Accurate prediction of ideal gas heat capacity C(T) using machine learning: data augmentation and ensemble learning for robust thermodynamic modeling

• Flexible framework using Mordred descriptors and neural networks for Cp prediction. • Cp normalized by infinite-temperature limit improves dataset balance. • High accuracy achieved (R² > 0.999) with bagged neural network ensemble. • Model generalizes well across chemical space, including complex molecules. Reliable predictions of ideal-gas molar heat capacities at constant pressure are a cornerstone of thermodynamic modeling. In the absence of a reliable model or correlation for the estimation of the ideal-gas heat capacity, an equation of state (EoS) alone cannot ensure thermodynamically consistent calculations of state-function variations, nor can it provide real-fluid heat capacities, since it only yields residual contributions. Changes in internal energy, enthalpy, entropy, exergy, and real-fluid heat capacities require the integration of the ideal-gas contribution which depends solely on the temperature-dependent ideal-gas heat capacity. Thus, coupling an EoS with an ideal-gas heat capacity model enables full thermodynamic property evaluation in process simulators. In this paper, we propose a deep-learning framework for continuous prediction of temperature-dependent ideal-gas molar heat capacities at constant pressure ( C p ) via artificial neural networks (ANNs) trained using molecular descriptors computed with Mordred and a high-quality experimental database comprising 1,471 compounds. The dataset was augmented using experimentally validated correlations, generating 100 C p data points per molecule over their corresponding temperature range and several ANN models were trained in a bagging ensemble configuration. The resulting model achieves excellent predictive accuracy (R² > 0.999 for both training and test sets) while demonstrating strong generalization to external datasets, indicating that the framework captures transferable structure-thermodynamic relationships. By explicitly treating temperature as an input variable, the proposed model provides smooth and continuous ideal-gas C p ( T ) predictions while preserving physically consistent thermodynamic trends across temperature ranges relevant to industrial applications. The developed prediction tool is available in a web interface (link).