Digital Twins Framework

write in paragraphs about this

The paper presents a comprehensive framework for building **digital twins of chemical processes** using a **knowledge graph–based architecture** that integrates process knowledge, mathematical models, databases, and artificial intelligence tools. The central idea is that chemical processes are too complex and heterogeneous to be effectively modeled using isolated equations or standalone simulations. Instead, the authors propose representing process knowledge in a structured, semantic way that allows models, data, and domain rules to be systematically connected and reused. At the core of the framework is a **knowledge graph that separates and organizes two main ontologies: OntoModel and OntoProcess**. OntoModel captures the mathematical and physical modeling side, including variables, laws, formulas, and their relationships, while OntoProcess describes real-world process context such as operating conditions, descriptors, and constraints. Mathematical relationships are encoded using standards like MathML, enabling formulas to be stored in a machine-interpretable form and later converted into executable code for simulation environments such as SciPy, Pyomo, or Julia. This structure allows models to be dynamically assembled rather than manually constructed from scratch. To operationalize the framework, the authors introduce **autonomous functional agents** that interact with the knowledge graph. These include model assembly agents, calibration agents, rule inference systems, database query agents, and AI/LLM-based assistants. Together, they enable both **bottom-up model construction**, where models are built from known physical phenomena and laws, and **top-down model selection**, where candidate models are evaluated and selected based on how well they match observed process behavior under specific conditions. This dual approach allows the system to handle both well-understood and partially unknown processes. The framework is demonstrated through multiple case studies. In an **annular microreactor**, the system successfully assembles a mixing and reaction model from fundamental physical laws and calibrates it against experimental data. In a **Taylor–Couette reactor**, the top-down approach selects appropriate dispersion models depending on flow regimes and operating conditions, showing that model validity is context-dependent. Finally, in a **flow chemistry optimization problem**, the framework integrates reaction kinetics, mass transfer, and transport phenomena to guide multi-objective optimization, outperforming purely data-driven methods. Overall, the work argues that digital twin development in chemical engineering can be significantly improved by shifting from isolated modeling approaches to a **knowledge-driven, graph-based ecosystem**. By combining ontologies, databases, physics-based modeling, and AI agents, the framework enables scalable, reusable, and adaptive digital representations of chemical processes. This moves the field toward more automated and interoperable process modeling, with implications for chemical manufacturing, process intensification, and AI-assisted scientific discovery.