Transfer Learning–Based Improvement of Low-Temperature Thermal Conductivity Prediction for N-Hydrogen and ANN-Based Thermophysical Property Modeling of Helium

The thermophysical properties of cryogenic gases, such as hydrogen, helium, and other low-temperature gases, play a crucial role in nuclear energy applications, particularly in reactor design, superconducting magnet cooling, cryogenic distribution systems, safety assessments, and thermal efficiency. The accurate estimation of these properties over a wide range of operating conditions is essential for reliable cryogenic process design and transient system analysis.Traditionally, thermophysical properties have been predicted using empirical correlations, theoretical equations of state, and experimentally derived formulations, which are implemented in established property packages such as RefProp®, CoolProp®, HEPAK®, and GASPAK®. While these approaches have proven highly effective, deviations may still arise in specific operating regions, particularly at cryogenic temperatures where experimental data are sparse and fluid behavior becomes highly nonlinear.In recent years, deep learning and artificial neural networks (ANNs) have emerged as powerful data-driven tools for modeling complex nonlinear relationships across scientific and engineering applications, including thermophysical property prediction. Rather than replacing established cryogenic property routines, ANN-based approaches may offer a complementary framework for incorporating newly available experimental data sets and refining predictions in operating regions where deviations from reference data are observed.The present work explores the possibility of utilizing the capabilities of ANNs and transfer learning techniques for cryogenic thermophysical property prediction. Residual transfer learning is applied to improve the thermal conductivity prediction for n-hydrogen (normal hydrogen) in the temperature range of 15 to 30 K using published experimental data sets. Helium properties are modeled as a proof-of-concept study to evaluate the capability of the ANN architecture to reproduce the global thermodynamic trends and overall nonlinear variations in cryogenic thermophysical properties, while also examining its limitations in regions with strong local property variations.Furthermore, In addition, a graphical user interface–based module, CRYOProp, has been developed to improve usability and accessibility. The present work highlights the applicability, limitations, and potential of the proposed methodology.