cs.AIMar 2, 2026

CeProAgents: A Hierarchical Agents System for Automated Chemical Process Development

Authors: Yuhang Yang, Ruikang Li, Jifei Ma, Kai Zhang, Qi Liu, Jianyu Han, Yonggan Bu, Jibin Zhou, 

Organizations: State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China · Dalian Institute of Chemical Physics, Chinese Academy of Sciences, Dalian, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The development of chemical processes, a cornerstone of chemical engineering, presents formidable challenges due to its multi-faceted nature, integrating specialized knowledge, conceptual design, and parametric simulation. Capitalizing on this, we propose CeProAgents, a hierarchical multi-agent system designed to automate the development of chemical process through collaborative division of labor. Our architecture comprises three specialized agent cohorts focused on knowledge, concept, and parameter respectively. To effectively adapt to the inherent complexity of chemical tasks, each cohort employs a novel hybrid architecture that integrates dynamic agent chatgroups with structured agentic workflows. To rigorously evaluate the system, we establish CeProBench, a multi-dimensional benchmark structured around three core pillars of chemical engineering. We design six distinct types of tasks across these dimensions to holistically assess the comprehensive capabilities of the system in chemical process development. The results not only confirm the effectiveness and superiority of our proposed approach but also reveal the transformative potential as well as the current boundaries of Large Language Models (LLMs) for industrial chemical engineering.

Figures & tables

Figure 1: Overview of CeProAgents and CeProBench . a, The system processes heterogeneous chemical engineering inputs, including scientific documents, PFD imagery defining equipment and connections, and parameter files for simulation. b, The CeProAgents architecture comprises three specialized cohorts: the Knowledge Cohort retrieves chemical engineering knowledge from heterogeneous sources; the Concept Cohort bridges text captions, structural graphs, and PFD imagery; and the Parameter Cohort optimizes operational configurations for yield and economic targets. c, Representative multi-modal engineering results span three domains: retrieved kinetics and property data; designed and parsed PFD schematics; and optimized industrial parameters for yield and cost. d, The CeProBench suite probes six atomic tasks across the dimensions of Knowledge (Extract and Augment), Concept (Parse, Complete, and Design), and Parameter (Optimize).
Figure 2: Chemical Knowledge Enhancement and Structured Extraction via CeProAgents . a, Quantitative assessment of knowledge enhancement performance across four dimensions: Correctness, Rationality, Clarity, and Completeness. b, Ablation studies demonstrating the impact of multi-source integration; scores are compared across Full configuration and versions excluding the Knowledge Base (w/o KB), Knowledge Graph (w/o KG), and Web Search (w/o Web). c, Case study of multi-agent collaborative retrieval for industrial production queries, illustrating the interplay between Web, KB, KG, and Report Agents. d, Statistical distribution of knowledge extraction performance, showing Entity Metrics (accuracy, recall, F1) and Graph Metrics (MEC, MED). e, Representative workflow of structured knowledge extraction, transitioning from unstructured academic text to a merged, disambiguated semantic graph.
Figure 3: Process Concept Parsing, Completion, and Design via CeProAgents . a, Quantitative assessment of semantic parsing performance, showing accuracy and recall metrics for equipment and connection identification across different model backbones. b, Evaluation of generative completion capabilities using Top-K accuracy metrics, highlighting the system’s proficiency in inferring missing process components compared to baseline models. c, Representative case study illustrating the complete workflow from semantic parsing of raw PFDs to the logical completion of underspecified topologies. d, Evaluation of generative design performance comparing the Full configuration against an ablation without the Correction Agent (w/o). e, Detailed case study of iterative generative design, demonstrating how the system refines initial proposals through a multi-turn correction loop to satisfy complex engineering constraints.
Figure 4: Industrial Process Parameter Optimization via CeProAgents . a, Quantitative assessment of multi-objective optimization performance. The radar chart illustrates the normalized scores across five critical dimensions including Cost, Yield, Purity, and Iteration efficiency. b, Comparative analysis of optimization efficacy stratified by distinct unit operation types (Separation, Synthesis, and Combined). c, Evaluation of convergence efficiency illustrating the distribution of total optimization iterations and the specific iteration steps where the global optimum was identified. The box plots coupled with scatter points highlight the search speed and decision-making decisiveness of different backbones. d-f, Case studies of autonomous parameter optimization across distinct chemical processes. d, Separation process for aromatics purification involving a multi-column distillation sequence. e, Synthesis process for MMA production. f, Combined process for butadiene hydrocyanation to produce 3-PN.
Figure 5: Overall performance assessment of CeProAgents on CeProBench . a, The Knowledge dimension evaluation comparing epistemic fidelity and structural extraction capabilities. The radar chart aggregates metrics for reasoning quality (Correctness, Rationality, Clarity, Completeness) and graph construction integrity (Accuracy, Recall, F1, MEC, MED). b, The Concept dimension evaluation assessing multimodal topological reasoning across three distinct tasks: semantic parsing (Equipment/Connection Accuracy and Recall), topological completion (Top-K Accuracy), and generative design (Valid Rate and Correct Rate). c, The Parameter dimension evaluation quantifying parameter optimization efficacy. The chart contrasts optimization outcomes (Yield, Purity, Effective, Cost, Overall) with computational efficiency (Total vs. Best Iterations) and the aggregate Effective Score.
Appendix figures & tables35 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Extended experimental analysis of knowledge enhancement and extraction performance. a, Stratified evaluation of the Knowledge Augment task across four specific chemical engineering sub-domains: Process Route Selection, Catalyst & Condition Selection, Production Method & Process Flow, and Separation Method Selection, displaying performance scores for Correctness, Rationality, Clarity, and Completeness across five foundation models. b, Statistical distribution of Knowledge Extraction metrics comparing different efficient backbone models, illustrating the density and variance for Predicted Entity Counts (left), Mapping-based Edge Distance (MED) Coverage (center), and Average Path Length of the constructed graphs (right).
Figure S2: Extended experimental analysis of concept parsing, completion, and generative design. a, Distribution of matched edges versus matched nodes for concept parsing tasks across individual samples. b, Average count of successfully matched nodes and edges for each model. c, Variation of parsing precision and recall metrics as a function of the similarity threshold ( τ\tau ). d, Parsing results for masked graph components in the topology completion task. e, Comparative distribution of topological complexity (connections versus equipment) between the initial generation step and the final iteration. f, Number of iterations required for the generative design process across distinct sample tasks.
Figure S3: Extended quantitative statistics for parameter optimization across distinct chemical process types. a-c, Stratified analysis of optimization performance and search efficiency: a, Combined processes; b, Separation processes; and c, Synthesis processes. In each panel, the left bar charts display the normalized performance scores across five key metrics, while the right scatter plots correlate the iteration step where the global optimum was identified versus the total iterations performed. d, Line graphs tracking the absolute performance trajectories for primary objectives, including Purity (Separation), Yield (Synthesis), and Effective Score (Combined), across individual benchmark cases. e, Longitudinal analysis of economic optimization displaying the raw Operating Cost values recorded across 20 distinct test cases for different model backbones.
Figure S4: Comprehensive statistical profiling of the CeProBench . a, Statistical characteristics of the Knowledge dimension corpus across industrial sub-domains, illustrating the distribution of document length, the density of entities and triples, and their correlation, validating semantic consistency. b, Composition of the Q&A dataset, detailing the distribution of task types and difficulty scores, highlighting the prevalence of high-complexity reasoning tasks in specific categories. c, Topological complexity of the PFD dataset, quantified by the count of equipment nodes and connection edges per sample. d, Frequency distribution of unit operations, contrasting the overall equipment distribution with the subset targeted for topology completion tasks, reflecting a focus on critical process units. e, Difficulty stratification of the generative text-to-design scenarios. f, Dimensionality analysis of the Simulation scenarios, showing the distribution of optimization parameters and objectives. g, Detailed breakdown of optimization parameters and objectives, emphasizing the benchmark’s focus on separation efficiency and economic optimization.
Figure 10
Figure 11
Figure 12
Figure 13
Figure 14
Figure 15
Figure 16
Figure 17
Figure 18
Figure 19
Figure 20
Figure 21
Figure 22
Figure 23
Figure 24
Figure 25
Figure 26
Figure 27
Figure 28
Figure 29
Figure 30
Figure 31
Figure 32
Figure 33
Figure 34
Figure 35
Figure 36
Figure 37
Figure 38
Figure 39
Figure 40
More in this topic

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,039 papers

Explore similar work

CardsList
Mar 13, 2026

Context is all you need: Towards autonomous model-based process design using agentic AI in flowsheet simulations

Agentic AI systems integrating large language models (LLMs) with reasoning and tooluse capabilities are transforming various domains - in particular, software development. In contrast, their application in chemical process flowsheet modelling remains largely unexplored. In this work, we present an agentic AI framework that delivers assistance in an industrial flowsheet simulation environment. To this end, we show the capabilities of GitHub Copilot (GitHub, Inc., 2026), when using state-of-the-art LLMs, such as Claude Opus 4.6 (Anthropic, PBC, 2026), to generate valid syntax for our in-house process modelling tool Chemasim using the technical documentation and a few commented examples as context. Based on this, we develop a multi-agent system that decomposes process development tasks with one agent solving the abstract problem using engineering knowledge and another agent implementing the solution as Chemasim code. We demonstrate the effectiveness of our framework for typical flowsheet modelling examples, including (i) a reaction/separation process, (ii) a pressure-swing distillation, and (iii) a heteroazeotropic distillation including entrainer selection. Along these lines, we discuss current limitations of the framework and outline future research directions to further enhance its capabilities.

Pascal Schäfer, Lukas J. Krinke, Martin Wlotzka +1

BASF SE, Ludwigshafen, Germany · RWTH Aachen University, Aachen, Germany

Agentic WorkflowsAgentic Code GenerationMulti-Agent LLM Systems
AskPDF
Aug 2, 2026

CRAFTS: Collaborative Role-Adaptive Fine-Tuning of LLM Agents for Chemical Process Simulation

Constructing an executable chemical-process model remains manually intensive. Chemical engineers translate underspecified requests into coupled decisions about unit operations, thermodynamics, streams, specifications, degrees of freedom (DoF), initialization, solver repair, and optimization; one error can invalidate the model. CRAFTS mirrors the staged workflow of chemical engineers by decomposing simulation building into bounded subtasks assigned to seven bounded roles, with deterministic IDAES/Pyomo gates between stages. Given a natural-language request, process flowsheet diagram (PFD) evidence, and curated chemical-engineering knowledge, Input Understanding and Intent recover requirements, constraints, and process semantics; visual, topology, and specification specialists translate them into typed simulator contracts; and Debug and Optimization support bounded repair and eligible optimization. Fine-tuning is applied to the three schema-critical visual, topology, and specification roles, while the remaining roles use untuned Qwen. The resulting VisualGraphIR, TopologyIR, SpecIR, BuildPlan, and SolveReport expose unit, port, thermodynamic, numerical, and execution decisions. Compatible constructors, property packages, and runners are attached only after semantic artifacts pass engineering gates. We introduce OpenIDAES-450, a 450-case IDAES process- simulation dataset, and evaluate the complete seven-role LangChain/LangGraph workflow through solve and eligible optimization on its frozen 82-case held-out split. CRAFTS completes the prescribed validation and execution contract for for 91.5% of cases and achieves unit, stream, and directed-connection F1 scores of 0.815, 0.791, and 0.782. These results demonstrate the effectiveness of role specialization, typed intermediate representations, and deterministic engineering gates for reliable automated process-model construction.

Ziyun Zhang, Yuxin Lin, Eldin Wee Chuan Lim +1

Department of Chemical & Biomolecular Engineering, National University of Singapore · 2Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, Fujian, China

Fine-TuningLLM Agent OrchestrationLLM-Based Program Synthesis
AskPDF
May 19, 2026

EngiAI: A Multi-Agent Framework and Benchmark Suite for LLM-Driven Engineering Design

Large Language Model (LLM) agents are increasingly applied to engineering design tasks, yet existing evaluation frameworks do not adequately address multi-agent systems that combine simulation, retrieval, and manufacturing preparation. We introduce a benchmark suite with three evaluation dimensions: (1) a workflow benchmark with seven prompt styles targeting distinct cognitive demands-including direct tool use, semantic disambiguation, conditional branching, and working-memory tasks; (2) a Retrieval-Augmented Generation (RAG) benchmark with gated scoring isolating retrieval contributions to parameter selection; and (3) an High Performance Computing (HPC) benchmark evaluating end-to-end ML training orchestration on a SLURM cluster. Alongside the benchmark we present EngiAI, a Multi-Agent System (MAS) reference implementation built on LangGraph that operationalizes the benchmark by coordinating seven specialized agents through a supervisor architecture, unifying topology optimization, document retrieval, HPC job orchestration, and 3D printer control. Across four LLM backends and two EngiBench problems, proprietary models achieve 96-97% average task completion on Beams2D, while open-source 4B-parameter models reach 55-78%, with clear generational improvement. Conditional branching proves most challenging, with task completion dropping to 20-53% for the conditional style on Photonics2D. RAG gating confirms near-perfect retrieval-augmented scores (about 1.0) versus near-zero without retrieval, validating the evaluation design. On HPC orchestration, one model completes all pipeline steps in 100% of runs while another drops to 50%, revealing that multi-step instruction following degrades over long-running workflows.

Gioele Molinari, Florian Felten, Soheyl Massoudi +1

IDEAL Chair of Artificial Intelligence in Engineering Design ETH Zurich Zurich, Switzerland · IDEAL, ETH Zurich, Zurich, Switzerland · Autom8.build, Brussels, Belgium

Multi-Agent LLM SystemsLLM Agent EvaluationRAG Evaluation
AskPDF