REFINERY — Explainable Industrial AI

REFINERY is an engineering AI prototype for diagnosing disturbances and evaluating corrective actions in a simulated petroleum refining unit. The system combines a physical process model, observation-based diagnosis, predictive simulation, feedback, and an operator workspace. It represents six interconnected process loops, including heating, distillation, condensation, reflux, cooling, and protection. Engineers can run synthetic scenarios, inspect operating signals, identify potential disturbances, and compare the simulated consequences of corrective actions before making decisions. The interface provides interactive charts, alarms, diagnostic hypotheses, calculation history, engineering reports, and a video wall. REFINERY integrates AMD Token Factory using Qwen3.8-27B to explain diagnostic results in natural language. This integration has been verified through a successful API request. Physical simulation and diagnosis execute locally on CPU. The project includes a reproducible demonstration, public source code, technical documentation, and a recorded walkthrough. REFINERY is a research and engineering prototype using synthetic process data. Industrial deployment requires further validation against real operating conditions.