Predicting second virial coefficients: a Gaussian process regression approach

Second virial coefficients represent a robust tool for describing the departure from ideality of a gas based on 2 body interactions in a system. The applications of second virial coefficients are diverse, ranging from the estimation of fluid properties to crystal growth, hence the broad interest in its calculation over the years. We propose the calculation of second virial coefficients within the new paradigm of data-intensive science, as an alternative to traditional methods. Our approach involves devising a set of predictors based on physical and chemical intuition, which encapsulates the most relevant features to 2-body interactions in a system. This is used jointly with Gaussian process regression to yield predictions of second virial coefficients for both organic and inorganic compounds. We explore the extrapolation and transferability capabilities of our model and show that good accuracy can be obtained in machine learning with the use of physico-chemical properties and molecular fingerprints solely selected using domain intuition and judgement, rather than canonical feature selection algorithms and universal descriptors.