Physics-Informed Neural Network Digital Twin for Dynamic Tray-Wise Modeling of Distillation Columns under Transient Operating Conditions

A research team has developed a novel Physics-Informed Neural Network (PINN) framework that creates a highly accurate digital twin for industrial distillation columns. Unlike standard data-driven models (LSTM, GRU, Transformer), this AI embeds fundamental chemical engineering physics—including vapor-liquid equilibrium, mass/energy balances, and the McCabe-Thiele method—directly into its neural network loss function. This forces the model to respect real-world thermodynamic constraints while learning from data. The system was trained and validated on a high-fidelity synthetic dataset comprising 961 measurements over 8 hours of transient operation, generated using the industry-standard Aspen HYSYS simulator for a 16-sensor binary distillation system.

The results are significant for industrial automation. The proposed PINN digital twin achieved a root mean square error (RMSE) of just 0.00143 for predicting key mole fractions, with an R² score of 0.9887. This performance represents a 44.6% reduction in error compared to the best purely data-driven baseline model. Crucially, it accurately captured complex dynamic behaviors like feed tray responses and pressure transients during operational upsets. This combination of high accuracy and built-in physics compliance makes the model a robust foundation for critical industrial applications, moving beyond simple prediction to reliable operational guidance.

This work demonstrates a practical path forward for AI in process industries. By creating a digital twin that is both data-informed and physics-consistent, the researchers have built a tool that engineers can trust for real-time decision support. The model's ability to simulate tray-by-tray temperature and composition profiles under changing conditions enables new capabilities in soft sensing (inferring unmeasured variables), model-predictive control, and early anomaly detection. This directly addresses the core challenge of applying AI to safety-critical physical systems: ensuring predictions are not just statistically plausible but physically possible.