A Coulomb Matrix Approach for Predicting the Melting Point of Organic Molecules
ECS Meeting Abstracts · 07.07.2026 · MA2026-01 (6) · 708-708
DOI: 10.1149/ma2026-016708mtgabs ↗Kurz gefasst
Fachartikel in der Zeitschrift ECS Meeting Abstracts, erschienen am 07.07.2026. Verfasst von Nicholas Furth und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Microemulsions (μEs), composed of water, oil, and surfactants in a single-phase bicontinuous liquid state, have been reported as promising electrolytes for redox flow batteries (RFBs). We have suggested that such μEs provide favorable aspects of aqueous solutions that have high ionic conductivity and that of non-aqueous solutions which have a wide electrochemical window and more varied redox active molecules. We have shown that μEs do not require ion conduction through the organic portion. As with all RFBs, solubility limits of the redox active species limit achievable energy density. Maximizing energy density in these systems requires the oil sub-phase have the highest possible concentration of redox-active moieties. While much of our work has employed toluene as the organic solvent, we recently have explored electrolytes employing redox active oils, with the expectation that a pure oil will represent a limit of concentration. This requires oils that remain liquid at room temperature. The use of solid or high-melting oils would necessitate co-solvents that dilute the redox active oil, reducing its volume fraction resulting in lower energy density. Accurately predicting melting points of candidate redox active oils is therefore a critical step in the rational design of such μEs. We are using machine learning (ML) to predict likely oils. However, melting point prediction using machine learning remains a challenging task. This challenge stems from the deficiencies of standard descriptors in capturing the intermolecular interactions needed for an ML model to predict melting point. Traditional computational approaches, including quantum chemical calculations, molecular mechanics, and thermodynamic modeling, are often computationally expensive or limited in generalizability across diverse chemical spaces. Our ML pipeline provides a promising alternative, capable of rapidly identifying candidate molecules from large chemical libraries. Model performance, however, strongly depends on molecular representation. In this work, we integrate Coulomb matrices which capture the electrostatic interactions between nuclei in a molecule. We augment the Coulomb matrices with molecular descriptors generated using RDKit and PaDEL descriptors, and Morgan fingerprints to provide a comprehensive molecular representation. These descriptors are used with various ML models; two neural networks and three gradient boosted tree models; XGBoost, CatBoost, and LightGBM framework to deliver fast and accurate melting point predictions, achieving over 10% improvement in predictive performance relative to existing ML methods. In addition to descriptor and model optimization, we have also compiled the largest organic melting point dataset, being over twice as large as publicly available dataset. We examine predicting redox active oils which are sourced for RFB compatibility to demonstrate model effectiveness on real-world applications. We will show how our data-driven approach enables high-throughput screening of molecules suitable for room-temperature liquid oils, accelerating the formulation of µEs and the development of high-energy density RFBs. Beyond energy storage, the melting point model proposed in this framework is broadly applicable to molecular design challenges in energy systems, manufacturing, and materials science. By combining multiple molecular representations in a ML pipeline, we demonstrate a method for accurately predicting complex molecular properties. This pipeline is useful for high-throughput screening, accelerating experimental trial-and-error and guiding the discovery of functional molecules across diverse technological domains.
Zitieren
APA
Furth, N., Imel, A., Joy, D., & Zawodzinski, T. A. (2026). A Coulomb Matrix Approach for Predicting the Melting Point of Organic Molecules. ECS Meeting Abstracts, MA2026-01(6), 708-708. https://doi.org/10.1149/ma2026-016708mtgabs
Vancouver
Furth N, Imel A, Joy D, Zawodzinski TA. A Coulomb Matrix Approach for Predicting the Melting Point of Organic Molecules. ECS Meeting Abstracts. 2026;MA2026-01(6):708-708. doi:10.1149/ma2026-016708mtgabs
BibTeX
@article{Furth2026,
author = {Furth, Nicholas and Imel, Adam and Joy, Daniel and Zawodzinski, Thomas A.},
title = {A Coulomb Matrix Approach for Predicting the Melting Point of Organic Molecules},
journal = {ECS Meeting Abstracts},
year = {2026},
volume = {MA2026-01},
number = {6},
pages = {708--708},
doi = {10.1149/ma2026-016708mtgabs}
}
EUVIMED