Theoretical Analyses and Performance Comparison of Physics-Based, Data-Driven, and Hybrid Models for Digital Twin Applications in Manufacturing

Digital twins (DT) are increasingly important in modern manufacturing systems to extend the physical system’s capabilities and improve manufacturing key performance indicators. Nevertheless, selecting the most effective modeling approach for DT services remains challenging. While physics-based models (PBM), data-driven models (DDM), and hybrid models are all utilized within DT implementations, there is a lack of theoretical understanding regarding their comparative performance under equivalent model input conditions. This research addresses this gap by developing a structured theoretical framework that relates model quality to key input resources: domain knowledge, modeling expertise, and data quality. The framework proposes that hybrid models can systematically outperform both PBM and DDM by optimizing the knowledge-data trade-off through the synergistic integration of physical principles with pattern recognition capabilities. To validate this theoretical framework, we implement and compare all three modeling approaches in an industrial stoneware floor tile polishing case study. Results reveal a performance hierarchy: the hybrid model (R² ≈ 0.55) significantly outperforms both the pure DDM approach (R² ≈ 0.40) and the PBM (R² ≈ 0.08) while providing superior transferability to different process conditions. The case study confirms that hybrid models can extract more value from identical input resources than either pure modeling paradigm alone. In addition to such comparisons, this research provides manufacturers with practical guidance for selecting appropriate modeling strategies in DT implementations, particularly valuable in scenarios with limited data or incomplete physical understanding of complex processes.