Current and future machine learning approaches for modeling atmospheric cluster formation

Predicting Composition Evolution for a Sulfuric Acid-Dimethylamine System from Monomer to Nanoparticle Using Machine Learning

Experimental and theoretical studies on the compositional changes of new particle formation in the nucleation and initial growth stages of acid-base systems (2 and 5 nm) are extremely challenging. This study proposes a machine learning method for predicting the composition change of the sulfuric acid-dimethylamine system in the transformation from monomer to nanoparticle by learning the structure and composition information on small-sized sulfuric acid (SA)-dimethylamine (DMA) molecular clusters. Based on this method and changes in components, we found that the sulfuric acid-dimethylamine growth was mainly through the alternate adsorption of (SA)1(DMA)1, (SA)1(DMA)2, and (SA)1 clusters at the early stage of nucleation, which accounted for about 70, 20, and 10%, respectively. This can explain the nature of possible changes in cluster acidity during the initial nucleation stage for the sulfuric acid-dimethylamine system. This method can also predict the base-stabilization mechanism of the sulfuric acid-dimethylamine system without relying on any experimental data, thereby yielding results that are consistent with those of previous experimental measurement.

Reparameterization of GFN1-xTB for atmospheric molecular clusters: applications to multi-acid-multi-base systems

Atmospheric molecular clusters, the onset of secondary aerosol formation, are a major part of the current uncertainty in modern climate models. Quantum chemical (QC) methods are usually employed in a funneling approach to identify the lowest free energy cluster structures. However, the funneling approach highly depends on the accuracy of low-cost methods to ensure that important low-lying minima are not missed. Here we present a reparameterized GFN1-xTB model based on the clusteromics I-V datasets for studying atmospheric molecular clusters (AMC), denoted AMC-xTB. The AMC-xTB model reduces the mean of electronic binding energy errors from 7-11.8 kcal mol-1 to roughly 0 kcal mol-1 and the root mean square deviation from 7.6-12.3 kcal mol-1 to 0.81-1.45 kcal mol-1. In addition, the minimum structures obtained with AMC-xTB are closer to the ωB97X-D/6-31++G(d,p) level of theory compared to GFN1-xTB. We employ the new parameterization in two new configurational sampling workflows that include an additional meta-dynamics sampling step using CREST with the AMC-xTB model. The first workflow, denoted the "independent workflow", is a commonly used funneling approach with an additional CREST step, and the second, the "improvement workflow", is where the best configuration currently known in the literature is improved with a CREST + AMC-xTB step. Testing the new workflow we find configurations lower in free energy for all the literature clusters with the largest improvement being up to 21 kcal mol-1. Lastly, by employing the improvement workflow we massively screened 288 new multi-acid-multi-base clusters containing up to 8 different species. For these new multi-acid-multi-base cluster systems we observe that the improvement workflow finds configurations lower in free energy for 245 out of 288 (85.1%) cluster structures. Most of the improvements are within 2 kcal mol-1, but we see improvements up to 8.3 kcal mol-1. Hence, we can recommend this new workflow based on the AMC-xTB model for future studies on atmospheric molecular clusters.

Computational Tools for Handling Molecular Clusters: Configurational Sampling, Storage, Analysis, and Machine Learning

Computational modeling of atmospheric molecular clusters requires a comprehensive understanding of their complex configurational spaces, interaction patterns, stabilities against fragmentation, and even dynamic behaviors. To address these needs, we introduce the Jammy Key framework, a collection of automated scripts that facilitate and streamline molecular cluster modeling workflows. Jammy Key handles file manipulations between varieties of integrated third-party programs. The framework is divided into three main functionalities: (1) Jammy Key for configurational sampling (JKCS) to perform systematic configurational sampling of molecular clusters, (2) Jammy Key for quantum chemistry (JKQC) to analyze commonly used quantum chemistry output files and facilitate database construction, handling, and analysis, and (3) Jammy Key for machine learning (JKML) to manage machine learning methods in optimizing molecular cluster modeling. This automation and machine learning utilization significantly reduces manual labor, greatly speeds up the search for molecular cluster configurations, and thus increases the number of systems that can be studied. Following the example of the Atmospheric Cluster Database (ACDB) of Elm (ACS Omega, 4, 10965-10984, 2019), the molecular clusters modeled in our group using the Jammy Key framework have been stored in an improved online GitHub repository named ACDB 2.0. In this work, we present the Jammy Key package alongside its assorted applications, which underline its versatility. Using several illustrative examples, we discuss how to choose appropriate combinations of methodologies for treating particular cluster types, including reactive, multicomponent, charged, or radical clusters, as well as clusters containing flexible or multiconformer monomers or heavy atoms. Finally, we present a detailed example of using the tools for atmospheric acid-base clusters.

Clusterome: A Comprehensive Data Set of Atmospheric Molecular Clusters for Machine Learning Applications

Formation and growth of atmospheric molecular clusters into aerosol particles impact the global climate and contribute to the high uncertainty in modern climate models. Cluster formation is usually studied using quantum chemical methods, which quickly becomes computationally expensive when system sizes grow. In this work, we present a large database of ∼250k atmospheric relevant cluster structures, which can be applied for developing machine learning (ML) models. The database is used to train the ML model kernel ridge regression (KRR) with the FCHL19 representation. We test the ability of the model to extrapolate from smaller clusters to larger clusters, between different molecules, between equilibrium structures and out-of-equilibrium structures, and the transferability onto systems with new interactions. We show that KRR models can extrapolate to larger sizes and transfer acid and base interactions with mean absolute errors below 1 kcal/mol. We suggest introducing an iterative ML step in configurational sampling processes, which can reduce the computational expense. Such an approach would allow us to study significantly more cluster systems at higher accuracy than previously possible and thereby allow us to cover a much larger part of relevant atmospheric compounds.