An interpretable molecular descriptor for machine learning predictions in atmospheric science

L. Lind, H. Sandström, P. Rinke · 2025

Read An interpretable molecular descriptor for machine learning predictions in atmospheric science by L. Lind, H. Sandström, P. Rinke (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-20465.txt.

Introduction

Linus Lind, Hilda Sandström, and Patrick Rinke make ATMOMACCS, a clear molecular code for machine learning on air science. Work on mist growth lacks codes fit for air-borne mixes. Stock code lists, tied to fixed parts of a molecule, miss big, much-oxidized mixes that fill the air. Their code joins the 166 MACCS keys with parts drawn from the SIMPOL way to gauge vapor push.

Nets on the new code cut miss rates for four traits: vapor push, split rates, glass shift heat, and boil heat loss. On four air data sets, cuts range from small to vast. Trait checks show vapor push and splits hinge on carbon count and oxygen marks. Glass and boil traits hinge more on bond types.

Catalog record

CREATORL. Lind, H. Sandström, P. Rinke
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-20465

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

An interpretable molecular descriptor for machine learning predictions in atmospheric science Preprint: AIP/123-QED L. Lind, H. Sandström, P. Rinke Abstract The study of aerosol formation and chemistry using machine learning is limited by the lack of molecular descriptors suited to atmospheric compounds. Interpretable models are particularly affected because they often rely on dictionary-based descriptors tied to specific molecular substructures, which currently fail to capture the full range of organic atmospheric compounds, including large, highly oxidized molecules common in the atmosphere. We introduce ATMOMACCS, an interpretable descriptor combining the 166 binary keys of the MACCS fingerprint with motifs inspired by the SIMPOL method for estimating saturation vapor pressures. We show that ATMOMACCS based models improve predictions of saturation vapor pressures (7-8% error reduction), equilibrium partition coefficients (5% and 9% error reduction), glass transition temperatures (22% error reduction), and enthalpy of vaporization (61% error reduction) on four datasets with atmospheric compounds. Feature analysis shows that saturation vapor pressure and partition coefficients are governed by carbon number and oxygen-related features, whereas other phase-transition properties (e.g., enthalpy of vaporization, glass transition temperature) depend on carbon–hydrogen bond types and the presence of heteroatoms other than oxygen. This highlights the generalizability of ATMOMACCS across different datasets and properties as an interpretable molecular descriptor. I Introduction Atmospheric aerosol particles contribute to climate change by scattering and absorbing sunlight and serving as cloud condensation nuclei. Pörtner et al. 2022 The presence of particles in the atmosphere also worsens air pollution and human health. Pozzer et al. 2023 Determining which compounds drive aerosol formation is an ongoing research challenge, particularly because the number of atmospheric compounds is estimated at 10 5 –10 6 . Goldstein and Galbally 2007 The process leading to aerosol formation and growth includes atmospheric compounds originating from natural and anthropogenic emissions. Hallquist et al. 2009 20-90 % of the total formed sub-micron particle mass can consist of organic compounds. Chen, Li, and Zhu 2022 ; Crippa et al. 2013 ; Zhang et al. 2011 ; Jimenez et al. 2009 ; Zhang et al. 2007 In the atmosphere, emissions of organic compounds undergo chemical transformations, particularly oxidation, producing a wide range of reaction products that are typically larger and chemically more complex in terms of elemental composition and functional groups. Bianchi et al. 2019 Consequently, formed oxidized compounds generally have lower volatility and a propensity to condense into the particulate phase, forming so-called secondary organic aerosols. Ehn et al. 2014 ; Kroll and Seinfeld 2008 ; Donahue et al. 2012 Figure 1: Molecular emissions from various sources enter the atmosphere, where they undergo reactions with molecules like ozone and hydroxyl radicals. These reactions produce oxidized compounds with diverse functional groups, which are key to understanding atmospheric processes leading to particle formation. One strategy to identify candidate atmospheric compounds that likely contribute to secondary aerosol formation is to screen their physicochemical properties related to particle formation, such as saturation vapor pressures ( P_{sat} ) and equilibrium partition coefficients. Besel et al. 2023 ; Besel et al. 2024 ; Li, Buchholz, and Hyttinen 2024 ; Krüger et al. 2022 ; Krüger et al. 2025 ; Lumiaro et al. 2021 However, experimentally characterizing these properties is challenging. Laboratory measurements are slow and labor intensive, producing datasets with only hundreds to thousands of species (e.g., Pankow and Asher 2008 ), far fewer than the estimated hundreds of thousands of atmospheric compounds. Goldstein and Galbally 2007 Computational methods provide an alternative route for high-throughput screening and property prediction. These approaches range from empirical models, which offer rapid but potentially crude estimates, to quantum chemistry calculations, which yield more precise predictions at higher computational cost. Klamt and Schüürmann 1993 ; Klamt et al. 1998 ; BIOVIA Dassault Systèmes ; Pankow and Asher 2008 ; Compernolle, Ceulemans, and Müller 2011 By suggesting molecules for further experimental studies, computational methods help bridge the gap between experimental feasibility and atmospheric complexity. Among computational approaches, group contribution methods are widely used in atmospheric science to estimate molecular properties by adding the effects of predefined structural groups. These methods have been applied to predict properties such as P_{sat} , enthalpies of vaporization ( \Delta H_{vap} ), refractive indices, molar volumes, densities, viscosities, and glass transition temperatures ( T_{g} ). Pankow and Asher 2008 ; Ramírez-Verduzco 2022 ; Li et al. 2020 ; Cai et al. 2017 ; Su, Zhao, and Deng 2017 A notable group contribution method is SIMPOL, which estimates P_{sat} based on 30 predefined structural group terms. Pankow and Asher 2008 Other P_{sat} estimation methods include EVAPORATION, Compernolle, Ceulemans, and Müller 2011 Nannoolal, Nannoolal, Rarey, and Ramjugernath 2008 Myrdal and Yalkowsky, Myrdal and Yalkowsky 1997 and Tochigi. Tochigi et al. 2010 Group contribution methods, though efficient, are limited by the small datasets on which they were parametrized Pankow and Asher 2008 ; Li et al. 2020 ; Compernolle, Ceulemans, and Müller 2011 and show reduced accuracy against experimental benchmarks, Bannan et al. 2017 ; Kurtén et al. 2016 partly because they ignore relative positioning of functional groups. A new wave of models instead uses machine learning for property prediction of atmospheric compounds. Bortolussi et al. 2025 ; Hyttinen, Pihlajamäki, and Häkkinen 2022 ; Lumiaro et al. 2021 ; Besel et al. 2023 ; Elm et al. 2020 ; Sandström and Rinke 2025 Unlike traditional group contribution methods, which are often linear, machine learning models can capture complex, nonlinear relationships between molecular structures and properties. The success of machine learning models depends on molecular representations, or descriptors, which convert chemical structures into numerical formats that models can process. Mikulskis, Alexander, and Winkler 2019 Molecular descriptors range from simple one-dimensional properties (e.g., molecular weight) and two-dimensional molecular fingerprint vectors to complex three-dimensional descriptors reflecting spatial atomic arrangements and interactions. Morgan 1965 ; Rogers and Hahn 2010 ; Carhart, Smith, and Venkataraghavan 1985 ; National Center for Biotechnology Information 2009 ; Landrum 2022 ; Accelrys 2011 ; Rupp et al. 2012 ; Huo and Rupp 2022 Dictionary-based molecular fingerprints, including MACCS, Accelrys 2011 PubChem, National Center for Biotechnology Information 2009 and Klekota-Roth, Klekota and Roth 2008 are valued for their interpretability because they encode the presence or absence of functional groups in a straightforward, human-readable way. Such interpretability is a highly-sought after trait in machine learning, providing both insights and confidence to model predictions. Many of the available molecular descriptors were originally designed for general organic chemistry and may overlook structural characteristics of atmospheric molecules. Sandström and Rinke 2025 Atmospheric compounds often contain many oxygen- and nitrogen-rich functional groups that influence their chemistry. Bianchi et al. 2019 Omitting these features can reduce predictive accuracy and limit understanding of model predictions. Thus, combining machine learning with group contribution approaches for atmospheric applications could improve predictive performance by leveraging the strengths of both strategies. For example, Krüger et al. Krüger et al. 2025 integrated SIMPOL group contributions with graph neural networks (GNNs), which learn molecular properties from molecular graphs, and achieved more accurate P_{sat} predictions. Despite this, such hybrid strategies remain largely unexplored in atmospheric chemistry. To address this need, we introduce ATMOMACCS, a molecular descriptor that combines the interpretability of the MACCS fingerprint with motifs derived from the SIMPOL group contribution method. Pankow and Asher 2008 By integrating atmospheric

Read the complete text