CN-121766160-B - Multi-agent-based reverse design method and system for cross-physical-field microstructure

CN121766160BCN 121766160 BCN121766160 BCN 121766160BCN-121766160-B

Abstract

The invention belongs to the technical field of intelligent design and calculation optimization, and provides a multi-agent-based reverse design method and system for a cross-physical-field microstructure, which are used for acquiring design intent and analyzing the design intent into a quantized cross-physical-field attribute target vector; the design method comprises the steps of utilizing a pre-trained neural network model to map design intentions into adjustment parameters in a potential representation space, decoding to generate geometric configurations of initial microstructures to construct an initial evolution population, conducting cross-physical field numerical simulation on the microstructures in the population, approximately identifying the evolution direction of a parameter space through local gradients, conducting elite screening through non-dominant ordering driven by pareto, dynamically adjusting the adjustment parameters of the potential space, guiding the evolution process to converge towards a physical effective pareto front, and obtaining a design scheme according to an optimization result. The invention realizes the efficient and accurate design of the microstructure under complex constraint.

Inventors

LV LIN
ZHAO ZHENYUAN
XING YU
XUE TIANYANG
CAO LINGXIN
YAN XIN

Assignees

山东大学

Dates

Publication Date: 20260508
Application Date: 20260305

Claims (10)

1. A multi-agent-based reverse design method for a cross-physical-field microstructure is characterized by comprising the following steps: Acquiring design intents containing natural language descriptions, generating and calling an external physical knowledge base through retrieval enhancement, and analyzing the design intents into quantized cross-physical field attribute target vectors; Based on the analyzed quantized cross-physical field attribute target vector, mapping the design intent into adjustment parameters in a potential representation space by utilizing a pre-trained neural network model, decoding to generate a geometric configuration of an initial microstructure, and constructing an initial evolution population; performing cross-physical field numerical simulation on microstructures in the population, and extracting performance data in real time; Based on numerical deviation obtained by numerical simulation across a physical field, the evolution direction of a parameter space is approximately identified through local gradients, elite screening is carried out by utilizing non-dominant ordering driven by pareto, the adjustment parameters of a potential space are dynamically adjusted, and the evolution process is guided to converge towards the physical effective pareto front; And screening an optimal microstructure geometric model meeting multi-objective constraint across a physical field according to the optimization result to obtain a design scheme.
2. The multi-agent-based reverse design method of cross-physical field microstructure according to claim 1, wherein the process of parsing the design intent into quantized cross-physical field attribute target vectors by retrieving an enhanced generation and retrieval external physical knowledge base comprises parsing physical field coupling relationships in user instructions by agents, determining physical feasible domain boundaries by knowledge graph retrieval, converting fuzzy semantics into quantized cross-physical field attribute target vectors containing target values, tolerances and priority weights, and translating semantic targets into standardized format constraint files.
3. The multi-agent based inverse design method of cross-physical field microstructures of claim 1, wherein mapping design intent to tuning parameters in potential representation space and decoding to generate geometry of initial microstructures, constructing initial evolutionary population comprises defining mechanical condition space as search space, generating condition vector of model by design variables, optimizing problem from finding optimal geometry Conversion to find optimal condition parameters I.e. Such that the properties are simulated Meeting the target value of the analysis definition; and searching an initial anchor point in a search space according to the semantic feature vector by utilizing a generator network of a depth generation model for jointly encoding the geometric features and the physical attributes so as to ensure that the rigidity property of the generated microstructure meets the condition.
4. The multi-agent-based reverse design method for the microstructure across the physical field of claim 1, wherein the process of identifying the evolution direction of the parameter space by local gradient approximation based on the numerical deviation obtained by numerical simulation across the physical field comprises the steps of randomly sampling in a search space or generating an initial population based on priori knowledge according to an analyzed target; The method comprises the steps of simulating microstructures generated by each individual in a population to obtain actual physical attributes, calculating residual vectors between the actual attributes and target attributes, monitoring performance improvement rates of targets in each physical field in an iterative optimization process, calling priority weights in the target vectors when iteration stagnation caused by cross-physical field conflicts is detected, and executing self-adaptive weight update, wherein for physical targets with the improvement rates lower than a threshold and higher priority weights, the weight coefficients in multi-target quantification are dynamically amplified.
5. The multi-agent-based reverse design method for cross-physical-field microstructures, as set forth in claim 1, is characterized in that the method comprises performing elite screening by utilizing non-dominant ordering driven by pareto, dynamically adjusting adjustment parameters of potential space, and guiding the evolution process to converge towards the physically effective pareto front, wherein the process comprises the steps of performing disturbance sampling near the current parameters, collecting performance feedback in the vicinity, constructing a local proxy model by utilizing a weighted least square method, and estimating pseudo-gradient direction under non-microlandscapes; guiding parameter updating by adopting a symbol function switching mechanism, searching along the gradient forward direction when the performance does not reach the standard, and correcting reversely when overshooting, so as to realize self-adaptive fine adjustment of potential space adjusting parameters; And performing pareto non-dominant sorting on the updated population, retaining leading-edge individuals with optimal performance in the multi-physical-field conflict, and eliminating individuals which do not meet the hard constraint to form a next-generation population until the iteration requirement is met.
6. The multi-agent based reverse design method across physical field microstructures as in claim 1 for design variables during dynamic adjustment of latent space tuning parameters Is the first of (2) The update formula is defined as: Wherein, the The value before the update is made, For the value of the physical attribute of the object, For the attribute values obtained for the current simulation, Is a sign function, is used to determine the direction of optimization, For the adaptive step size, greater than the set point at the early stage of the search to encourage exploration, reduced later to fine tune, Noise is explored for gaussian to prevent searches from falling into local optima.
7. The multi-agent based reverse design method for micro-structures across physical fields according to claim 1, wherein the iteration requirement is to calculate single candidates while satisfying And (3) the proportion of each physical target, evaluating the proportion reaching a preset error range in all independent target tasks, evaluating the average relative error, and integrating the three to obtain a final score, and stopping iteration if the final score tends to be stable or reaches the upper limit of the evolution algebra.
8. A multi-agent-based trans-physical field microstructure reverse design system is characterized by comprising: The semantic intention analysis and task collaborative planning module is configured to acquire a design intention containing natural language description, generate and call an external physical knowledge base through retrieval enhancement, and analyze the design intention into a quantized cross-physical field attribute target vector; The symbolic design space initialization module is configured to map design intentions into adjustment parameters in a potential representation space by utilizing a pre-trained neural network model based on the analyzed quantized cross-physical field attribute target vector, and decode and generate a geometric configuration of an initial microstructure to construct an initial evolution population; The multi-physical-field closed-loop simulation module is configured to perform cross-physical-field numerical simulation on microstructures in the population, and extract performance data in real time; the simulation-aware evolution search optimization module is configured to identify the evolution direction of a parameter space through local gradient approximation based on numerical deviation obtained by cross-physical field numerical simulation, perform elite screening by utilizing non-dominant ordering driven by pareto, dynamically adjust the adjustment parameters of a potential space, and guide the evolution process to converge towards the physical effective pareto front; And the optimization result output module is configured to screen an optimal microstructure geometric model meeting multi-objective constraint across a physical field according to the optimization result to obtain a design scheme.
9. A computer readable storage medium storing computer instructions which, when executed by a processor, perform the steps of the method of any of claims 1-7.
10. An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps in the method of any one of claims 1-7.

Description

Multi-agent-based reverse design method and system for cross-physical-field microstructure Technical Field The invention belongs to the technical field of intelligent design and calculation optimization, and particularly relates to a multi-agent-based reverse design method and system for a microstructure across a physical field. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. The porous microstructure material can realize light weight and multifunction through a geometric structure. Has been applied to the fields of aerospace lightweight structures, energy absorbing devices, biomedical implants, and the like. For such materials, it is often necessary to reverse the microstructure topology or geometry parameters from the target properties, i.e., reverse design of the microstructure, given a number of physical property index constraints such as volume fraction, stiffness, thermal conductivity, etc. However, cross-physical field microstructure design is a very challenging optimization problem in that its search space is high-dimensional and discontinuous, and the objective function derived from coupling cross-physical field simulation tends to be non-micromanipulable and very costly to evaluate. The current microstructure design method mainly comprises a structure generation method based on a depth generation model or a proxy model, an iteration solving method based on topological optimization and a searching method based on genetic algorithm or other black box optimization. Existing depth generation models typically employ "single" probability mapping, lacking the inherent ability to infer to handle conflicting cross-physical field constraints, i.e., the generated structure, while visually reasonable, is physically ineffective. Traditional topological optimization has strict theoretical deduction, but cannot be solved when facing to non-tiny physical targets, and is easy to fall into a local optimal solution for multi-target optimization and has huge calculation cost. Genetic algorithms or other black box optimized search methods can be optimized using simulation feedback, but their operators often lack the ability to express and constrain the design semantics or high-level intent, making it difficult to stably map a designer-given functional preference or language description to a controllable structural parameter, resulting in a serious mismatch between the designer's natural language description's design intent and the ultimately generated parameters. Furthermore, existing design processes typically rely on expert empirical intervention and high-frequency high-fidelity simulation, which greatly limits the scalability and interactive exploration of the design. Most existing learning methods operate under open loop settings, rely heavily on proxy model approximation rather than direct physical verification, and have difficulty in robustly performing coupled cross-physical field constraints. Disclosure of Invention In order to solve the problems, the invention provides a multi-agent-based reverse design method and system for a microstructure across a physical field, the invention realizes the reverse design of the structure based on the collaborative search of multiple agents and the combination of simulation feedback, and realizes the efficient and accurate design of the microstructure under complex constraint. According to some embodiments, the present invention employs the following technical solutions: A multi-agent-based reverse design method for a cross-physical-field microstructure comprises the following steps: Acquiring design intents containing natural language descriptions, generating and calling an external physical knowledge base through retrieval enhancement, and analyzing the design intents into quantized cross-physical field attribute target vectors; Based on the analyzed quantized cross-physical field attribute target vector, mapping the design intent into adjustment parameters in a potential representation space by utilizing a pre-trained neural network model, decoding to generate a geometric configuration of an initial microstructure, and constructing an initial evolution population; performing cross-physical field numerical simulation on microstructures in the population, and extracting performance data in real time; Based on numerical deviation obtained by cross-physical field numerical simulation, the evolution direction of a parameter space is approximately identified through local gradients, elite screening is carried out by utilizing non-dominant ordering driven by pareto, the adjustment parameters of potential space are dynamically adjusted, and the evolution process is guided to converge towards the physical effective pareto front; And screening an optimal microstructure geometric model meeting multi-objective constraint across a physical field according to the optimizatio