Physics-informed modular modeling of a gas turbine with component-level thermodynamic constraints and system-level consistency

Accurate and physically consistent modeling of gas turbine systems is essential for improving performance prediction, reliability, and energy efficiency. Conventional data-driven approaches often lack physical interpretability, while purely physics-based models struggle to capture complex nonlinear system behavior. This study presents a modular physics-informed modeling framework for gas turbine performance and emission prediction, in which thermodynamic constraints are embedded at the component level and preserved during system-level integration. The proposed approach incorporates differential thermodynamic relations for compressor and turbine components, along with an energy balance formulation for the combustor, directly into the learning process. A coordinated system-level fine-tuning strategy is introduced to enforce consistency across interconnected components and mitigate the accumulation of integration errors. The framework is evaluated using three datasets with varying levels of complexity, including a high-fidelity simulated turbofan engine dataset based on Pratt & Whitney PW-4056, the National Aeronautics and Space Administration (NASA) Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) benchmark dataset, and a real industrial gas turbine Predictive Emission Monitoring System (PEMS) dataset for emission prediction. At the component level, the physics-informed formulation achieves mean absolute percentage errors (MAPE) below 0.28% for key variables, outperforming conventional neural network models. At the system level, the proposed integration strategy reduces prediction errors by up to 86.39% for thrust, 81.85% for thrust-specific fuel consumption, and 75.70% for overall pressure ratio relative to baseline data-driven models. On the C-MAPSS dataset, the approach maintains MAPE values below 0.33% under limited sensor conditions. For the industrial PEMS dataset, the model achieves mean absolute errors of 0.8536 mg/m 3 for Carbon monoxide (CO) and 4.8419 mg/m 3 for Nitrogen oxides (NOx) under real operating conditions. These results demonstrate that integrating thermodynamic constraints with data-driven learning improves prediction accuracy, robustness, and physical consistency, particularly under varying operating conditions across gas turbine systems. The proposed framework provides a foundation for developing physics-consistent digital twins for condition monitoring and predictive maintenance applications.