating mechanical component as the controlled execution object, combined with... Figure 1 This paper elaborates on the specific execution and flow process of the method based on the internalization of multidimensional parameterized general model and high-fidelity physical simulation data and physical feedback closed loop (steps S1 to S6).

[0035] The design of complex rotating machinery components is essentially a high-dimensional, nonlinear, multidisciplinary design optimization problem: under specific speed conditions, it is necessary to reduce the overall system weight to achieve lightweighting while ensuring that the principal inertial torque remains high to suppress transient torsional vibrations and speed fluctuations in the transmission system. When subjected to maximum speed operation, the maximum stress in the geometric solid region must not exceed the material's fatigue yield limit, and it must pass a stringent ultimate speed test. This invention completely reconstructs this reverse optimization process using a large-scale artificial intelligence language model and a closed-loop system of industrial software in the loop:

[0036] S1: Construct a parameterized universal template that covers high-dimensional features and rigorously define the boundary of the "total parameterized range" to prevent interference.

[0037] In the initial design phase, the system performs topology reduction and feature decoupling on the target mechanical parts. This method comprehensively considers the independent design parameters that determine the machining modeling and engineering constraints of the parts, and constructs a design variable matrix.

[0038] Specifically, the matrix covers:

[0039] (1) Geometric dimensions and shape parameters (continuous variables): such as the diameter of the main mounting surface, the outer diameter of the part, the diameter of the main bearing hole, the distance from the mounting surface to the outer end face, the thickness of the critical bearing flange, the diameter of the guide hole, the thickness of each step, and the dimensions of the rounded corners and chamfers at various points in the internal cavity;

[0040] (2) Design and manufacturing Boolean features (discrete variables): such as whether a weight-reducing step needs to be added to the mounting surface, whether spare assembly holes are needed, and whether positioning and guiding features are needed.

[0041] The core principle of this invention lies in the fact that traditional random point generation parameter combinations are prone to conflict, leading to geometric interference, topological surface breaks, or feature generation failures when subsequently imported into CAD software. To address this issue, this system, considering the constraints of actual machining processes and assembly interference space, sets a strict "total parameterized range" (i.e., effective value boundary conditions) for each of the aforementioned parameters. For example, the underlying logic forces the definition domain of the outer diameter of a part to be limited to a specific tolerance zone, and the thickness of the main bearing hole is strictly limited to the allowable range of the process.

[0042] The system embeds all the aforementioned high-dimensional parameters and total range constraints into a universal template for fully parameterized dynamic driving within the underlying kernel of 3D design software (such as CREO). This fundamentally ensures that during subsequent data sampling, as long as the parameter combinations of each dimension fall within the set multi-dimensional parameterized total range, the 3D solid digital model generated by subsequent automated driving reconstruction will be topologically valid and will not experience geometric boundary violations or surface breakage errors.

[0043] S2: Construction of Automated Constrained Point Scattering and High-Fidelity Simulation Datasets

[0044] The system performs fully automated multidimensional parameter sampling and batch simulation solutions on servers or computing clusters.

[0045] (1) Boundary-bounded sampling: The Latin hypercube sampling algorithm (LHS) is used to enforce the strictly parameterized boundary defined in step S1. Discrete "points" are generated within the matrix to produce tens of thousands of uniform size matrices that satisfy topological rules.

[0046] (2) Fully automated pipeline flow across software: The scheduling middleware automatically calls the application programming interface of the 3D design software, reads the size matrix one by one, performs parameter dynamic injection and model reconstruction; after successful reconstruction, the mesh is automatically generated and transferred to the finite element simulation software.

[0047] (3) Loading and performance index extraction: The system imports core load parameters (such as setting specific limit operating speeds) into the solver. After the solution converges, the system extracts the corresponding core target performance index for each set of input parameters:

[0048] Component weight (strictly control the bottom line target for lightweighting);

[0049] Principal inertial moment (to assess the target of suppressing dynamic torsional vibration);

[0050] Maximum stress (verifying the static yield red line under extreme conditions);

[0051] Fatigue life (evaluating long-term fatigue life under alternating load cycles).

[0052] Limiting speed (assessing the safety threshold for instability and damage under extreme overspeed conditions).

[0053] After batch processing of massive pipelines, the system successfully constructed a massive high-fidelity structured dataset containing a mapping and alignment of "[size / design parameter matrix] + [operating condition parameters] → [the above six core physical performance indicators]".

[0054] S3: Fine-tuning of network weights in a large domain model and implicit internalization of underlying physical mechanisms

[0055] The aforementioned high-fidelity dataset was transformed into a multimodal engineering instruction dialogue corpus. Efficient parameter fine-tuning techniques (such as the low-rank adaptive LoRA algorithm) were employed, freezing the general pre-trained weights of the base large language model, and guiding network parameter updates by injecting trainable low-rank decomposition matrices. In this process, the large model implicitly internalized the complex nonlinear mechanical rules between input and output. For example, the model learned through data training that: "Under specific speed conditions, reducing the thickness of key load-bearing components or setting the transition fillet too small can reduce weight, but will cause local stress concentration, leading to the maximum stress exceeding the safety limit, thus lowering the ultimate speed." The fine-tuned large language model became a "cognitive-level intelligent design model" with intuition regarding mechanical engineering constraints.

[0056] S4: Knowledge Retrieval Enhancement and High-Dimensional Parameter Inverse Derivation

[0057] In practical interactive R&D scenarios, engineers input natural language instructions with expected goals into the front-end system (e.g., input specific speed parameters, set the lower limit of the principal inertial torque, and require the maximum stress and fatigue life to meet the safety margin threshold, with the objective function being the minimization of the total weight of the components). The system first triggers the Knowledge Retrieval Enhancement (RAG) module to perform semantic vectorization processing on the initial instructions and conduct similarity matching retrieval in a pre-built vertical heterogeneous knowledge base (covering national standards, industry design specifications, and enterprise historical experience data). It accurately extracts implicit constraint boundary conditions that match the design intent and uses them as contextual enhancement features, fusing them with the user's initial instructions using prompts. The system then inputs the fused complete instructions into the cognitive-level intelligent design model (e.g., ...). Figure 2 (As shown). It directly invokes the internalized nonlinear physical and mechanical mapping laws in the hidden layer of its neural network to complete the inverse feature optimization in the high-dimensional solution space in a very short time. The model outputs a precise numerical combination covering all dimensions of geometry, topology, and manufacturing process, including key mating diameters, outer diameters, and apertures at various locations, as the initial design scheme.

[0058] S5: Physical closed-loop recalculation process across underlying industrial software

[0059] To avoid the "parameter illusion," the system extracts the preliminary predicted design dimension parameter combination output from step S4 and automatically imports it into the underlying CAD / CAE co-simulation engine for real physical retesting and verification (the process is described in reference [reference]). Figure 2 The physical closed-loop and anti-hallucination error correction module (D) is used in this system. Specifically, the 3D design software reconstructs the solid model based on the predicted dimensional values, and the finite element simulation software loads the corresponding operating condition parameters for solution. After the solution is completed, the system can obtain the actual physical verification indicators of the predicted scheme in the actual simulation environment (such as actual weight, principal moment of inertia, maximum stress, etc.).

[0060] S6: Feedback Based on Real Physical Error Costs and Large-Scale Model Anti-Hallucination Self-Evolution

[0061] The adaptive physical error comparison module quantifies and compares the "real physical verification performance index" calculated in step S5 with the "expected target performance index" set in step S4, and calculates the specific physical error value. If the system determines that the error value exceeds the preset engineering tolerance range (for example, the model pursues the ultimate lightweight of the parts, causing its actual maximum stress to exceed the material's safe fatigue limit, or some key stiffness / rotation speed indicators fail to meet the standards), then the system determines that the current large model generation strategy has triggered an "engineering mechanics illusion" (i.e., generated parameters that violate physical laws).

[0062] Subsequently, the physics feedback error correction module converts the physics error values ​​into a negative feedback signal in reinforcement learning algorithms (such as the Proximal Policy Optimization algorithm PPO). The system uses this signal to correct the weight parameters of the large model, suppressing and pruning the "aggressive optimization strategy that deviates from mechanical constraints" in the model from the algorithm's underlying layer.

[0063] After the large model completes parameter correction, the system triggers a restart of the large model simulation to generate next-generation component size schemes that adjust structural thickness or alleviate stress concentration, and automatically re-triggers closed-loop retesting steps S4 to S6. This "generation-verification-feedback" closed-loop process will automatically iterate until the complete set of three-dimensional parameter arrays generated by a certain generation of large model has all core performance indicators that converge smoothly within the set safety engineering tolerance domain.

[0064] Ultimately, the system unlocks the verification defenses and outputs full-size 3D digital drawings of components to the manufacturing end, taking into account expected goals (such as lightweighting and specific dynamic performance) and conforming to the laws of engineering mechanics. This mechanism effectively solves the pain point of AI-generated parameters being difficult to implement in vertical industries, enabling the intelligent design system to continuously and adaptively evolve within an interactive closed loop.

[0065] Patent Standard Disclaimer and Extension Statement

[0066] The above description is merely a typical embodiment of rotating machinery equipment that discloses the core closed-loop control concept of this invention, but the scope of protection of this invention is not limited thereto. Those skilled in the art should understand that, without departing from the technical essence of this invention, any extension of this control method to the reverse feature optimization of other complex transmission or load-bearing components should be covered within the intellectual property protection scope of this invention.

Claims

1. A method for intelligent design of mechanical parts based on a large model, characterized in that, Includes the following steps: S1: Construct a parametric general template: Extract the set of dimensional parameters and associated constraints of the target mechanical parts, and set effective boundary conditions for each dimensional parameter to avoid geometric interference and reconstruction failure. Solidify these into the corresponding parametric general template in the 3D design software to ensure that the generated 3D model has a valid topology and no geometric interference. S2: Automated construction of high-fidelity simulation datasets: Spatial discrete sampling is performed under strict valid value boundary conditions to generate multiple sets of discrete and model-valid size parameter matrices; The automated 3D design software is used to batch convert the dimensional parameter matrix into 3D model files, which are then imported into the finite element simulation software. Physical simulation is performed under the set working parameters to accurately obtain multiple target performance indicators corresponding to the model. The dimensional parameter matrix, working parameters, and target performance indicators are feature-aligned to construct a high-fidelity simulation dataset. S3: Domain-wide large model internalization fine-tuning: The network parameters of the base large language model are updated using efficient parameter fine-tuning technology, so that it implicitly internalizes the complex nonlinear physical simulation mapping law and obtains a cognitive-level intelligent design model. S4: Inverse parameter optimization: Receive the expected target performance indicators and working status instructions input by the user, call the cognitive-level intelligent design model to perform inverse parameter deduction, and output the preliminary predicted combination of design size parameters; S5: Cross-software physical closed-loop verification: intercept the preliminary predicted combination of design dimension parameters, automatically drive the 3D design software to reconstruct the physical verification model through the underlying interface, and drive the finite element simulation software to perform physical solution to obtain the real physical verification performance index under the dimension combination. S6: Physical Feedback Reinforcement Learning and Continuous Self-Evolution Correction: Compare the actual physical verification performance index with the expected target performance index, calculate the physical error value. If the physical error value exceeds the preset engineering tolerance range, convert the physical error value into an error feedback signal, use the reinforcement learning algorithm to reverse correct the weight parameters of the model, and trigger the re-execution of steps S4 to S6 until the error value is within the preset engineering tolerance range, and output the final component design parameters.

2. The method according to claim 1, characterized in that, In step S4, after receiving the user input instruction, the method further includes: using retrieval-enhanced generation (RAG) technology to perform semantic-level retrieval in a pre-built vertical heterogeneous knowledge base of mechanical equipment, extracting industry design specifications and constraints that match the design intent, and inputting them as contextual prompts into the cognitive-level intelligent design model.

3. The method according to claim 1, characterized in that, In step S2, the spatial discrete sampling adopts the Latin hypercube sampling algorithm or the Monte Carlo sampling algorithm to ensure the uniformity of parameter distribution throughout the entire effective solution space; the scheduling script is a middleware that requires no manual intervention and is used to realize the fully automated workflow of model updating, intermediate format export, mesh generation and solver calculation.

4. The method according to claim 1, characterized in that, In step S3, the parameter efficient fine-tuning technique employs a low-rank adaptive (LoRA) fine-tuning algorithm to freeze the pre-trained weights of the base large language model and fit the evolution law of the underlying mechanical partial differential equations by injecting a trainable low-rank decomposition matrix.

5. The method according to claim 1, characterized in that, In step S6, the construction logic of the error feedback signal is as follows: a reward function is constructed based on the absolute deviation between the actual physical verification performance index and the expected target performance index. The greater the deviation, the higher the negative reward weight is assigned. Algorithms such as Proximal Policy Optimization (PPO) are used as the reinforcement learning algorithm to update the model policy gradient.

6. A method for intelligent design of mechanical parts based on a large model, characterized in that, include: Offline data generation module: used to perform steps S1 and S2; Model training and knowledge internalization module: used to perform step S3; Knowledge retrieval and reverse reasoning module: used to execute step S4; Physical closed-loop and anti-hallucination error correction module: used to execute steps S5 and S6.

Images