Hybrid Physics-Informed Machine Learning Frameworks for Predictive Thermodynamic Modeling

Recent advances in machine learning (ML) have opened new frontiers for modeling complex thermodynamic systems. Traditional thermodynamic property predictions often rely on methods limited in accuracy or scope. This research explores the potential of ML methods—specifically support vector regression (SVR) and physics-informed neural networks (PINNs)—to improve predictive accuracy for thermodynamic properties. We propose a hybrid framework combining these supervised learning algorithms with classical thermodynamic modeling concepts. Accompanying Python code examples using scikit-learn and TensorFlow demonstrate model training, cross-validation, and the application of basic physics-informed constraints using synthetic datasets. Execution of these examples shows that SVR can achieve high accuracy (𝑅2 ≈0.96, MAE ≈0.15, RMSE ≈0.23) on synthetic entropy data, a physics-informed neural network with non-negativity constraints (not a PDE PINN) attains 𝑅2 ≈0.78 (MAE≈0.52, RMSE ≈0.88) for heat capacity prediction, and a hybrid model with learned Shomate-like correction substantially improves performance to 𝑅2 ≈0.88 (MAE≈0.43, RMSE ≈0.65), all on synthetic data. While the provided code focuses on property prediction, the broader conceptual framework discussed herein extends to potential digital twin integration and reinforcement learning for operational optimization in complex energy systems, representing avenues for future development.