Machine Learning Model to Predict Saturation Vapor Pressures of Atmospheric Aerosol Constituents

Abstract

We present a novel machine learning (ML) model for predicting saturation vapor pressures (p_sat), a physical property of use to describe transport, distribution, mass transfer, and fate of environmental toxins and contaminants. The ML model uses σ-profiles from the conductor-like screening model (COSMO) as molecular descriptors. The main advantages in using σ-profiles instead of other types of molecular representations are the relatively small size of the descriptor and the fact that the addition of new elements does not affect the size of the descriptor. The ML model was trained separately for liquid and solid compounds using experimental vapor pressures at various temperatures. The 95% confidence intervals of the error in the liquid- and solid-phase log₁₀(p_sat/Pa) are 1.02 and 1.4, respectively. Especially our solid-phase model outperforms all group-contribution models in predicting experimental sublimation pressures of solid compounds. To demonstrate its applicability, the model was used to predict p_sat of atmospherically relevant species, and the values were compared with those obtained from a new experimental method. Here, our model provided a tool for a better description of this critical property and gave a higher confidence in the measurements.

Accurate saturation vapor pressure estimates of environmental contaminants are lacking in the low volatility range. Our quantum chemistry-based machine learning model provides a novel tool for predicting vapor pressure.