AI Turns Chemical Sketches into Working Simulations

Chemical engineers spend countless hours manually converting process diagrams into simulation models, a bottleneck that slows design and innovation. This tedious task requires simulator-specific expertise and iterative refinement, often taking days or weeks for complex industrial flowsheets. Now, researchers have developed an AI system that automates this conversion, transforming raw visual diagrams directly into executable Aspen HYSYS simulations with high accuracy. The breakthrough could accelerate engineering workflows, support digitalization in process systems engineering, and make simulation more accessible by reducing the need for specialized manual coding.

The key finding is that a multi-agent large language model system can successfully generate executable process simulation models from visual diagrams across a range of complexities. In four case studies, the system produced working HYSYS models in all instances, achieving perfect structural fidelity (F1 = 1.00) for simpler processes like desalting and Merox sweetening. For more complex industrial flowsheets, such as atmospheric crude oil distillation and aromatic production, performance remained high, with connection consistency at 0.93 and 0.98, and stream consistency at 1.00 and 0.96, respectively. This demonstrates that the AI can handle dense recycle loops and intricate topologies, though minor errors occur in the most challenging scenarios.

Ology employs a three-layer multi-agent architecture to decompose the task into specialized stages. The Diagram Parsing and Interpretation Layer uses a multimodal LLM, Gemini 3 Flash, to describe and extract elements from the diagram, creating an intermediate graph representation. The Simulation Model Synthesis Layer then translates this graph into executable code via coding agents like Qwen2.5-Coder-7B and Qwen3-Coder-30B, which interact with Aspen HYSYS through its COM interface. The Multi-level Validation Layer adds checks at each step, including description validation and execution fixing, to improve reliability and error localization. This modular approach limits hallucination risks and keeps failure modes transparent, as shown in Figure 1 of the paper.

Analysis reveals strong performance across case studies of increasing complexity. For the desalting process (Case Study 1), the system achieved full consistency in units, streams, connections, and materials, as shown in Figure 7, with the model executing without errors. In the Merox process (Case Study 2), it correctly inferred missing connections, maintaining perfect scores. For the atmospheric distillation (Case Study 3), connection consistency dropped slightly to 0.93 due to simulator-interface limitations with side-draw streams, not interpretation errors. The aromatic production process (Case Study 4), the most complex, saw minor deviations in stream and connection consistency but still executed successfully, indicating robustness even in industrial-scale flowsheets. Ablation studies confirmed that each agent contributes meaningfully, with sensitivity increasing alongside diagram complexity.

Are significant for chemical engineering practice, as this system bridges a long-standing gap between diagram understanding and executable model generation. By automating a manual bottleneck, it could reduce design time, enhance reliability, and support emerging paradigms like digital twins and real-time optimization. The approach is industrially relevant, targeting Aspen HYSYS, a widely used commercial simulator, and handles real-world diagrams with variations in notation and completeness. However, the paper notes that performance depends on diagram quality, with implicit elements or dense annotations posing s, and simulator constraints like sequential model construction can affect executability.

Limitations include sensitivity to diagram formatting, such as unlabeled units or implicit connections, which can increase ambiguity. Simulator-specific restrictions, like predefined templates for distillation columns, may limit dynamic stream assignments through the automation interface. The system also relies on carefully designed prompts and instruction files, so transferring it to other simulators or domains may require adaptation. Infrastructure-wise, multimodal inference is computationally intensive, and cloud deployment introduces latency concerns. Future work should focus on extending the framework to noisier industrial artifacts, developing simulator-agnostic abstractions, and improving self-correction mechanisms for broader applicability.