A collaborative research team has published a multi-agent large language model framework within Engineering, offering a structured workflow to mitigate persistent reliability gaps in existing LLM-assisted computer-automated design (CAutoD) pipelines. The work, titled LLM-Driven Framework for Industrial Design Automation, addresses core limitations stemming from the black-box characteristics of general large language models, which often lack consistent explainability and robustness when deployed for iterative industrial design tasks spanning conceptual drafting, knowledge-driven refinement and digital prototyping.

The research introduces the LLM-IDA system, a tiered multi-agent architecture split into three vertical functional layers corresponding to distinct design phases. The bottom L0 layer relies on multimodal foundational black-box models to process user natural language input, complete requirement analysis and generate initial conceptual designs through multimodal perception and long-sequence content generation. The middle L1 grey-box layer embeds specialized CAE engineering knowledge graphs and optimization algorithms to translate abstract concepts into constraint-compliant design schematics, narrowing the search space for subsequent simulation-based tuning by encoding physical manufacturing and functional rules into agent decision logic. The top L2 white-box layer establishes direct connections with CAD and CAE software application programming interfaces, deploying LLM code generation capabilities to build parametric digital prototypes and run finite element analysis performance evaluations with automated feedback loops.

Functionally, the full automation pipeline is split into two core modules: a specification module and a code-generation module that incorporates quantification, surrogate and prototyping sub-units. The specification module formalizes unstructured natural language user requests into standardized optimization problem tuples containing defined design variables, solution boundaries, objective functions and constraint sets, distributing segmented tasks to downstream code-generation agents. The three submodules operate in sequential coordination: the quantification module generates executable simulation scripts via iterative coding and visual testing loops based on PyAnsys libraries; the surrogate module builds neural network approximation models through uniform sampling and adaptive data normalization to cut repeated high-fidelity simulation overhead; the prototyping module integrates pre-verified simulation and surrogate interfaces with multi-objective optimization templates such as NSGA-II to produce converged Pareto front solutions.

To counter the hallucination risks that disrupt domain-specific API calls in standard retrieval-augmented generation (RAG) workflows, the team constructed a “Task Analysis–Code Generation–Code Feedback” closed-loop logic. This mechanism maps high-dimensional natural language problem spaces onto low-dimensional structured intent spaces to reduce information entropy, then applies iterative code validation with execution error feedback to lift the probability of generating fully functional engineering scripts. The team adopted the pass@10 evaluation metric across a curated dataset of industrial design cases to benchmark LLM-IDA against conventional RAG workflows, with ablation tests verifying the individual performance contributions of built-in templates, task analysis logic and code testing feedback modules across multiple mainstream LLMs.

The framework’s modular structure maintains domain agnosticism, with researchers noting that substitution of underlying physics simulation kernels allows adaptation to electronic design, energy system and materials engineering use cases. The paper also outlines two primary current limitations: weak spatial geometry perception for complex CAD assemblies and long-chain reasoning instability in general-purpose LLMs. Planned follow-up work includes targeted fine-tuning of open-source language models on CAD and CAE datasets, alongside the development of a unified engineering-focused world model to strengthen physical reasoning capacity for industrial design automation tasks.

The paper “LLM-Driven Framework for Industrial Design Automation,” is authored by Sicheng He, Xiaoxu Wang, Zeke Chen, Jianxing Liao, Bo Wang, Junyan Xu, Xiaohong Guan, Shui Yu, Yun Li. Full text of the open access paper: https://doi.org/10.1016/j.eng.2026.04.009. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.