Predicting vapor pressures using neural networks

Abstract

Calculating surface vapor pressures of volatile inorganic components, nitric acid, hydrochloric acid and ammonia, is essential for modeling condensation and evaporation processes occurring in atmospheric aerosols. The vapor pressure of these compounds depends on temperature, relative humidity, phase state, and particle composition, and their calculation consumes an enormous amount of computer time in Eulerian photochemical/aerosol models. Here we use a thermodynamic model to generate a large set of vapor pressure data as a function of aerosol composition, relative humidity, and temperature. These data are then used as a training set for neural networks. Once the networks memorize the data, interpolation of vapor pressures for intermediate compositions, temperatures and relative humidities is automatic. The neural network models are able to reproduce the values predicted by the thermodynamic models accurately and are 4–1200 times faster depending on atmospheric conditions and the assumptions employed in the thermodynamic calculations.