Mechanism–AI Fusion in Refining Simulation Software: A Module-Level Review

Refining simulation software has remained functionally static for decades: property calculation, reaction kinetics, unit-operation modeling, and computational fluid dynamics (CFD) solving modules are hard-coded, unidirectional, and cannot be calibrated online against drifting plant conditions. Purely mechanistic approaches provide irreplaceable rigor but are constrained by computational cost; purely data-driven approaches offer speed and adaptivity but lack the interpretability, generalization, and physical guarantees required for safety-critical deployment. This review examines how mechanism–AI fusion—encompassing physics-informed neural networks (PINNs), reduced-order models (ROMs), operator learning, and closure-model augmentation—is reshaping these functional modules at the architectural level. We propose a module-level taxonomy that maps each fusion strategy to the specific simulation module it transforms, and we distinguish this paradigm from related concepts (hybrid modeling, physics-informed machine learning, grey-box modeling, and digital twins) by specifying the bidirectional information flows between mechanistic and AI components. A systematic literature survey following PRISMA-style methodology identified 133 references from Web of Science, Scopus, and IEEE Xplore (2000–2025). We provide enhanced comparative tables that include training cost, data requirements, extrapolation ability, uncertainty quantification (UQ), scalability, and industrial maturity for each method class. Industrial software benchmarking is presented as a qualitative comparison rather than a subjective ranking. Three industrial case studies (gasifier, FCC unit, distillation column) are analyzed with explicit baselines, deployment durations, and performance indicators, clearly identifying vendor-reported versus independently verified data. We discuss emerging topics—AI-native scientific software, foundation models for scientific computing, autonomous engineering agents, trustworthy and explainable AI, verification & validation (V&V), and MLOps for scientific applications—that will shape the next decade. The review concludes with graded findings (verified, supported, emerging, conceptual, and policy-level) to clearly distinguish established results from industrial observations and conceptual proposals. The central finding is that mechanism–AI fusion transforms refining simulation software from a fixed-algorithm tool into an adaptive multi-fidelity platform, but the transition from advisory and optimization applications to safety-critical closed-loop control requires standardized V&V protocols, UQ frameworks, and process-industry-specific safety governance that do not yet exist.