A Physically Consistent Framework for AI-Native Process Simulation: Autonomous Flowsheet Construction and Solver-Aware Convergence

Abstract: Process simulation requires engineers to translate process objectives into flowsheet topology, thermodynamic models, equipment specifications, and numerical solution procedures. This work proposes a framework in which generative AI constructs and revises steady-state flowsheets while a deterministic process engine calculates process states and evaluates convergence. Machine-validatable data contracts define components, streams, unit operations, thermodynamic methods, operating conditions, and connectivity. These contracts allow proposed flowsheets to be checked for missing inputs, incompatible connections, and unsupported specifications before execution. The initial implementation focuses on equation-of-state-based thermodynamics, flash calculations, fundamental unit operations, graph-based flowsheet execution, and explicit material and energy balance residuals. During simulation, the engine returns structured diagnostics for incomplete specifications and numerical failures. The AI layer uses this feedback to revise equipment settings or flowsheet configurations, while calculated stream properties and equipment outputs remain the responsibility of the process engine. Representative cases are used to evaluate construction accuracy, specification validity, successful execution, conservation residuals, agreement with reference calculations, and recovery from invalid or incomplete configurations. By separating flowsheet generation from numerical calculation and making solver failures explicit, the framework aims to support more autonomous process simulation without obscuring the basis of its results.

READ ABSTRACT PDF🔗