Machine learning-based prediction of Joule-Thomson coefficient for cryogenic and industrial gases

This study investigates the application of machine learning techniques for predicting the Joule–Thomson coefficient of several industrial gases over a wide range of operating conditions, with relevance to cryogenic and gas-processing applications. Five different models were implemented and tested, including decision tree, extra trees, feedforward neural network, gradient boosting, and random forest, and were based on thermodynamic data obtained through Aspen HYSYS. The gases investigated include oxygen, nitrogen, carbon dioxide, hydrogen, helium, methane, ethane, argon, water, and hydrogen sulfide, covering a range of thermophysical behaviors relevant to industrial gas processing and cryogenic applications. The results demonstrate the capability of machine learning models to capture the nonlinear thermodynamic behavior of real gases. Both the gradient boosting ( R 2 = 0.9988) and random forest ( R 2 = 0.9737) models showed the highest predictive accuracy, with gradient boosting achieving near-perfect prediction coupled with an impressively low mean squared error. Furthermore, the gas-specific analysis showed that helium exhibited the largest prediction errors, likely due to its unique low-temperature thermophysical behavior and weak intermolecular interactions. Overall, the findings indicate that machine learning provides a robust and computationally efficient framework for estimating Joule-Thomson coefficients in cryogenic and gas-processing applications.