• Quantify and map thermodynamic irreversibilities in the GE9FA gas turbine by rigorously applying first and second law analyses to establish a benchmark entropy-generation profile that guides subsequent optimization efforts. • Design and execute a statistically rigorous benchmark that ranks seven state-of-the-art optimizers (e.g., Stochastic Gradient Descent (SGD), Root Mean Square Propagation (RMSprop), Adaptive Gradient (Adagrad), Adadelta, Adaptive Moment Estimation (Adam), Adamax, and Nesterov-accelerated Adam (Nadam) for training physics-informed deep neural networks on high-fidelity gas-turbine data. • Develop a multi-criteria evaluation protocol that contrasts four families of data-driven surrogates, deep neural network with Adam optimizer, gradient boosting, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), on identical transient gas-turbine datasets. • Create and validate a physics-informed deep-learning digital twin that reproduces the full transient thermal dynamics across all six critical performance categories, transforming the engine into a continuously self-diagnosing asset. • Enable AI-driven generation of interactive T-S and P-V diagrams with sub-second latency and embed these visualizations in an advanced monitoring dashboard that translates thermodynamic anomalies into actionable maintenance prescriptions to reduce unplanned downtime Gas turbine technology is essential for resolving the energy trilemma while facilitating renewable energy integration, yet existing physics-based models lack adaptability to real-world variations and purely data-driven approaches frequently violate fundamental thermodynamic laws. This study presents a novel physics-constrained deep learning framework that uniquely integrates first-principles thermodynamics with operational data assimilation to create a high-fidelity digital twin of a GE 9FA heavy-duty gas turbine, distinguishing itself from conventional approaches by ensuring thermodynamic consistency while adapting to measured performance. The digital twin employs neural network architectures regularized by conservation laws and thermodynamic cycle constraints to forecast and interactively visualize T–S and P–V trajectories, translating subtle efficiency variations into actionable operational insights. Comprehensive validation across six key operating regimes demonstrates low predictive error and robust performance, confirming that physics constraints enhance generalization compared to standard machine learning baselines. These capabilities support proactive maintenance strategies and long-term efficiency optimization, representing a significant advancement toward intelligent, self-optimizing turbine systems for reliable energy management.
Shah et al. (Wed,) studied this question.