Multimodal LLM approach digitizes legacy P&ID diagrams into process graphs

A team of researchers has introduced a novel method for digitizing legacy piping and instrumentation diagrams (P&IDs) using multimodal large language models. P&IDs are critical engineering documents that encode the functional structure of process plants, but they remain underutilized for digital twins due to heterogeneous drawing standards and brittle rule-based extraction methods. The proposed two-stage workflow first performs visual extraction of equipment tags and symbols, then separately infers process topology—a departure from end-to-end digitization that treats both tasks as a single black box. By leveraging chemical engineering process knowledge as a reasoning guide, the model produces more accurate and structurally consistent process graphs.

Evaluated on two ANSI-standard P&ID case studies of increasing complexity, the approach demonstrated clear advantages over conventional methods. The decomposition of visual and topological reasoning stages led to higher fidelity in capturing plant connectivity and equipment relationships. This work highlights the potential of knowledge-guided, language-model-based workflows for scalable and semantically reliable P&ID digitization. For industries relying on legacy plant documentation (e.g., oil & gas, chemical processing), this could unlock faster and more accurate digital twin creation, reducing manual engineering effort and enabling intelligent operational decision support.