DIGITAL TWINS / 2026
Digital Twins for Complex Industrial Systems
How physics-informed digital twins enable real-time optimization of metallurgical processes without disrupting production.
ABSTRACT
Digital Twins for Complex Industrial Systems
Digital twins have emerged as a critical technology for industrial process optimization. This paper presents a physics-informed approach to digital twin construction for metallurgical processes, combining first-principles models with data-driven corrections to achieve real-time accuracy without prohibitive computational cost. We demonstrate the approach on electric arc furnace operations, showing 15% reduction in specific energy consumption and 23% improvement in tap-to-tap time predictability.
ARTICLE
MAIN CONTENT
The metallurgical industry faces a fundamental challenge: processes operate at extreme temperatures and pressures where direct measurement is limited, yet precise control is essential for quality and efficiency. Traditional process models — either purely physics-based or purely data-driven — each have significant limitations. Physics-based models (CFD, FEM) capture first principles but are computationally expensive and require simplifying assumptions. Data-driven models (ML, neural networks) are fast but require large training datasets and lack extrapolation guarantees. Our approach constructs a hybrid digital twin: a reduced-order physics model provides the structural backbone, while neural network correction terms learn the discrepancy between simplified physics and reality. This architecture preserves physical consistency while achieving data-driven accuracy.
FIGURE
Hybrid Digital Twin Architecture
Physics-informed neural network correcting a reduced-order furnace model.
CONTINUED
DISCUSSION
Validation on production EAF data demonstrates the approach achieves 94% accuracy in temperature prediction with 50ms inference time — suitable for real-time control. The digital twin enables what-if scenario evaluation: operators can simulate the effect of power profile changes, scrap mix variations, and oxygen injection strategies before implementation. Key results include: 15% reduction in specific energy consumption through optimized power profiles, 23% improvement in tap-to-tap time predictability, and 31% reduction in electrode consumption via optimized current density profiles. The system is currently deployed at two pilot sites with continuous learning from operational data.
CONCLUSION
CONCLUSION
Physics-informed digital twins represent a practical path to industrial AI deployment — combining the reliability of first-principles modeling with the accuracy of data-driven methods. Future work extends the approach to continuous casting and rolling processes, and explores federated learning across multiple sites while preserving data sovereignty.
RELATED RESEARCH
RELATED\nPUBLICATIONS.
Explore additional research in industrial intelligence and digital twins.
CONTACT
DISCUSS THIS\nRESEARCH.
We welcome collaboration inquiries, peer review, and industrial validation partnerships.