The search for sparse data in molecular datasets: Application of active learning to identify extremely low volatile organic compounds

Abstract

The formation of aerosol particles in the atmosphere is driven by the gas to particle conversion of extremely low volatile organic compounds (ELVOC), organic compounds with a particularly low saturation vapor pressure ($p_{\mathrm{Sat}}$). Identifying ELVOCs and their chemical structures is both experimentally and theoretically challenging: Measuring the very low $p_{\mathrm{Sat}}$ of ELVOCs is extremely difficult, and computing $p_{\mathrm{Sat}}$ for these often large molecules is computationally costly. Moreover, ELVOCs are underrepresented in available datasets of atmospheric organic species, which reduces the value of statistical models built on such data. We propose an active learning (AL) approach to efficiently identify ELVOCs in a data pool of atmospheric organic species with initially unknown $p_{\mathrm{Sat}}$. We assess the performance of our AL approach by comparing it to traditional machine learning regression methods, as well as ELVOC classification based on molecular properties. AL proves to be a highly efficient method for ELVOC identification with limitations on the type of ELVOC it can identify. We also show that traditional machine learning or molecular property-based methods can be adequate tools depending on the available data and desired degree of efficiency.