From Sketch to Simulation: Revolutionizing Chemical Process Design with Multi-Agent AI

      The modern chemical engineering landscape relies heavily on process simulation for everything from initial design to operational optimization. Yet, the critical step of transforming early-stage process sketches into functional, executable simulation models has remained a significant bottleneck. This manual, expertise-intensive process often demands countless hours of work, deep simulator-specific knowledge, and repetitive refinements, creating a considerable hurdle to efficiency and innovation.

      Recent advancements in Artificial Intelligence (AI) have introduced exciting possibilities, particularly in the realm of generative AI for interpreting engineering diagrams and Large Language Model (LLM)-assisted content creation. However, these two powerful areas have largely operated in isolation. Diagram interpretation tools typically extract visual elements but stop short of creating a full simulation, while text-to-simulation workflows require already structured data, not raw visual information. The core challenge lies in bridging this gap: converting the often-ambiguous visual information of a diagram into precise, machine-interpretable simulation components that adhere to strict creation, connectivity, and initialization rules.

The Challenge: Bridging Sketch to Simulation

      Process diagrams serve as the universal language for engineers, visually encoding crucial information such as unit operations, material streams, and the overall layout of a process system. They are designed for human understanding, facilitating communication of design intent and structural topology. However, the information presented in these diagrams often differs significantly from the explicit, structured definitions required by process simulation software like Aspen HYSYS. Key details might be missing, implied, or ambiguous in a sketch, demanding significant human interpretation, inference, and iterative refinement to produce a valid simulation.

      This divergence means that merely extracting symbols and labels from a diagram is insufficient. A successful automated system must perform a structured synthesis process, organizing, supplementing, and translating the diagram’s content into a format fully compatible with the chosen simulation environment. This requirement underscores the need for sophisticated AI that can not only "see" but also "understand" and "engineer" based on visual inputs, a critical step toward broader digitalization in process systems engineering (PSE).

Introducing Multi-Agent LLMs for Process Automation

      To address this complex challenge, researchers have developed an innovative end-to-end multi-agent large language model system. This system is engineered to directly convert process diagrams into executable Aspen HYSYS flowsheets, offering a transformative solution for automating a historically manual process. By deploying a network of specialized AI agents working in concert, the system aims to automate what previously required extensive human intervention, accelerating design cycles and enhancing reliability.

      This multi-agent approach represents a significant leap in AI’s capability to handle intricate engineering tasks. Unlike monolithic AI systems, a multi-agent architecture breaks down the problem into smaller, manageable sub-tasks, each handled by an expert agent. This decomposition is crucial for complex industrial applications, allowing for greater transparency in decision-making and more robust error handling. For organizations seeking custom AI solutions that navigate intricate data landscapes and specific operational protocols, this model of specialized AI collaboration provides a powerful framework.

How the System Works: A Layered Approach

      The system operates through three carefully coordinated layers, each with distinct functions:

• Diagram Parsing and Interpretation: This initial layer focuses on visually interpreting the raw process diagram. Specialized agents analyze the sketch, identifying unit operations, streams, and their labels. It's like having an expert eye that can discern the individual components and their initial relationships from a hand-drawn or digital image. This goes beyond simple image recognition, aiming to extract the underlying engineering intent.
• Simulation Model Synthesis: Once the visual information is interpreted, the system constructs a structured, graph-based intermediate representation. Think of this as creating a detailed digital blueprint of the process, translating visual elements into machine-understandable data points. Following this, other agents generate the specific code required to build the simulation model within the Aspen HYSYS COM interface. This involves applying simulator-specific rules for object creation, connectivity, and initialization, ensuring the generated model is not just a collection of parts but a functionally coherent system.
• Multi-level Validation: The final, crucial layer involves executing the generated HYSYS model and performing rigorous structural verification. This step is vital for "execution grounding," meaning the AI's output is constantly checked against real-world simulation rules. This multi-level validation significantly reduces the risk of AI "hallucination"—where the AI generates plausible but incorrect information—by pinpointing errors explicitly, making the system's failure modes transparent and easier to debug. This focus on verifiable output is a hallmark of robust AI deployment in critical industrial settings.

Real-World Impact and Verified Performance

      The effectiveness of this multi-agent framework was rigorously tested on four chemical engineering case studies, ranging from a basic desalting process to a complex industrial aromatic production flowsheet featuring multiple recycle loops. The results were compelling:

• The system successfully produced executable Aspen HYSYS models for all cases.
• For the two simpler cases, it achieved complete structural fidelity (an F1 score of 1.00 across all metrics), meaning the AI model perfectly replicated the diagram's structure.
• Even for the more complex cases, the system maintained high performance, demonstrating connection consistency of at least 0.93 and stream consistency of at least 0.96. This indicates a very high accuracy in correctly establishing connections and defining material flows within the simulation.

      An ablation analysis, which involves testing the system by removing individual components, confirmed that each architectural element plays a significant role in the overall robustness and reliability, with its importance increasing with diagram complexity. These findings highlight the viability of end-to-end sketch-to-simulation automation, positioning it as a key driver for digital transformation in process systems engineering. ARSA, with its team experienced since 2018, understands the critical importance of such verifiable and robust AI systems in production environments across various industries.

Future Horizons and Remaining Hurdles

      While the multi-agent LLM system marks a significant milestone, the research acknowledges ongoing challenges that offer fertile ground for future development. These include effectively handling:

• Dense Recycle Structures: Complex processes often involve intricate feedback loops, which remain difficult for AI to fully interpret and synthesize accurately.
• Implicit Diagram Semantics: Many engineering diagrams contain subtle cues or unwritten conventions that human engineers understand intuitively but are challenging for AI to infer.
• Simulator-Interface Constraints: The limitations and specific requirements of how simulation software interacts with external tools can still pose integration difficulties.

      Overcoming these challenges will further refine the automation capabilities, pushing the boundaries of what is possible in AI-driven process design.

      The advent of multi-agent LLM systems for automating flowsheet generation promises a future where chemical engineers can dramatically reduce manual effort, accelerate design cycles, and focus on higher-value tasks. By leveraging advanced AI to bridge the gap between conceptual sketches and executable simulations, industries can move closer to fully digitalized, highly efficient, and more reliable engineering workflows. This kind of practical AI, transforming visual data into actionable intelligence, mirrors the core mission of companies like ARSA, who deploy advanced technologies, such as AI Video Analytics and edge AI systems, to solve complex operational challenges.

      Source: Bahamdan, A., Pajak, E., Hedengren, J. D., & del Rio Chanona, A. (2026). Sketch2Simulation: Automating Flowsheet Generation via Multi Agent Large Language Models. arXiv preprint arXiv:2603.24629.

      Ready to explore how advanced AI and IoT solutions can transform your operational challenges into strategic advantages? Contact ARSA today for a consultation.