Efficient and Accurate Machine Learning of Thermophysical Properties from Small Data

Thermophysical properties of working fluids play a crucial role in advanced propulsion systems, directly influencing heat transfer and fluid flow, and phase behavior under extreme conditions. Accurate and efficient evaluations of those thermophysical properties are essential for large-scale modeling and high-fidelity simulations. However, data-driven prediction remains challenging under small and highly nonlinear data sets. This study proposes a novel modeling framework integrating Gaussian process regression (GPR) with an artificial neural network (ANN) for accurate and rapid predictions of thermophysical properties. The framework first calibrates a GPR model on a small dataset, which subsequently generates high-fidelity synthetic data for ANN training. An adaptive data augmentation strategy is employed to iteratively expand the training set before the prescribed prediction accuracy is achieved. Three representative cases are demonstrated: vapor–liquid equilibrium of binary mixtures, density of supercritical carbon dioxide, and energy and dipole moment of water molecules. Compared with an ANN using the original small data set, an ANN with linearly interpolated data set, and GPR models, the proposed model achieves superior accuracy and comparable predictive performance to GPR, with up to an order-of-magnitude faster inference speed. The results highlight its potential as a robust and scalable tool for accurate and efficient modeling of complex thermophysical properties in propulsion-related applications.