Large Language Model Agent for User-friendly Chemical Process Simulations

Abstract: Modern process simulators enable detailed process design, simulation, and optimization; however, constructing and interpreting simulations is time-consuming and requires expert knowledge. This limits early exploration by inexperienced users. To address this, a LLM agent is integrated with AVEVA Process Simulation (APS) via Model Context Protocol (MCP), allowing natural language interaction with rigorous process simulations. An MCP server toolset enables the LLM to communicate programmatically with APS using Python, allowing it to execute complex simulation tasks from plain-language instructions. Two water-methanol separation case studies assess the framework across different task complexities and interaction modes. The first shows the agent autonomously analyzing flowsheets, finding improvement opportunities, and iteratively optimizing, extracting data, and presenting results clearly. The framework benefits both educational purposes, by translating technical concepts and demonstrating workflows, and experienced practitioners by automating data extraction, speeding routine tasks, and supporting brainstorming. The second case study assesses autonomous flowsheet synthesis through both a step-by-step dialogue and a single prompt, demonstrating its potential for novices and experts alike. The step-by-step mode gives reliable, guided construction suitable for educational contexts; the single-prompt mode constructs fast baseline flowsheets for later refinement. While current limitations such as oversimplification, calculation errors, and technical hiccups mean expert oversight is still needed, the framework's capabilities in analysis, optimization, and guided construction suggest LLM-based agents can become valuable collaborators.

• The paper introduces an LLM-driven agent integrated with APS via MCP, enabling conversational chemical process simulation.
• It demonstrates robust flowsheet analysis and autonomous optimization in a water-methanol separation benchmark using detailed tool calls.
• Results show practical performance gains and energy savings while emphasizing the need for expert oversight and error mitigation.

LLM-Driven Agents for Chemical Process Simulation: Framework, Evaluation, and Outlook

Introduction and Motivation

Contemporary chemical process simulators are foundational in engineering practice, enabling high-fidelity design, analysis, and optimization of industrial systems. Despite mature simulator technologies, such as AVEVA Process Simulation (APS), these platforms present steep learning curves and high entry barriers due to complex configuration requirements and the need for deep domain expertise. This often limits accessibility for junior engineers and interdisciplinary teams. The rapid emergence of LLMs, particularly their deployment within agentic AI systems, presents an opportunity to democratize access to these simulators by translating natural language instructions into programmatic simulation tasks.

The paper "LLM Agent for User-friendly Chemical Process Simulations" (2601.11650) addresses the integration of LLM-driven agents with rigorous chemical process simulators. The central thesis is the design and evaluation of an LLM agent, coupled with APS via the Model Context Protocol (MCP), to support both experienced and novice users in process simulation through conversational, tool-using AI.

Framework Architecture and Integration

The proposed architecture replaces direct, manual interaction with the process simulator by a conversational LLM interface leveraging a strictly defined protocol and curated toolset. The MCP provides the glue between the LLM agent (acting as MCP client) and the process simulator (exposed programmatically via a Python-based MCP server). This design ensures clear separation between the cognitive functions (LLM reasoning, context management, tool orchestration) and deterministic simulation logic (handled by the APS backend). Claude Sonnet 4.0 is selected as the LLM agent for its multi-step tool use reliability and broad context window. The FastMCP Python framework enables rapid MCP server development and tool definition.

Figure 1: Schematic view of the LLM-based agentic framework, highlighting the division between conversational reasoning (LLM), protocol interface (MCP), and deterministic simulation (APS).

The MCP server exposes a suite of high-level tools to the LLM agent. These abstract over detailed simulator API calls and are readily extensible for future workflows (e.g., optimization routines or troubleshooting). This architecture both enforces engineering rigor—leveraging APS's physics-based solvers for all calculations—and ensures modularity and future-proofing as LLM and simulation technologies evolve.

Benchmark Case: Water-Methanol Separation Process

As a canonical testbed, the authors utilize the "C1 - Water Methanol Separation" process, modeled via distillation in APS. This binary system is chosen for its relevance (industrial prevalence of methanol/water mixtures), underlying non-ideal VLE, and absence of azeotrope, which collectively stress the agent's interpretive, analytical, and synthesis capabilities. The flowsheet comprises a feed source, a distillation column with internal condenser and reboiler, and product sinks for distillate and bottoms.

Figure 2: Detailed APS simulation flowsheet for "C1 - Water Methanol Separation" highlighting model connectivity and key process modules.

Case Study 1: Flowsheet Analysis and Autonomous Optimization

Simulation Understanding and Summarization

The agent is tasked to extract flowsheet topology, process variables, and deliver a consolidated technical summary solely from a minimal high-level prompt. Results demonstrate that the agent autonomously identifies equipment, stream connectivity, operating conditions, and process parameters by sequencing tool calls (seven for initial summary; accessing 356 out of 2006 available simulation variables).

Critically, all reported quantitative values were numerically correct and the variable selection was contextually appropriate. The summary included equipment configuration, process data (stream compositions, flows), energy demands, and a convergence check. However, the agent displayed limitations: it ascribed "optimal" status to the feed stage location without optimization, overstated separation quality, and did not critique unrealistic tray efficiency assumptions. This illustrates that, while the agent outperforms simple retrieval, user oversight remains essential to avoid over-interpretation or unsupported process evaluations.

Process Improvement Recommendations

When asked for separation improvement strategies, the agent exhibited high autonomy, broad technical reasoning, and output diversity. Suggestions fell into four classes: configuration (e.g., increase stages, feed stage tuning), operating parameters (reflux, pressure, feed conditioning), advanced flowsheeting (heat integration, double effect distillation), and hardware (tray design, control systems). Notably, all 11 distinct recommendations were technically sound or, at minimum, provided reasonable directions, with only a single suggestion classified as potentially misleading and none as fundamentally incorrect.

The agent's prioritization of "quick wins" (feed preheat, reflux optimization) aligns with standard engineering practice, and it offered estimated performance improvements, e.g., energy savings of 15–25% for heat integration and purity improvement to >95% methanol in distillate. However, some quantitative claims were not simulation-verified and likely derived from training data priors; this lack of explicit uncertainty annotation is a key limitation.

Autonomous Optimization

Upon being directed to adjust the reflux ratio to achieve >95% methanol purity, the agent orchestrated a feedback-driven, multi-step optimization: querying baseline data, incrementally updating the reflux ratio, and summarizing achieved purity, flows, and energy demands. Convergence was reached with a reflux ratio of 1.45, achieving 95.1% methanol purity and a 6% increase in energy duty relative to the base case, at the cost of a modest reduction in distillate yield. This demonstrates robust tool use and experiment planning. However, small arithmetic errors in change calculations and non-validated economic impact claims highlight the vulnerability of LLMs to minor but impactful mistakes without external computational tool support.

Case Study 2: Flowsheet Synthesis—Stepwise vs. Autonomous Modes

Two synthesis modes were benchmarked:

1. Stepwise Dialogue Construction: The user incrementally instructs the agent (tool call granularity: single action per turn). The agent reliably mapped these instructions to APS tool invocations, maintained conversational context, and successfully produced a functioning process flowsheet with correct stream and model connectivity.

   Figure 3: Simulation flowsheet generated via agentic step-by-step user-guided interaction, confirming correct component addition and connections.

2. Single-Prompt Autonomous Construction: The agent was provided a high-level target process (feed specification, desired separation) and required to autonomously plan simulator setup, model instantiation, fluid package configuration, and process parameters. The agent succeeded in producing a valid flowsheet with minimal user intervention but exhibited several technical errors: setting unspecifiable variables, minor deviations from APS workflow conventions, and overlooking the interdependence of some configuration steps. The agent adopted reasonable default parameters and provided operation summaries, but, as in the analysis case, some intermediate quantitative reasoning lacked robustness or was contextually misplaced.

Key Observations and Strong Claims

• The agent was able to extract, synthesize, and communicate information drawn from hundreds of simulation parameters, filtering for engineering relevance, and all reported values were correct when strictly obtained via simulator tools.
• All improvement and optimization recommendations were at least broadly valid or practically useful, with the majority matching best-practices in process design.
• Single-prompt (autonomous) synthesis is achievable but currently bounded by technical errors in action planning and inability to recover from workflow deviations or ambiguous simulator settings.
• Systematic, iterative, multi-tool workflows (e.g. parameter sweeps for optimization) were realized autonomously without explicit intermediate user guidance.

Limitations and Areas for Improvement

• Numerical calculations not directly tied to tool outputs (e.g. arithmetic on retrieved values) are error prone. This calls for LLM integration with explicit computational tools or interpreter modules to remove such failure points.
• The agent sometimes overclaims or omits uncertainty in performance improvements, e.g., reporting industry-typical but not simulation-specific energy savings, undermining the trustworthiness for high-stakes design.
• The framework is constrained in addressing flowsheet synthesis failure modes (e.g., convergence issues, ill-posed specifications) and cannot autonomously resolve troubleshooting beyond the exposed toolset.

Vision: Multi-Agent Extension and Industrial-Grade Scalability

To overcome the scalability and reliability limitations observed in single-agent workflows, the authors propose an extension towards structured multi-agent frameworks. These would separate concerns (e.g., thermodynamic property assignment, equipment configuration, troubleshooting) across specialized agents, potentially coordinated via orchestration protocols and augmented with RAG (retrieval-augmented generation) layers for dynamic domain knowledge and process memory incorporation.

Figure 4: Extended architecture proposal utilizing MAS and RAG for scalable, robust, and context-aware process simulation agents.

Such advancements could enable context-filtered reasoning, persistent design memory, advanced troubleshooting, and seamless interface with optimization engines, towards full-lifecycle, scalable simulation copilot systems.

Implications and Future Directions

The successful integration of LLM agents with deterministic process simulators has direct educational and industrial implications. For novices, these systems can guide users through complex workflows, lower knowledge barriers, and serve as interactive pedagogical agents. For practitioners, they enable routine automation, accelerate data extraction and preliminary optimization, and support ideation with broad process improvement suggestions. However, expert oversight remains mandatory, both to counteract common LLM failure modes (hallucinations, spurious generalizations) and ensure compliance with domain rigor and safety constraints—concerns widely echoed in the literature [bender_dangers_2021, huang_survey_2025].

The demonstration on a binary separation flowchart represents only the early stage. Open research avenues include scaling to multi-unit, recycle-intensive flowsheets; integration with superstructure generation and stochastic optimization [quaglia_systematic_2015, vollmer_synergistic_2022, al_stochastic_2020]; and deployment of robust, MAS-driven architectures for industrial applications [rupprecht_multi-agent_2025; zeng_llm-guided_2025]. Advances in tool-use infrastructure, context management, and domain-specific RAG are required to close the reliability and accuracy gap for fully autonomous engineering design.

Conclusion

This study provides clear evidence that LLM-driven, tool-using agent frameworks can significantly enhance the interpretability, accessibility, and preliminary automation of commercial chemical process simulators, while maintaining deterministic safety via programmatic protocol boundaries. Strong numerical results in process understanding, solution synthesis, and iterative optimization evidence both immediate practitioner utility and the promise for deeper, automated engineering workflows. However, deployment beyond supervised scenarios necessitates architectural advances, improved error mitigation, and persistent human oversight. The framework cements LLM agents as collaborative process engineering tools—complementary co-pilots rather than autonomous replacements—for the foreseeable future.