A Multi-Agent Framework for Zero-Dimensional Reduced-Order Model Planning

Jul 16, 2026

Upload by Eden

High-end complex equipment relies on multi-dimensional workflows in which zero-dimensional reduced-order models (0D ROMs) are central. Traditional 0D model planning depends on manual experience, suffers from limited topology exploration and long iterations, and even genetic-algorithm approaches mainly achieve local parameter tuning. Large-language-model agents enable rapid exploration, yet single-agent planning struggles with long-horizon, tightly coupled 0D ROM tasks. This work introduces Z-COPA, a multi-agent framework combining SAGE graph representation with MILP-guided navigation, validated on aerospace, IEEE power and water-network benchmarks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages47

Reading Time1 h 58 min

Abstract

Keywords

Introduction

arXiv :2607 . 10994v 1 [ cs .LG] 13 Jul 2026 A Multi-Agent Framework for Zero-Dimensional Reduced-Order Model Planning Bingteng Sun 1,2 , Hao Yin 1,2†, Yiling Chen 1,2†, Renjie Xiao 1,2 , Lei Xie 1,2 , Shanyou Wang 1,2 , Ruonan Wang 1,2 , Shubao Chen 1,2 , Qingzong Xu 1,2 , Lin Lu6 , Qiang Du 1,2,3,4,5*, Junqiang Zhu 1,2 1 Advanced Gas Turbine Laboratory, Institute of Engineering Thermophysics, Chinese Academy of Sciences, Beijing, 100190, China. 2 National Key Laboratory of Science and Technology on Advanced Light-duty Gas-turbine, Beijing, 100190, China. 3 , University of Chinese Academy of Sciences, Beijing, 100190, China. 4 , Qingdao Institute of Aeronautical Technology, Qingdao, China. 5 , Nanjing Future Energy System Research Institute, Nanjing, Jiangsu, 211135, China. 6 School of Computer Science and Technology, Shandong University, 72 Binhai Road, Qingdao, 266237, Shandong Province, China. *Corresponding author(s). E-mail(s): [duqiang@iet.cn](duqiang@iet.cn) ; Contributing authors: [sunbingteng@iet.cn](sunbingteng@iet.cn) ; [yinhao@iet.cn](yinhao@iet.cn) ; [chenyiling@iet.cn](chenyiling@iet.cn) ; [xiaorenjie@iet.cn](xiaorenjie@iet.cn) ; [xielei@iet.cn](xielei@iet.cn) ; [wangshanyou@iet.cn](wangshanyou@iet.cn) ; [wangruonan@iet.cn](wangruonan@iet.cn) ; [chenshubao@iet.cn](chenshubao@iet.cn) ; [xuqingzong@iet.cn](xuqingzong@iet.cn) ; [llu@sdu.edu.cn](llu@sdu.edu.cn) ; [zhujunqiang24@sina.com](zhujunqiang24@sina.com) ; †These authors contributed equally to this work. Abstract High-end complex equipment generally adopts multi-dimensional design workflows, among which the development of zero-dimensional reduced-order models (0D ROMs) serves as the core of the entire design process. Currently, this model planning process heavily relies on manual experience, facing bottlenecks such as a limited exploration space for network topologies and long iteration periods. Even with the introduction of traditional optimization methods like the Genetic Algorithm (GA), the process is often restricted to local parameter optimization. In recent years, Large Language Model (LLM) agents, equipped with 1 powerful decision-making and tool-use capabilities, have demonstrated the ability to rapidly explore massive sample spaces, bringing new opportunities to engineering design. Although the Chain of Thought (CoT) and Reason and Act (ReAct) frameworks significantly improve the reliability of agent reasoning, and Retrieval-Augmented Generation (RAG) technology effectively overcomes domain knowledge barriers, a single agent still falls short when dealing with the long-horizon and highly coupled nature of complex 0D ROM planning tasks. To address these challenges, this paper proposes a Zero-dimensional reduced-order model CO-Planning framework based on a multi-Agent architecture (Z-COPA) . This framework is equipped with a Symbolic Action Graph Engine (SAGE) and a MILP-Guided Navigation (MGN) optimizer. Its core innovation lies in proposing a dedicated graph representation method to accurately describe the 0D flow network topology, thereby innovatively converting the traditional empirical model planning process into a rigorous graph structure optimization problem. We validated the forward and inverse design capabilities as well as generalization performance of Z-COPA on three types of cases, including two real aircraft engine secondary-air system design cases, two IEEE power-distribution reconfiguration benchmarks, and two water-distribution network design benchmarks. The results show that our multi-agent system achieves superior task completion quality, obtaining the best performance in both forward and reverse design of air systems. It also cuts the active power loss of the IEEE 69-bus distribution network by 55 .46% and reduces the cost for the Two-Loop water network by 90 .45% . The proposal of Z-COPA disrupts the traditional 0D model planning paradigm, providing a brand-new technical approach for explori

FAQs

What limitations in current 0D reduced-order model planning does Z-COPA target?

It targets manual-experience dependence, limited exploration of network topologies, long iteration cycles, and the tendency of traditional optimization (e.g., GA) to focus on local parameter optimization rather than global planning.

How does Z-COPA represent and optimize 0D flow network topologies?

It introduces a dedicated graph representation method to accurately encode the 0D flow network topology, converting the empirical planning process into a rigorous graph-structure optimization problem.

What results does the framework achieve on benchmark cases?

Z-COPA demonstrates strong forward and inverse design capability and generalization. It yields best performance on air-system forward/reverse design, reduces active power loss of the IEEE 69-bus network by 55.46%, and lowers the cost for the Two-Loop water network by 90.45%.

Jul 16, 2026

Upload by Eden