"Operational Digital Twin for Coke Drum Skirt Thermal Stress: Thermocouple-Informed FEA and PINN Transfer Learning Approach"

"This paper presents a practical framework for an operational digital twin for continuous thermal stress evaluation of coke drum skirts. Finite Element Analysis (FEA) delivers accurate predictions but is too computationally intensive and slow for real-time use. Physics-Informed Neural Networks (PINNs), by embedding governing physical laws into their training, provide a faster, physics-consistent alternative. We calibrate FEA models with thermocouple data to generate a spectrum of synthetic cycles, then apply transfer learning to train a PINN surrogate. Case studies on simulated coke drum cycles demonstrated that the PINN reduced computation time from 4-6 hours per FEA transient run to under 20 seconds per cycle while maintaining prediction errors within 3-5% relative to calibrated FEA outputs and within + 5 \u00b0C agreement with thermocouple measurements. The framework reduces computational effort by more than two orders of magnitude, enables continuous monitoring, and supports early anomaly detection, enhancing reliability and predictive maintenance. Operationally, this reduces failure risk and unplanned outages while improving process safety and unit uptime. "