A group contribution-based machine learning model to estimate the triple-point temperature

Abstract

This manuscript reports a new thermodynamic model for calculating triple-point temperature using a machine-learning algorithm and group contribution theory. The model was developed using the Auto-Machine Learning approach available in the auto-sklearn library in Python to identify the best algorithm for estimating this relevant property of pure compounds. Different input variables and ensembles of machine learning algorithms were assessed. The limitations and gaps in the proposed model are highlighted for different chemical families and functional groups. The results demonstrate that the Gradient Boosting algorithm achieved the best performance in estimating the triple-point temperature. The average absolute relative deviation $\left(\text{AARD}\right)$ of this model ranged from 0.85 to 5.73 % for the main chemical families included in the data analysis. The proposed model is reliable for calculating the triple-point temperatures of alkanes, alkynes, alcohols, cycloalkenes, polyols, nitriles, and anhydrides. However, the estimation of the triple-point temperature was challenging for polar compounds containing halogens and NO₂, which showed a non-ideal thermodynamic behavior. This study represents an initial step towards the development of an improved thermodynamic framework based on machine learning algorithms and group contribution theory for the accurate estimation of triple-point temperature.